power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
/* SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
|
* Copyright(c) 2010-2018 Intel Corporation
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include <string.h>
|
|
|
|
|
|
|
|
|
|
#include <rte_lcore.h>
|
|
|
|
|
#include <rte_cycles.h>
|
|
|
|
|
#include <rte_atomic.h>
|
|
|
|
|
#include <rte_malloc.h>
|
2018-10-26 13:38:24 +01:00
|
|
|
#include <inttypes.h>
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
|
|
|
|
|
#include "rte_power.h"
|
|
|
|
|
#include "rte_power_empty_poll.h"
|
|
|
|
|
|
|
|
|
|
#define INTERVALS_PER_SECOND 100 /* (10ms) */
|
|
|
|
|
#define SECONDS_TO_TRAIN_FOR 2
|
|
|
|
|
#define DEFAULT_MED_TO_HIGH_PERCENT_THRESHOLD 70
|
|
|
|
|
#define DEFAULT_HIGH_TO_MED_PERCENT_THRESHOLD 30
|
|
|
|
|
#define DEFAULT_CYCLES_PER_PACKET 800
|
|
|
|
|
|
|
|
|
|
static struct ep_params *ep_params;
|
|
|
|
|
static uint32_t med_to_high_threshold = DEFAULT_MED_TO_HIGH_PERCENT_THRESHOLD;
|
|
|
|
|
static uint32_t high_to_med_threshold = DEFAULT_HIGH_TO_MED_PERCENT_THRESHOLD;
|
|
|
|
|
|
|
|
|
|
static uint32_t avail_freqs[RTE_MAX_LCORE][NUM_FREQS];
|
|
|
|
|
|
|
|
|
|
static uint32_t total_avail_freqs[RTE_MAX_LCORE];
|
|
|
|
|
|
|
|
|
|
static uint32_t freq_index[NUM_FREQ];
|
|
|
|
|
|
|
|
|
|
static uint32_t
|
|
|
|
|
get_freq_index(enum freq_val index)
|
|
|
|
|
{
|
|
|
|
|
return freq_index[index];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
set_power_freq(int lcore_id, enum freq_val freq, bool specific_freq)
|
|
|
|
|
{
|
|
|
|
|
int err = 0;
|
|
|
|
|
uint32_t power_freq_index;
|
|
|
|
|
if (!specific_freq)
|
|
|
|
|
power_freq_index = get_freq_index(freq);
|
|
|
|
|
else
|
|
|
|
|
power_freq_index = freq;
|
|
|
|
|
|
|
|
|
|
err = rte_power_set_freq(lcore_id, power_freq_index);
|
|
|
|
|
|
|
|
|
|
return err;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
exit_training_state(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
RTE_SET_USED(poll_stats);
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
enter_training_state(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
poll_stats->iter_counter = 0;
|
|
|
|
|
poll_stats->cur_freq = LOW;
|
|
|
|
|
poll_stats->queue_state = TRAINING;
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
enter_normal_state(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
/* Clear the averages arrays and strs */
|
|
|
|
|
memset(poll_stats->edpi_av, 0, sizeof(poll_stats->edpi_av));
|
|
|
|
|
poll_stats->ec = 0;
|
|
|
|
|
memset(poll_stats->ppi_av, 0, sizeof(poll_stats->ppi_av));
|
|
|
|
|
poll_stats->pc = 0;
|
|
|
|
|
|
|
|
|
|
poll_stats->cur_freq = MED;
|
|
|
|
|
poll_stats->iter_counter = 0;
|
|
|
|
|
poll_stats->threshold_ctr = 0;
|
|
|
|
|
poll_stats->queue_state = MED_NORMAL;
|
|
|
|
|
RTE_LOG(INFO, POWER, "Set the power freq to MED\n");
|
|
|
|
|
set_power_freq(poll_stats->lcore_id, MED, false);
|
|
|
|
|
|
|
|
|
|
poll_stats->thresh[MED].threshold_percent = med_to_high_threshold;
|
|
|
|
|
poll_stats->thresh[HGH].threshold_percent = high_to_med_threshold;
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
enter_busy_state(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
memset(poll_stats->edpi_av, 0, sizeof(poll_stats->edpi_av));
|
|
|
|
|
poll_stats->ec = 0;
|
|
|
|
|
memset(poll_stats->ppi_av, 0, sizeof(poll_stats->ppi_av));
|
|
|
|
|
poll_stats->pc = 0;
|
|
|
|
|
|
|
|
|
|
poll_stats->cur_freq = HGH;
|
|
|
|
|
poll_stats->iter_counter = 0;
|
|
|
|
|
poll_stats->threshold_ctr = 0;
|
|
|
|
|
poll_stats->queue_state = HGH_BUSY;
|
|
|
|
|
set_power_freq(poll_stats->lcore_id, HGH, false);
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
enter_purge_state(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
poll_stats->iter_counter = 0;
|
|
|
|
|
poll_stats->queue_state = LOW_PURGE;
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
set_state(struct priority_worker *poll_stats,
|
|
|
|
|
enum queue_state new_state)
|
|
|
|
|
{
|
|
|
|
|
enum queue_state old_state = poll_stats->queue_state;
|
|
|
|
|
if (old_state != new_state) {
|
|
|
|
|
|
|
|
|
|
/* Call any old state exit functions */
|
|
|
|
|
if (old_state == TRAINING)
|
|
|
|
|
exit_training_state(poll_stats);
|
|
|
|
|
|
|
|
|
|
/* Call any new state entry functions */
|
|
|
|
|
if (new_state == TRAINING)
|
|
|
|
|
enter_training_state(poll_stats);
|
|
|
|
|
if (new_state == MED_NORMAL)
|
|
|
|
|
enter_normal_state(poll_stats);
|
|
|
|
|
if (new_state == HGH_BUSY)
|
|
|
|
|
enter_busy_state(poll_stats);
|
|
|
|
|
if (new_state == LOW_PURGE)
|
|
|
|
|
enter_purge_state(poll_stats);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
set_policy(struct priority_worker *poll_stats,
|
|
|
|
|
struct ep_policy *policy)
|
|
|
|
|
{
|
|
|
|
|
set_state(poll_stats, policy->state);
|
|
|
|
|
|
|
|
|
|
if (policy->state == TRAINING)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
poll_stats->thresh[MED_NORMAL].base_edpi = policy->med_base_edpi;
|
|
|
|
|
poll_stats->thresh[HGH_BUSY].base_edpi = policy->hgh_base_edpi;
|
|
|
|
|
|
|
|
|
|
poll_stats->thresh[MED_NORMAL].trained = true;
|
|
|
|
|
poll_stats->thresh[HGH_BUSY].trained = true;
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
update_training_stats(struct priority_worker *poll_stats,
|
|
|
|
|
uint32_t freq,
|
|
|
|
|
bool specific_freq,
|
|
|
|
|
uint32_t max_train_iter)
|
|
|
|
|
{
|
|
|
|
|
RTE_SET_USED(specific_freq);
|
|
|
|
|
|
|
|
|
|
uint64_t p0_empty_deq;
|
|
|
|
|
|
|
|
|
|
if (poll_stats->cur_freq == freq &&
|
|
|
|
|
poll_stats->thresh[freq].trained == false) {
|
|
|
|
|
if (poll_stats->thresh[freq].cur_train_iter == 0) {
|
|
|
|
|
|
|
|
|
|
set_power_freq(poll_stats->lcore_id,
|
|
|
|
|
freq, specific_freq);
|
|
|
|
|
|
|
|
|
|
poll_stats->empty_dequeues_prev =
|
|
|
|
|
poll_stats->empty_dequeues;
|
|
|
|
|
|
|
|
|
|
poll_stats->thresh[freq].cur_train_iter++;
|
|
|
|
|
|
|
|
|
|
return;
|
|
|
|
|
} else if (poll_stats->thresh[freq].cur_train_iter
|
|
|
|
|
<= max_train_iter) {
|
|
|
|
|
|
|
|
|
|
p0_empty_deq = poll_stats->empty_dequeues -
|
|
|
|
|
poll_stats->empty_dequeues_prev;
|
|
|
|
|
|
|
|
|
|
poll_stats->empty_dequeues_prev =
|
|
|
|
|
poll_stats->empty_dequeues;
|
|
|
|
|
|
|
|
|
|
poll_stats->thresh[freq].base_edpi += p0_empty_deq;
|
|
|
|
|
poll_stats->thresh[freq].cur_train_iter++;
|
|
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
if (poll_stats->thresh[freq].trained == false) {
|
|
|
|
|
poll_stats->thresh[freq].base_edpi =
|
|
|
|
|
poll_stats->thresh[freq].base_edpi /
|
|
|
|
|
max_train_iter;
|
|
|
|
|
|
|
|
|
|
/* Add on a factor of 0.05%
|
|
|
|
|
* this should remove any
|
|
|
|
|
* false negatives when the system is 0% busy
|
|
|
|
|
*/
|
|
|
|
|
poll_stats->thresh[freq].base_edpi +=
|
|
|
|
|
poll_stats->thresh[freq].base_edpi / 2000;
|
|
|
|
|
|
|
|
|
|
poll_stats->thresh[freq].trained = true;
|
|
|
|
|
poll_stats->cur_freq++;
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline uint32_t
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
update_stats(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
uint64_t tot_edpi = 0, tot_ppi = 0;
|
|
|
|
|
uint32_t j, percent;
|
|
|
|
|
|
|
|
|
|
struct priority_worker *s = poll_stats;
|
|
|
|
|
|
|
|
|
|
uint64_t cur_edpi = s->empty_dequeues - s->empty_dequeues_prev;
|
|
|
|
|
|
|
|
|
|
s->empty_dequeues_prev = s->empty_dequeues;
|
|
|
|
|
|
|
|
|
|
uint64_t ppi = s->num_dequeue_pkts - s->num_dequeue_pkts_prev;
|
|
|
|
|
|
|
|
|
|
s->num_dequeue_pkts_prev = s->num_dequeue_pkts;
|
|
|
|
|
|
|
|
|
|
if (s->thresh[s->cur_freq].base_edpi < cur_edpi) {
|
|
|
|
|
|
|
|
|
|
/* edpi mean empty poll counter difference per interval */
|
|
|
|
|
RTE_LOG(DEBUG, POWER, "cur_edpi is too large "
|
2018-10-26 13:38:24 +01:00
|
|
|
"cur edpi %"PRId64" "
|
|
|
|
|
"base edpi %"PRId64"\n",
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
cur_edpi,
|
|
|
|
|
s->thresh[s->cur_freq].base_edpi);
|
|
|
|
|
/* Value to make us fail need debug log*/
|
|
|
|
|
return 1000UL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
s->edpi_av[s->ec++ % BINS_AV] = cur_edpi;
|
|
|
|
|
s->ppi_av[s->pc++ % BINS_AV] = ppi;
|
|
|
|
|
|
|
|
|
|
for (j = 0; j < BINS_AV; j++) {
|
|
|
|
|
tot_edpi += s->edpi_av[j];
|
|
|
|
|
tot_ppi += s->ppi_av[j];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
tot_edpi = tot_edpi / BINS_AV;
|
|
|
|
|
|
|
|
|
|
percent = 100 - (uint32_t)(((float)tot_edpi /
|
|
|
|
|
(float)s->thresh[s->cur_freq].base_edpi) * 100);
|
|
|
|
|
|
|
|
|
|
return (uint32_t)percent;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2020-02-09 13:27:49 +01:00
|
|
|
static __rte_always_inline void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
update_stats_normal(struct priority_worker *poll_stats)
|
|
|
|
|
{
|
|
|
|
|
uint32_t percent;
|
|
|
|
|
|
|
|
|
|
if (poll_stats->thresh[poll_stats->cur_freq].base_edpi == 0) {
|
|
|
|
|
|
|
|
|
|
enum freq_val cur_freq = poll_stats->cur_freq;
|
|
|
|
|
|
|
|
|
|
/* edpi mean empty poll counter difference per interval */
|
2018-10-26 13:38:24 +01:00
|
|
|
RTE_LOG(DEBUG, POWER, "cure freq is %d, edpi is %"PRIu64"\n",
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
cur_freq,
|
|
|
|
|
poll_stats->thresh[cur_freq].base_edpi);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
percent = update_stats(poll_stats);
|
|
|
|
|
|
|
|
|
|
if (percent > 100) {
|
|
|
|
|
/* edpi mean empty poll counter difference per interval */
|
|
|
|
|
RTE_LOG(DEBUG, POWER, "Edpi is bigger than threshold\n");
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (poll_stats->cur_freq == LOW)
|
|
|
|
|
RTE_LOG(INFO, POWER, "Purge Mode is not currently supported\n");
|
|
|
|
|
else if (poll_stats->cur_freq == MED) {
|
|
|
|
|
|
|
|
|
|
if (percent >
|
|
|
|
|
poll_stats->thresh[MED].threshold_percent) {
|
|
|
|
|
|
|
|
|
|
if (poll_stats->threshold_ctr < INTERVALS_PER_SECOND)
|
|
|
|
|
poll_stats->threshold_ctr++;
|
|
|
|
|
else {
|
|
|
|
|
set_state(poll_stats, HGH_BUSY);
|
|
|
|
|
RTE_LOG(INFO, POWER, "MOVE to HGH\n");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
/* reset */
|
|
|
|
|
poll_stats->threshold_ctr = 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
} else if (poll_stats->cur_freq == HGH) {
|
|
|
|
|
|
|
|
|
|
if (percent <
|
|
|
|
|
poll_stats->thresh[HGH].threshold_percent) {
|
|
|
|
|
|
|
|
|
|
if (poll_stats->threshold_ctr < INTERVALS_PER_SECOND)
|
|
|
|
|
poll_stats->threshold_ctr++;
|
|
|
|
|
else {
|
|
|
|
|
set_state(poll_stats, MED_NORMAL);
|
|
|
|
|
RTE_LOG(INFO, POWER, "MOVE to MED\n");
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
/* reset */
|
|
|
|
|
poll_stats->threshold_ctr = 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
empty_poll_training(struct priority_worker *poll_stats,
|
|
|
|
|
uint32_t max_train_iter)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
if (poll_stats->iter_counter < INTERVALS_PER_SECOND) {
|
|
|
|
|
poll_stats->iter_counter++;
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
update_training_stats(poll_stats,
|
|
|
|
|
LOW,
|
|
|
|
|
false,
|
|
|
|
|
max_train_iter);
|
|
|
|
|
|
|
|
|
|
update_training_stats(poll_stats,
|
|
|
|
|
MED,
|
|
|
|
|
false,
|
|
|
|
|
max_train_iter);
|
|
|
|
|
|
|
|
|
|
update_training_stats(poll_stats,
|
|
|
|
|
HGH,
|
|
|
|
|
false,
|
|
|
|
|
max_train_iter);
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (poll_stats->thresh[LOW].trained == true
|
|
|
|
|
&& poll_stats->thresh[MED].trained == true
|
|
|
|
|
&& poll_stats->thresh[HGH].trained == true) {
|
|
|
|
|
|
|
|
|
|
set_state(poll_stats, MED_NORMAL);
|
|
|
|
|
|
2018-10-26 13:38:24 +01:00
|
|
|
RTE_LOG(INFO, POWER, "LOW threshold is %"PRIu64"\n",
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
poll_stats->thresh[LOW].base_edpi);
|
|
|
|
|
|
2018-10-26 13:38:24 +01:00
|
|
|
RTE_LOG(INFO, POWER, "MED threshold is %"PRIu64"\n",
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
poll_stats->thresh[MED].base_edpi);
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 13:38:24 +01:00
|
|
|
RTE_LOG(INFO, POWER, "HIGH threshold is %"PRIu64"\n",
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
poll_stats->thresh[HGH].base_edpi);
|
|
|
|
|
|
|
|
|
|
RTE_LOG(INFO, POWER, "Training is Complete for %d\n",
|
|
|
|
|
poll_stats->lcore_id);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_empty_poll_detection(struct rte_timer *tim, void *arg)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
uint32_t i;
|
|
|
|
|
|
|
|
|
|
struct priority_worker *poll_stats;
|
|
|
|
|
|
|
|
|
|
RTE_SET_USED(tim);
|
|
|
|
|
|
|
|
|
|
RTE_SET_USED(arg);
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < NUM_NODES; i++) {
|
|
|
|
|
|
|
|
|
|
poll_stats = &(ep_params->wrk_data.wrk_stats[i]);
|
|
|
|
|
|
|
|
|
|
if (rte_lcore_is_enabled(poll_stats->lcore_id) == 0)
|
|
|
|
|
continue;
|
|
|
|
|
|
|
|
|
|
switch (poll_stats->queue_state) {
|
|
|
|
|
case(TRAINING):
|
|
|
|
|
empty_poll_training(poll_stats,
|
|
|
|
|
ep_params->max_train_iter);
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case(HGH_BUSY):
|
|
|
|
|
case(MED_NORMAL):
|
|
|
|
|
update_stats_normal(poll_stats);
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
case(LOW_PURGE):
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
int
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_power_empty_poll_stat_init(struct ep_params **eptr, uint8_t *freq_tlb,
|
|
|
|
|
struct ep_policy *policy)
|
|
|
|
|
{
|
|
|
|
|
uint32_t i;
|
|
|
|
|
/* Allocate the ep_params structure */
|
|
|
|
|
ep_params = rte_zmalloc_socket(NULL,
|
|
|
|
|
sizeof(struct ep_params),
|
|
|
|
|
0,
|
|
|
|
|
rte_socket_id());
|
|
|
|
|
|
|
|
|
|
if (!ep_params)
|
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
|
|
if (freq_tlb == NULL) {
|
|
|
|
|
freq_index[LOW] = 14;
|
|
|
|
|
freq_index[MED] = 9;
|
|
|
|
|
freq_index[HGH] = 1;
|
|
|
|
|
} else {
|
|
|
|
|
freq_index[LOW] = freq_tlb[LOW];
|
|
|
|
|
freq_index[MED] = freq_tlb[MED];
|
|
|
|
|
freq_index[HGH] = freq_tlb[HGH];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
RTE_LOG(INFO, POWER, "Initialize the Empty Poll\n");
|
|
|
|
|
|
|
|
|
|
/* Train for pre-defined period */
|
|
|
|
|
ep_params->max_train_iter = INTERVALS_PER_SECOND * SECONDS_TO_TRAIN_FOR;
|
|
|
|
|
|
|
|
|
|
struct stats_data *w = &ep_params->wrk_data;
|
|
|
|
|
|
|
|
|
|
*eptr = ep_params;
|
|
|
|
|
|
|
|
|
|
/* initialize all wrk_stats state */
|
|
|
|
|
for (i = 0; i < NUM_NODES; i++) {
|
|
|
|
|
|
|
|
|
|
if (rte_lcore_is_enabled(i) == 0)
|
|
|
|
|
continue;
|
|
|
|
|
/*init the freqs table */
|
|
|
|
|
total_avail_freqs[i] = rte_power_freqs(i,
|
|
|
|
|
avail_freqs[i],
|
|
|
|
|
NUM_FREQS);
|
|
|
|
|
|
|
|
|
|
RTE_LOG(INFO, POWER, "total avail freq is %d , lcoreid %d\n",
|
|
|
|
|
total_avail_freqs[i],
|
|
|
|
|
i);
|
|
|
|
|
|
|
|
|
|
if (get_freq_index(LOW) > total_avail_freqs[i])
|
|
|
|
|
return -1;
|
|
|
|
|
|
2020-10-15 15:57:19 -07:00
|
|
|
if (rte_get_main_lcore() != i) {
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
w->wrk_stats[i].lcore_id = i;
|
|
|
|
|
set_policy(&w->wrk_stats[i], policy);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
void
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_power_empty_poll_stat_free(void)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
RTE_LOG(INFO, POWER, "Close the Empty Poll\n");
|
|
|
|
|
|
|
|
|
|
if (ep_params != NULL)
|
|
|
|
|
rte_free(ep_params);
|
|
|
|
|
}
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
int
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_power_empty_poll_stat_update(unsigned int lcore_id)
|
|
|
|
|
{
|
|
|
|
|
struct priority_worker *poll_stats;
|
|
|
|
|
|
|
|
|
|
if (lcore_id >= NUM_NODES)
|
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
|
|
poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
|
|
|
|
|
|
|
|
|
|
if (poll_stats->lcore_id == 0)
|
|
|
|
|
poll_stats->lcore_id = lcore_id;
|
|
|
|
|
|
|
|
|
|
poll_stats->empty_dequeues++;
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
int
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
struct priority_worker *poll_stats;
|
|
|
|
|
|
|
|
|
|
if (lcore_id >= NUM_NODES)
|
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
|
|
poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
|
|
|
|
|
|
|
|
|
|
if (poll_stats->lcore_id == 0)
|
|
|
|
|
poll_stats->lcore_id = lcore_id;
|
|
|
|
|
|
|
|
|
|
poll_stats->num_dequeue_pkts += nb_pkt;
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
uint64_t
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_power_empty_poll_stat_fetch(unsigned int lcore_id)
|
|
|
|
|
{
|
|
|
|
|
struct priority_worker *poll_stats;
|
|
|
|
|
|
|
|
|
|
if (lcore_id >= NUM_NODES)
|
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
|
|
poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
|
|
|
|
|
|
|
|
|
|
if (poll_stats->lcore_id == 0)
|
|
|
|
|
poll_stats->lcore_id = lcore_id;
|
|
|
|
|
|
|
|
|
|
return poll_stats->empty_dequeues;
|
|
|
|
|
}
|
|
|
|
|
|
2019-06-29 13:58:52 +02:00
|
|
|
uint64_t
|
power: add traffic pattern aware power control
1. Abstract
For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:
* No indication of overload conditions.
* User does not know how much real load is on a system, resulting
in wasted energy as no power management is utilized.
Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.
When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.
2. Proposed solution
The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.
In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.
2.1 Power state definition:
LOW: Not currently used, reserved for future use.
MED: the frequency is used to process modest traffic workload.
HIGH: the frequency is used to process busy traffic workload.
2.2 There are two phases to establish the power management system:
a.Initialization/Training phase. The training phase is necessary
in order to figure out the system polling baseline numbers from
idle to busy. The highest poll count will be during idle, where
all polls are empty. These poll counts will be different between
systems due to the many possible processor micro-arch, cache
and device configurations, hence the training phase.
In the training phase, traffic is blocked so the training
algorithm can average the empty-poll numbers for the LOW, MED and
HIGH power states in order to create a baseline.
The core's counter are collected every 10ms, and the Training
phase will take 2 seconds.
Training is disabled as default configuration. The default
parameter is applied. Sample App still can trigger training
if that's needed. Once the training phase has been executed once on
a system, the application can then be started with the relevant
thresholds provided on the command line, allowing the application
to start passing start traffic immediately
b.Normal phase. Traffic starts immediately based on the default
thresholds, or based on the user supplied thresholds via the
command line parameters. The run-time poll counts are compared with
the baseline and the decision will be taken to move to MED power
state or HIGH power state. The counters are calculated every 10ms.
3. Proposed API
1. rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
2. rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
3. rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
4. rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
5. rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
6. rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.
7. rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
2018-10-19 12:07:18 +01:00
|
|
|
rte_power_poll_stat_fetch(unsigned int lcore_id)
|
|
|
|
|
{
|
|
|
|
|
struct priority_worker *poll_stats;
|
|
|
|
|
|
|
|
|
|
if (lcore_id >= NUM_NODES)
|
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
|
|
poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
|
|
|
|
|
|
|
|
|
|
if (poll_stats->lcore_id == 0)
|
|
|
|
|
poll_stats->lcore_id = lcore_id;
|
|
|
|
|
|
|
|
|
|
return poll_stats->num_dequeue_pkts;
|
|
|
|
|
}
|
