This modification allows for the branch ratio threshold to be set
per core rather than system wide. This gives greater flexibility to
the branch ratio monitoring allowing it to manage different
workloads with different characteristics on the same system.
Signed-off-by: Rory Sexton <rory.sexton@intel.com>
Reviewed-by: David Hunt <david.hunt@intel.com>
Acked-by: Reshma Pattan <reshma.pattan@intel.com>
The Unix socket path may be as long as UNIX_PATH_MAX.
This constant is supposed to be defined in sys/un.h.
On Linux, it appears to be in linux/un.h.
This constant was re-defined locally, based on a variable declaration.
It is breaking compilation with -fno-common (default in GCC 10)
We could avoid the variable declaration by using NULL struct,
but it looks simpler not redefining this system constant.
As the power library and its examples are restricted to Linux only,
the Linux header file is directly included.
Fixes: 0d74597c1b4f ("examples/vm_power: fix max length of unix socket path")
Cc: stable@dpdk.org
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: David Marchand <david.marchand@redhat.com>
Acked-by: Kevin Traynor <ktraynor@redhat.com>
The branch ratio algorithm in the vm_power_manager sample application
can be very sensitive at patricular loads in a workload, causing
oscillations between min and max frequency. For example, if a
workload is at 50%, scaling up may change the ratio
enough that it immediately thinks it needs to scale down again.
This patch introduces a sliding window recording the scale up/down
direction for the last 32 samples, and scales up if any samples indicate
we should scale up, otherwise scale down. Each core has it's own window.
Fixes: 4b1a631b8a8a ("examples/vm_power: add oob monitoring functions")
Cc: stable@dpdk.org
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
The freqs array in freq_info struct has RTE_MAX_LCORE_FREQS elements,
yet the code can attempt to look at the index at RTE_MAX_LCORE,
which may be greater than RTE_MAX_LCORE_FREQS. Fix to limit index to
RTE_MAX_LCORE_FREQS.
Coverity issue: 337660
Fixes: d26c18c93260 ("examples/vm_power: cpu frequency in host")
Cc: stable@dpdk.org
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Reshma Pattan <reshma.pattan@intel.com>
A previous change removed the limit of 64 cores by
moving away from 64-bit masks to char arrays. However
this left a buffer overrun issue, where the max channels
was defined as 64, and max cores was defined as 256. These
should all be consistently set to RTE_MAX_LCORE.
The #defines being removed are CHANNEL_CMDS_MAX_CPUS,
CHANNEL_CMDS_MAX_CHANNELS, POWER_MGR_MAX_CPUS, and
CHANNEL_CMDS_MAX_VM_CHANNELS, and are being replaced
with RTE_MAX_LCORE for consistency and simplicity.
Coverity issue: 337672, 337673, 337678
Fixes: fd73630e95c1 ("examples/power: change 64-bit masks to arrays")
Cc: stable@dpdk.org
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
The vm_power_manager starts by setting the environment to acpi
using rte_power_set_env(PM_ENV_ACPI_CPUFREQ). This causes a problem
starting vm_power_manager when the system is using the intel_pstate
driver. The env should be set to none, or not called at all, because
the library now auto-detects the environment to be either acpi or
intel_pstate. This patch sets the environment to none so that the
library can successfully auto-detect.
Fixes: e6c6dc0f96c8 ("power: add p-state driver compatibility")
Signed-off-by: David Hunt <david.hunt@intel.com>
Since we're moving to allowing greater than 64 cores, the mask functions
that use uint64_t to perform functions on a masked set of cores are no
longer needed, so removing them.
Signed-off-by: David Hunt <david.hunt@intel.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
The vm_power_manager app was not respecting the POWER_MGR_MAX_CPUS
during initialisation, so if there were more CPUs than this value (64),
it would lead to buffer overruns of there were more then 64 cores in
the system.
Added in a check during init and un-init to only initialise up to
lcore_id 63.
This raises the question as to why not simply increase the value of
POWER_MGR_MAX_CPUS. Well, it's not that simple, as many of the APIs take
a uint64_t as a parameter for the core mask, and this will not work for
cores greater than 63. So some work needs to be done in the future to
remove this limitation. For now we'll fix the memory corruption.
Also, the patch that this fixes says "allow greater than 64 cores" but
that's not across the entire application, it's only for the out-of-band
monitoring. I'll add a notice for an API change in the next release to
clean this up, i.e. depricate any API calls that use masks.
Fixes: 6453b9284b64 ("examples/vm_power: allow greater than 64 cores")
Cc: stable@dpdk.org
Signed-off-by: David Hunt <david.hunt@intel.com>
For different workloads and poll loops, the theshold
may be different for when you want to scale up and down.
This patch allows changing of the default branch ratio
by using the -b command line argument (or --branch-ratio=)
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Radu Nicolau <radu.nicolau@intel.com>
To facilitate more info per core, change the global_cpu_mask
from a uint64_t to an array. This also removes the limit on
64 cores, allocing the aray at run-time based on the number of
cores found in the system.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Radu Nicolau <radu.nicolau@intel.com>
Add in the '-l' command line parameter (also --core-list)
So the user can now pass --corelist=4,6,8-10 and it will
expand out to 4,6,8,9,10 using the parse function provided
in parse.c (parse_set).
This list of cores is then used to enable out-of-band monitoring
to scale up and down these cores based on the ratio of branch
hits versus branch misses. The ratio will be low when a poll
loop is spinning with no packets being received, so the frequency
will be scaled down.
Also , as part of this change, we introduce a core_info struct
which keeps information on each core in the system, and whether
we're doing out of band monitoring on them.
Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Radu Nicolau <radu.nicolau@intel.com>
Replace the BSD license header with the SPDX tag for files
with only an Intel copyright on them.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Add extra commands to command line to allow enable/disable of
per-core turbo.
When a core has turbo enabled, calling for max frequency will allow it to
go to a turbo frequency (P0n).
When a core has turbo disabled, calling for max frequency will allow it to
go to the maximum non-turbo frequency (P1), but not beyond.
Signed-off-by: David Hunt <david.hunt@intel.com>
The file rte_config.h is automatically generated and included.
No need to #include it.
The example performance-thread needs a makefile fix to avoid
overwriting the default cflags.
Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
A wrapper around librte_power(using ACPI cpufreq), providing locking around the
non-threadsafe library, allowing for frequency changes based on core masks and
core numbers from both the CLI thread and epoll monitor thread.
Signed-off-by: Alan Carew <alan.carew@intel.com>
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
