d/numam-dpdk
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

examples/l3fwd: fix exact match performance

Browse Source 
It seems that for the most use cases, previous hash_multi_lookup provides
better performance, and more, sequential lookup can cause significant
performance drop.

This patch sets previously optional hash_multi_lookup method as default.
It also provides some minor optimizations such as queue drain only on used
tx ports.

Fixes: 94c54b4158 ("examples/l3fwd: rework exact-match")
Fixes: dc81ebbaca ("lpm: extend IPv4 next hop field")
Fixes: 64d3955de1 ("examples/l3fwd: fix ARM build")

Reported-by: Qian Xu <qian.q.xu@intel.com>
Signed-off-by: Tomasz Kulasek <tomaszx.kulasek@intel.com>
This commit is contained in:
Tomasz Kulasek 2016-03-18 14:31:46 +01:00 committed by Thomas Monjalon
parent fdefa0a932
commit 52c97adc1f
6 changed files with 38 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
examples/l3fwd/l3fwd.h
View File

@ -40,6 +40,10 @@
#define RTE_LOGTYPE_L3FWD RTE_LOGTYPE_USER1
#if !defined(NO_HASH_MULTI_LOOKUP) && defined(RTE_MACHINE_CPUFLAG_NEON)
#define NO_HASH_MULTI_LOOKUP 1
#endif
#define MAX_PKT_BURST 32
#define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */
@ -86,6 +90,8 @@ struct lcore_rx_queue {
struct lcore_conf {
uint16_t n_rx_queue;
struct lcore_rx_queue rx_queue_list[MAX_RX_QUEUE_PER_LCORE];
uint16_t n_tx_port;
uint16_t tx_port_id[RTE_MAX_ETHPORTS];
uint16_t tx_queue_id[RTE_MAX_ETHPORTS];
struct mbuf_table tx_mbufs[RTE_MAX_ETHPORTS];
void *ipv4_lookup_struct;

8
examples/l3fwd/l3fwd_em.c
View File

@ -250,7 +250,7 @@ em_mask_key(void *key, xmm_t mask)
return _mm_and_si128(data, mask);
}
#elif defined(__ARM_NEON)
#elif defined(RTE_MACHINE_CPUFLAG_NEON)
static inline xmm_t
em_mask_key(void *key, xmm_t mask)
{
@ -320,7 +320,7 @@ em_get_ipv6_dst_port(void *ipv6_hdr, uint8_t portid, void *lookup_struct)
* buffer optimization i.e. ENABLE_MULTI_BUFFER_OPTIMIZE=1.
*/
#if defined(__SSE4_1__)
#ifndef HASH_MULTI_LOOKUP
#if defined(NO_HASH_MULTI_LOOKUP)
#include "l3fwd_em_sse.h"
#else
#include "l3fwd_em_hlm_sse.h"
@ -568,8 +568,8 @@ em_main_loop(__attribute__((unused)) void *dummy)
diff_tsc = cur_tsc - prev_tsc;
if (unlikely(diff_tsc > drain_tsc)) {
for (i = 0; i < qconf->n_rx_queue; i++) {
portid = qconf->rx_queue_list[i].port_id;
for (i = 0; i < qconf->n_tx_port; ++i) {
portid = qconf->tx_port_id[i];
if (qconf->tx_mbufs[portid].len == 0)
continue;
send_burst(qconf,

28
examples/l3fwd/l3fwd_em_hlm_sse.h
View File

@ -34,17 +34,9 @@
#ifndef __L3FWD_EM_HLM_SSE_H__
#define __L3FWD_EM_HLM_SSE_H__
/**
* @file
* This is an optional implementation of packet classification in Exact-Match
* path using rte_hash_lookup_multi method from previous implementation.
* While sequential classification seems to be faster, it's disabled by default
* and can be enabled with HASH_LOOKUP_MULTI global define in compilation time.
*/
#include "l3fwd_sse.h"
static inline void
static inline __attribute__((always_inline)) void
em_get_dst_port_ipv4x8(struct lcore_conf *qconf, struct rte_mbuf *m[8],
uint8_t portid, uint32_t dst_port[8])
{
@ -168,7 +160,7 @@ get_ipv6_5tuple(struct rte_mbuf *m0, __m128i mask0,
key->xmm[2] = _mm_and_si128(tmpdata2, mask1);
}
static inline void
static inline __attribute__((always_inline)) void
em_get_dst_port_ipv6x8(struct lcore_conf *qconf, struct rte_mbuf *m[8],
uint8_t portid, uint32_t dst_port[8])
{
@ -322,17 +314,17 @@ l3fwd_em_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
} else {
dst_port[j] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+1] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+2] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+3] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+4] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+5] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+6] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+7] = em_get_dst_port(qconf, pkts_burst[j], portid);
dst_port[j+1] = em_get_dst_port(qconf, pkts_burst[j+1], portid);
dst_port[j+2] = em_get_dst_port(qconf, pkts_burst[j+2], portid);
dst_port[j+3] = em_get_dst_port(qconf, pkts_burst[j+3], portid);
dst_port[j+4] = em_get_dst_port(qconf, pkts_burst[j+4], portid);
dst_port[j+5] = em_get_dst_port(qconf, pkts_burst[j+5], portid);
dst_port[j+6] = em_get_dst_port(qconf, pkts_burst[j+6], portid);
dst_port[j+7] = em_get_dst_port(qconf, pkts_burst[j+7], portid);
}
}
for (; j < n; j++)
for (; j < nb_rx; j++)
dst_port[j] = em_get_dst_port(qconf, pkts_burst[j], portid);
send_packets_multi(qconf, pkts_burst, dst_port, nb_rx);

9
examples/l3fwd/l3fwd_em_sse.h
View File

@ -34,6 +34,15 @@
#ifndef __L3FWD_EM_SSE_H__
#define __L3FWD_EM_SSE_H__
/**
* @file
* This is an optional implementation of packet classification in Exact-Match
* path using sequential packet classification method.
* While hash lookup multi seems to provide better performance, it's disabled
* by default and can be enabled with NO_HASH_LOOKUP_MULTI global define in
* compilation time.
*/
#include "l3fwd_sse.h"
static inline __attribute__((always_inline)) uint16_t

4
examples/l3fwd/l3fwd_lpm.c
View File

@ -159,8 +159,8 @@ lpm_main_loop(__attribute__((unused)) void *dummy)
diff_tsc = cur_tsc - prev_tsc;
if (unlikely(diff_tsc > drain_tsc)) {
for (i = 0; i < qconf->n_rx_queue; i++) {
portid = qconf->rx_queue_list[i].port_id;
for (i = 0; i < qconf->n_tx_port; ++i) {
portid = qconf->tx_port_id[i];
if (qconf->tx_mbufs[portid].len == 0)
continue;
send_burst(qconf,

7
examples/l3fwd/main.c
View File

@ -791,6 +791,7 @@ main(int argc, char **argv)
unsigned lcore_id;
uint32_t n_tx_queue, nb_lcores;
uint8_t portid, nb_rx_queue, queue, socketid;
uint8_t nb_tx_port;
/* init EAL */
ret = rte_eal_init(argc, argv);
@ -830,6 +831,7 @@ main(int argc, char **argv)
rte_exit(EXIT_FAILURE, "check_port_config failed\n");
nb_lcores = rte_lcore_count();
nb_tx_port = 0;
/* Setup function pointers for lookup method. */
setup_l3fwd_lookup_tables();
@ -906,8 +908,13 @@ main(int argc, char **argv)
qconf = &lcore_conf[lcore_id];
qconf->tx_queue_id[portid] = queueid;
queueid++;
qconf->n_tx_port = nb_tx_port;
qconf->tx_port_id[qconf->n_tx_port] = portid;
}
printf("\n");
nb_tx_port++;
}
for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {