Skip to contentSkip to conversation

l3fwd improvements

PATCH Superseded Ruifeng Wang 7 participants 28 messages Mar 18, 2021 – Jul 5, 2021

lore
Share & tools
lore
#0 lore raw

cover l3fwd improvements

This series of patches include changes to l3fwd example application. Some improvements are made for better usage of CPU cycles and memory.

Ruifeng Wang (4):
examples/l3fwd: tune prefetch for better performance examples/l3fwd: eliminate unnecessary calculations examples/l3fwd: eliminate unnecessary reloads in loop examples/l3fwd: make data struct to be memory efficient

 examples/l3fwd/l3fwd.h          | 12 ++++++------
 examples/l3fwd/l3fwd_common.h   |  4 ++--
 examples/l3fwd/l3fwd_em.c       |  6 +++---
 examples/l3fwd/l3fwd_lpm.c      | 16 +++++++++-------
 examples/l3fwd/l3fwd_lpm_neon.h | 20 ++++++++++----------
 5 files changed, 30 insertions(+), 28 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/4 examples/l3fwd: tune prefetch for better performance

Packet header is prefetched before packet processing for better memory access performance. As L2 header will be updated by l3fwd, using of prefetch for store hint will set cache line to proper status and reduce cache maintenance overhead.

With this change, 12.9% performance uplift was measured on N1SDP platform with MLX5 NIC.

---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_lpm_neon.h +5 -5
index d6c0ba64a..ae8840694 100644
@@ -97,13 +97,13 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 
 	if (k) {
 		for (i = 0; i < FWDSTEP; i++) {
-			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
+			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[i],
 						struct rte_ether_hdr *) + 1);
 		}
 
 		for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
 			for (i = 0; i < FWDSTEP; i++) {
-				rte_prefetch0(rte_pktmbuf_mtod(
+				rte_prefetch0_write(rte_pktmbuf_mtod(
 						pkts_burst[j + i + FWDSTEP],
 						struct rte_ether_hdr *) + 1);
 			}
@@ -124,17 +124,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 		/* Prefetch last up to 3 packets one by one */
 		switch (m) {
 		case 3:
-			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
 						struct rte_ether_hdr *) + 1);
 			j++;
 			/* fallthrough */
 		case 2:
-			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
 						struct rte_ether_hdr *) + 1);
 			j++;
 			/* fallthrough */
 		case 1:
-			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
 						struct rte_ether_hdr *) + 1);
 			j++;
 		}
Signature · 1 line
2.25.1
#2 lore raw

Re: [dpdk-dev] [PATCH 1/4] examples/l3fwd: tune prefetch for better performance

Quoted 5 lines · Ruifeng Wang

Packet header is prefetched before packet processing for better memory access performance. As L2 header will be updated by l3fwd, using of prefetch for store hint will set cache line to proper status and reduce cache maintenance overhead.

The code does read the cache line too. Right?

Quoted 8 lines

With this change, 12.9% performance uplift was measured on N1SDP platform with MLX5 NIC.

On the octeontx2 platform, It is 2% regression.

Looks like micro architecture-specific item of handing write hint on the memory the area that does read and write.

I am testing the LPM lookup miss case.

My test command:
./build/examples/dpdk-l3fwd -c 0x0100 -- -p 0x1 --config="(0,0,8)" -P

Quoted 48 lines
---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_lpm_neon.h +5 -6
index d6c0ba64a..ae8840694 100644
@@ -97,13 +97,13 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 
        if (k) {
                for (i = 0; i < FWDSTEP; i++) {
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
+                       rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[i],
                                                struct rte_ether_hdr *) + 1);
                }
 
                for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
                        for (i = 0; i < FWDSTEP; i++) {
-                               rte_prefetch0(rte_pktmbuf_mtod(
+                               rte_prefetch0_write(rte_pktmbuf_mtod(
                                                pkts_burst[j + i + FWDSTEP],
                                                struct rte_ether_hdr *) + 1);
                        }
@@ -124,17 +124,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
                /* Prefetch last up to 3 packets one by one */
                switch (m) {
                case 3:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+                       rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                        /* fallthrough */
                case 2:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+                       rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                        /* fallthrough */
                case 1:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+                       rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                }
--

2.25.1

#3 lore raw

Re: [dpdk-dev] [PATCH 1/4] examples/l3fwd: tune prefetch for better performance

<snip>

Quoted 14 lines
Quoted 5 lines · Ruifeng Wang

Packet header is prefetched before packet processing for better memory access performance. As L2 header will be updated by l3fwd, using of prefetch for store hint will set cache line to proper status and reduce cache maintenance overhead.

The code does read the cache line too. Right?

Quoted 4 lines

With this change, 12.9% performance uplift was measured on N1SDP platform with MLX5 NIC.

Quoted 7 lines
Quoted 4 lines

On the octeontx2 platform, It is 2% regression.

Thanks Jerin for testing.
It would be good to know the results from others with A72 platforms.

Quoted 66 lines

Looks like micro architecture-specific item of handing write hint on the memory the area that does read and write.

I am testing the LPM lookup miss case.

My test command:
./build/examples/dpdk-l3fwd -c 0x0100 -- -p 0x1 --config="(0,0,8)" -P

examples/l3fwd/l3fwd_lpm_neon.h l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf

Quoted 27 lines
---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
(quoted diff) +0 -0

b/examples/l3fwd/l3fwd_lpm_neon.h index d6c0ba64a..ae8840694 100644

examples/l3fwd/l3fwd_lpm_neon.h +3 -2
@@ -97,13 +97,13 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf
 **pkts_burst,
 
        if (k) {
                for (i = 0; i < FWDSTEP; i++) {
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[i],
                                                struct rte_ether_hdr *) + 1);
                }
 
                for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
                        for (i = 0; i < FWDSTEP; i++) {
-                               rte_prefetch0(rte_pktmbuf_mtod(
+                               rte_prefetch0_write(rte_pktmbuf_mtod(
                                                pkts_burst[j + i + FWDSTEP],
                                                struct rte_ether_hdr *) + 1);
                        }
@@ -124,17 +124,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf

**pkts_burst,

Quoted 26 lines
examples/l3fwd/l3fwd_lpm_neon.h +6 -4
                /* Prefetch last up to 3 packets one by one */
                switch (m) {
                case 3:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                        /* fallthrough */
                case 2:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                        /* fallthrough */
                case 1:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                }
--

2.25.1

#4 lore raw

Re: [dpdk-dev] [PATCH 1/4] examples/l3fwd: tune prefetch for better performance

In reply to Jerin Jacob
Quoted 14 lines

-----Original Message-----
From: Jerin Jacob <[email protected]>
Sent: Wednesday, April 14, 2021 2:51 AM
To: Ruifeng Wang <[email protected]>
Cc: [email protected]; [email protected]; Ferruh Yigit <[email protected]>; [email protected]; David Marchand <[email protected]>; dpdk-dev <[email protected]>; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>
Subject: Re: [dpdk-dev] [PATCH 1/4] examples/l3fwd: tune prefetch for better performance

Quoted 1 line · Ruifeng Wang
Quoted 7 lines
Quoted 4 lines

Packet header is prefetched before packet processing for better memory access performance. As L2 header will be updated by l3fwd, using of prefetch for store hint will set cache line to proper status and reduce cache maintenance overhead.

The code does read the cache line too. Right?

Yes, the code also read the cache line.
And prefetch to write helps writes. It saves snooping cost.

Quoted 6 lines
Quoted 6 lines

With this change, 12.9% performance uplift was measured on N1SDP platform with MLX5 NIC.

Quoted 9 lines
Quoted 2 lines

On the octeontx2 platform, It is 2% regression.

Looks like micro architecture-specific item of handing write hint on the memory the area that does read and write.

OK. Performance impact of the write hint may be different on various micro architecture implementations.
The 12% measured on N1 is not a small enough number to ignore. How about using a flag to distinguish prefetches invoked on different SoCs? A compile time flag like RTE_USE_PREFETCH_WRITE is introduced. And SoCs enable it based on need.

Quoted 63 lines

I am testing the LPM lookup miss case.

My test command:
./build/examples/dpdk-l3fwd -c 0x0100 -- -p 0x1 --config="(0,0,8)" -P

examples/l3fwd/l3fwd_lpm_neon.h l3fwd_lpm_send_packets(int nb_rx, struct

Quoted 9 lines
---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
(quoted diff) +0 -0

b/examples/l3fwd/l3fwd_lpm_neon.h index d6c0ba64a..ae8840694 100644

examples/l3fwd/l3fwd_lpm_neon.h +0 -0
@@ -97,13 +97,13 @@ l3fwd_lpm_send_packets(int nb_rx, struct

rte_mbuf

examples/l3fwd/l3fwd_lpm_neon.h l3fwd_lpm_send_packets(int nb_rx, struct

Quoted 18 lines
examples/l3fwd/l3fwd_lpm_neon.h +3 -2
 **pkts_burst,
 
        if (k) {
                for (i = 0; i < FWDSTEP; i++) {
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[i],
                                                struct rte_ether_hdr *) + 1);
                }
 
                for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
                        for (i = 0; i < FWDSTEP; i++) {
-                               rte_prefetch0(rte_pktmbuf_mtod(
+                               rte_prefetch0_write(rte_pktmbuf_mtod(
                                                pkts_burst[j + i + FWDSTEP],
                                                struct rte_ether_hdr *) + 1);
                        }
@@ -124,17 +124,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct

rte_mbuf **pkts_burst,

Quoted 26 lines
examples/l3fwd/l3fwd_lpm_neon.h +6 -4
                /* Prefetch last up to 3 packets one by one */
                switch (m) {
                case 3:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                        /* fallthrough */
                case 2:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                        /* fallthrough */
                case 1:
-                       rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
+
+ rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
                                                struct rte_ether_hdr *) + 1);
                        j++;
                }
--

2.25.1

#5 lore raw

PATCH 2/4 examples/l3fwd: eliminate unnecessary calculations

In reply to Ruifeng Wang

Both L2 and L3 headers will be used in forward processing. And these two headers are in the same cache line. It has the same effect for prefetching with L2 header address and prefetching with L3 header address.

Changed to use L2 header address for prefetching. The change showed no measurable performance improvement, but it definitely removed unnecessary instructions for address calculation.

---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_lpm_neon.h +5 -5
index ae8840694..1650ae444 100644
@@ -98,14 +98,14 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 	if (k) {
 		for (i = 0; i < FWDSTEP; i++) {
 			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[i],
-						struct rte_ether_hdr *) + 1);
+							void *));
 		}
 
 		for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
 			for (i = 0; i < FWDSTEP; i++) {
 				rte_prefetch0_write(rte_pktmbuf_mtod(
 						pkts_burst[j + i + FWDSTEP],
-						struct rte_ether_hdr *) + 1);
+						void *));
 			}
 
 			processx4_step1(&pkts_burst[j], &dip, &ipv4_flag);
@@ -125,17 +125,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 		switch (m) {
 		case 3:
 			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 			/* fallthrough */
 		case 2:
 			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 			/* fallthrough */
 		case 1:
 			rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 		}
 
Signature · 1 line
2.25.1
Acked-by on v1 2/4
#6 lore raw

Re: [dpdk-dev] [PATCH 2/4] examples/l3fwd: eliminate unnecessary calculations

Quoted 8 lines · Ruifeng Wang

Both L2 and L3 headers will be used in forward processing. And these two headers are in the same cache line. It has the same effect for prefetching with L2 header address and prefetching with L3 header address.

Changed to use L2 header address for prefetching. The change showed no measurable performance improvement, but it definitely removed

Same here.

Quoted 5 lines

unnecessary instructions for address calculation.

Quoted 49 lines
---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_lpm_neon.h +5 -6
index ae8840694..1650ae444 100644
@@ -98,14 +98,14 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
        if (k) {
                for (i = 0; i < FWDSTEP; i++) {
                        rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[i],
-                                               struct rte_ether_hdr *) + 1);
+                                                       void *));
                }
 
                for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
                        for (i = 0; i < FWDSTEP; i++) {
                                rte_prefetch0_write(rte_pktmbuf_mtod(
                                                pkts_burst[j + i + FWDSTEP],
-                                               struct rte_ether_hdr *) + 1);
+                                               void *));
                        }
 
                        processx4_step1(&pkts_burst[j], &dip, &ipv4_flag);
@@ -125,17 +125,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
                switch (m) {
                case 3:
                        rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
-                                               struct rte_ether_hdr *) + 1);
+                                                       void *));
                        j++;
                        /* fallthrough */
                case 2:
                        rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
-                                               struct rte_ether_hdr *) + 1);
+                                                       void *));
                        j++;
                        /* fallthrough */
                case 1:
                        rte_prefetch0_write(rte_pktmbuf_mtod(pkts_burst[j],
-                                               struct rte_ether_hdr *) + 1);
+                                                       void *));
                        j++;
                }
 
--

2.25.1

#7 lore raw

PATCH 3/4 examples/l3fwd: eliminate unnecessary reloads in loop

In reply to Ruifeng Wang

Number of rx queue and number of rx port in lcore config are constants during the period of l3 forward application running. But compiler has no this information.

Copied values from lcore config to local variables and used the local variables for iteration. Compiler can see that the local variables are not changed, so qconf reloads at each iteration can be eliminated.

The change showed 1.8% performance uplift in single core, single port, single queue test on N1SDP platform with MLX5 NIC.

---
 examples/l3fwd/l3fwd_lpm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
examples/l3fwd/l3fwd_lpm.c +6 -4
index 3dcf1fef1..d338590b9 100644
@@ -190,14 +190,16 @@ lpm_main_loop(__rte_unused void *dummy)
 	lcore_id = rte_lcore_id();
 	qconf = &lcore_conf[lcore_id];
 
-	if (qconf->n_rx_queue == 0) {
+	uint16_t n_rx_q = qconf->n_rx_queue;
+	uint16_t n_tx_p = qconf->n_tx_port;
+	if (n_rx_q == 0) {
 		RTE_LOG(INFO, L3FWD, "lcore %u has nothing to do\n", lcore_id);
 		return 0;
 	}
 
 	RTE_LOG(INFO, L3FWD, "entering main loop on lcore %u\n", lcore_id);
 
-	for (i = 0; i < qconf->n_rx_queue; i++) {
+	for (i = 0; i < n_rx_q; i++) {
 
 		portid = qconf->rx_queue_list[i].port_id;
 		queueid = qconf->rx_queue_list[i].queue_id;
@@ -216,7 +218,7 @@ lpm_main_loop(__rte_unused void *dummy)
 		diff_tsc = cur_tsc - prev_tsc;
 		if (unlikely(diff_tsc > drain_tsc)) {
 
-			for (i = 0; i < qconf->n_tx_port; ++i) {
+			for (i = 0; i < n_tx_p; ++i) {
 				portid = qconf->tx_port_id[i];
 				if (qconf->tx_mbufs[portid].len == 0)
 					continue;
@@ -232,7 +234,7 @@ lpm_main_loop(__rte_unused void *dummy)
 		/*
 		 * Read packet from RX queues
 		 */
-		for (i = 0; i < qconf->n_rx_queue; ++i) {
+		for (i = 0; i < n_rx_q; ++i) {
 			portid = qconf->rx_queue_list[i].port_id;
 			queueid = qconf->rx_queue_list[i].queue_id;
 			nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
Signature · 1 line
2.25.1
#8 lore raw

Re: [dpdk-dev] [PATCH 3/4] examples/l3fwd: eliminate unnecessary reloads in loop

Quoted 11 lines · Ruifeng Wang

Number of rx queue and number of rx port in lcore config are constants during the period of l3 forward application running. But compiler has no this information.

Copied values from lcore config to local variables and used the local variables for iteration. Compiler can see that the local variables are not changed, so qconf reloads at each iteration can be eliminated.

The change showed 1.8% performance uplift in single core, single port, single queue test on N1SDP platform with MLX5 NIC.

At least, in octeontx2, I dont see any performance improvement. But change looks good. Please find below a comment.

Quoted 11 lines
---
 examples/l3fwd/l3fwd_lpm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
examples/l3fwd/l3fwd_lpm.c +0 -0
index 3dcf1fef1..d338590b9 100644

examples/l3fwd/l3fwd_lpm.c lpm_main_loop(__rte_unused void *dummy) · line 194

Quoted 6 lines
examples/l3fwd/l3fwd_lpm.c +2 -1
@@ -190,14 +190,16 @@ lpm_main_loop(__rte_unused void *dummy)
        lcore_id = rte_lcore_id();
        qconf = &lcore_conf[lcore_id];
 
-       if (qconf->n_rx_queue == 0) {
+       uint16_t n_rx_q = qconf->n_rx_queue;
+       uint16_t n_tx_p = qconf->n_tx_port;

How about adding const?

#9 lore raw

Re: [dpdk-dev] [PATCH 3/4] examples/l3fwd: eliminate unnecessary reloads in loop

Quoted 42 lines

-----Original Message-----
From: Jerin Jacob <[email protected]>
Sent: Wednesday, April 14, 2021 1:43 AM
To: Ruifeng Wang <[email protected]>
Cc: [email protected]; [email protected]; Ferruh Yigit <[email protected]>; [email protected]; David Marchand <[email protected]>; dpdk-dev <[email protected]>; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>
Subject: Re: [dpdk-dev] [PATCH 3/4] examples/l3fwd: eliminate unnecessary reloads in loop

Quoted 11 lines · Ruifeng Wang

Number of rx queue and number of rx port in lcore config are constants during the period of l3 forward application running. But compiler has no this information.

Copied values from lcore config to local variables and used the local variables for iteration. Compiler can see that the local variables are not changed, so qconf reloads at each iteration can be eliminated.

The change showed 1.8% performance uplift in single core, single port, single queue test on N1SDP platform with MLX5 NIC.

At least, in octeontx2, I dont see any performance improvement. But change looks good. Please find below a comment.

Quoted 14 lines
---
 examples/l3fwd/l3fwd_lpm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
examples/l3fwd/l3fwd_lpm.c +0 -0
index 3dcf1fef1..d338590b9 100644
@@ -190,14 +190,16 @@ lpm_main_loop(__rte_unused void *dummy)
        lcore_id = rte_lcore_id();
        qconf = &lcore_conf[lcore_id];
 
Quoted 5 lines

examples/l3fwd/l3fwd_lpm.c lpm_main_loop(__rte_unused void *dummy) · line 194

Quoted 3 lines
examples/l3fwd/l3fwd_lpm.c +2 -1
-       if (qconf->n_rx_queue == 0) {
+       uint16_t n_rx_q = qconf->n_rx_queue;
+       uint16_t n_tx_p = qconf->n_tx_port;

How about adding const?

Ack. The values are not expected to be changed.
Will update in next version.

#10 lore raw

PATCH 4/4 examples/l3fwd: make data struct to be memory efficient

In reply to Ruifeng Wang

There are some holes in data struct lcore_conf. The holes are due to alignment requirement.

For struct lcore_rx_queue, there is no need to make every element of this type to be cache line aligned, because the data is not shared between cores.

Member len of struct mbuf_table can be moved out. So data can be packed and there will be no need to load an extra cache line when mbuf table is empty.

The change showed slight performance improvement on N1SDP platform.

---
 examples/l3fwd/l3fwd.h        | 12 ++++++------
 examples/l3fwd/l3fwd_common.h |  4 ++--
 examples/l3fwd/l3fwd_em.c     |  6 +++---
 examples/l3fwd/l3fwd_lpm.c    |  6 +++---
 4 files changed, 14 insertions(+), 14 deletions(-)
examples/l3fwd/l3fwd.h +6 -6
index 2cf06099e..f3a301e12 100644
@@ -57,22 +57,22 @@
 #define HASH_ENTRY_NUMBER_DEFAULT	4
 
 struct mbuf_table {
-	uint16_t len;
 	struct rte_mbuf *m_table[MAX_PKT_BURST];
 };
 
 struct lcore_rx_queue {
 	uint16_t port_id;
 	uint8_t queue_id;
-} __rte_cache_aligned;
+};
 
 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];
+	uint16_t tx_mbuf_len[RTE_MAX_ETHPORTS];
 	struct mbuf_table tx_mbufs[RTE_MAX_ETHPORTS];
+	uint16_t n_rx_queue;
+	uint16_t n_tx_port;
 	void *ipv4_lookup_struct;
 	void *ipv6_lookup_struct;
 } __rte_cache_aligned;
@@ -122,7 +122,7 @@ send_single_packet(struct lcore_conf *qconf,
 {
 	uint16_t len;
 
-	len = qconf->tx_mbufs[port].len;
+	len = qconf->tx_mbuf_len[port];
 	qconf->tx_mbufs[port].m_table[len] = m;
 	len++;
 
@@ -132,7 +132,7 @@ send_single_packet(struct lcore_conf *qconf,
 		len = 0;
 	}
 
-	qconf->tx_mbufs[port].len = len;
+	qconf->tx_mbuf_len[port] = len;
 	return 0;
 }
 
examples/l3fwd/l3fwd_common.h +2 -2
index 7d83ff641..05e03dbfc 100644
@@ -183,7 +183,7 @@ send_packetsx4(struct lcore_conf *qconf, uint16_t port, struct rte_mbuf *m[],
 {
 	uint32_t len, j, n;
 
-	len = qconf->tx_mbufs[port].len;
+	len = qconf->tx_mbuf_len[port];
 
 	/*
 	 * If TX buffer for that queue is empty, and we have enough packets,
@@ -258,7 +258,7 @@ send_packetsx4(struct lcore_conf *qconf, uint16_t port, struct rte_mbuf *m[],
 		}
 	}
 
-	qconf->tx_mbufs[port].len = len;
+	qconf->tx_mbuf_len[port] = len;
 }
 
 #endif /* _L3FWD_COMMON_H_ */
examples/l3fwd/l3fwd_em.c +3 -3
index 9996bfba3..1970e0376 100644
@@ -662,12 +662,12 @@ em_main_loop(__rte_unused void *dummy)
 
 			for (i = 0; i < qconf->n_tx_port; ++i) {
 				portid = qconf->tx_port_id[i];
-				if (qconf->tx_mbufs[portid].len == 0)
+				if (qconf->tx_mbuf_len[portid] == 0)
 					continue;
 				send_burst(qconf,
-					qconf->tx_mbufs[portid].len,
+					qconf->tx_mbuf_len[portid],
 					portid);
-				qconf->tx_mbufs[portid].len = 0;
+				qconf->tx_mbuf_len[portid] = 0;
 			}
 
 			prev_tsc = cur_tsc;
examples/l3fwd/l3fwd_lpm.c +3 -3
index d338590b9..e62139a0e 100644
@@ -220,12 +220,12 @@ lpm_main_loop(__rte_unused void *dummy)
 
 			for (i = 0; i < n_tx_p; ++i) {
 				portid = qconf->tx_port_id[i];
-				if (qconf->tx_mbufs[portid].len == 0)
+				if (qconf->tx_mbuf_len[portid] == 0)
 					continue;
 				send_burst(qconf,
-					qconf->tx_mbufs[portid].len,
+					qconf->tx_mbuf_len[portid],
 					portid);
-				qconf->tx_mbufs[portid].len = 0;
+				qconf->tx_mbuf_len[portid] = 0;
 			}
 
 			prev_tsc = cur_tsc;
Signature · 1 line
2.25.1
#11 lore raw

Re: [dpdk-dev] [PATCH 4/4] examples/l3fwd: make data struct to be memory efficient

Quoted 12 lines · Ruifeng Wang

There are some holes in data struct lcore_conf. The holes are due to alignment requirement.

For struct lcore_rx_queue, there is no need to make every element of this type to be cache line aligned, because the data is not shared between cores.

Member len of struct mbuf_table can be moved out. So data can be packed and there will be no need to load an extra cache line when mbuf table is empty.

Quoted 5 lines

The change showed slight performance improvement on N1SDP platform.

This change alone is OK in the octeontx2 platform.(No difference in performance) combining with 3/4 shows some regression. Probably is due to prefetch or 128B cache line tuning specifics.

Quoted 121 lines
---
 examples/l3fwd/l3fwd.h        | 12 ++++++------
 examples/l3fwd/l3fwd_common.h |  4 ++--
 examples/l3fwd/l3fwd_em.c     |  6 +++---
 examples/l3fwd/l3fwd_lpm.c    |  6 +++---
 4 files changed, 14 insertions(+), 14 deletions(-)
examples/l3fwd/l3fwd.h +6 -6
index 2cf06099e..f3a301e12 100644
@@ -57,22 +57,22 @@
 #define HASH_ENTRY_NUMBER_DEFAULT      4
 
 struct mbuf_table {
-       uint16_t len;
        struct rte_mbuf *m_table[MAX_PKT_BURST];
 };
 
 struct lcore_rx_queue {
        uint16_t port_id;
        uint8_t queue_id;
-} __rte_cache_aligned;
+};
 
 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];
+       uint16_t tx_mbuf_len[RTE_MAX_ETHPORTS];
        struct mbuf_table tx_mbufs[RTE_MAX_ETHPORTS];
+       uint16_t n_rx_queue;
+       uint16_t n_tx_port;
        void *ipv4_lookup_struct;
        void *ipv6_lookup_struct;
 } __rte_cache_aligned;
@@ -122,7 +122,7 @@ send_single_packet(struct lcore_conf *qconf,
 {
        uint16_t len;
 
-       len = qconf->tx_mbufs[port].len;
+       len = qconf->tx_mbuf_len[port];
        qconf->tx_mbufs[port].m_table[len] = m;
        len++;
 
@@ -132,7 +132,7 @@ send_single_packet(struct lcore_conf *qconf,
                len = 0;
        }
 
-       qconf->tx_mbufs[port].len = len;
+       qconf->tx_mbuf_len[port] = len;
        return 0;
 }
 
examples/l3fwd/l3fwd_common.h +2 -2
index 7d83ff641..05e03dbfc 100644
@@ -183,7 +183,7 @@ send_packetsx4(struct lcore_conf *qconf, uint16_t port, struct rte_mbuf *m[],
 {
        uint32_t len, j, n;
 
-       len = qconf->tx_mbufs[port].len;
+       len = qconf->tx_mbuf_len[port];
 
        /*
         * If TX buffer for that queue is empty, and we have enough packets,
@@ -258,7 +258,7 @@ send_packetsx4(struct lcore_conf *qconf, uint16_t port, struct rte_mbuf *m[],
                }
        }
 
-       qconf->tx_mbufs[port].len = len;
+       qconf->tx_mbuf_len[port] = len;
 }
 
 #endif /* _L3FWD_COMMON_H_ */
examples/l3fwd/l3fwd_em.c +3 -3
index 9996bfba3..1970e0376 100644
@@ -662,12 +662,12 @@ em_main_loop(__rte_unused void *dummy)
 
                        for (i = 0; i < qconf->n_tx_port; ++i) {
                                portid = qconf->tx_port_id[i];
-                               if (qconf->tx_mbufs[portid].len == 0)
+                               if (qconf->tx_mbuf_len[portid] == 0)
                                        continue;
                                send_burst(qconf,
-                                       qconf->tx_mbufs[portid].len,
+                                       qconf->tx_mbuf_len[portid],
                                        portid);
-                               qconf->tx_mbufs[portid].len = 0;
+                               qconf->tx_mbuf_len[portid] = 0;
                        }
 
                        prev_tsc = cur_tsc;
examples/l3fwd/l3fwd_lpm.c +3 -4
index d338590b9..e62139a0e 100644
@@ -220,12 +220,12 @@ lpm_main_loop(__rte_unused void *dummy)
 
                        for (i = 0; i < n_tx_p; ++i) {
                                portid = qconf->tx_port_id[i];
-                               if (qconf->tx_mbufs[portid].len == 0)
+                               if (qconf->tx_mbuf_len[portid] == 0)
                                        continue;
                                send_burst(qconf,
-                                       qconf->tx_mbufs[portid].len,
+                                       qconf->tx_mbuf_len[portid],
                                        portid);
-                               qconf->tx_mbufs[portid].len = 0;
+                               qconf->tx_mbuf_len[portid] = 0;
                        }
 
                        prev_tsc = cur_tsc;
--

2.25.1

#12 lore raw

Re: [dpdk-dev] [PATCH 4/4] examples/l3fwd: make data struct to be memory efficient

Quoted 12 lines · Jerin Jacob
Quoted 11 lines · Ruifeng Wang

There are some holes in data struct lcore_conf. The holes are due to alignment requirement.

For struct lcore_rx_queue, there is no need to make every element of this type to be cache line aligned, because the data is not shared between cores.

Member len of struct mbuf_table can be moved out. So data can be packed and there will be no need to load an extra cache line when mbuf table is empty.

Quoted 8 lines
Quoted 5 lines

The change showed slight performance improvement on N1SDP platform.

This change alone is OK in the octeontx2 platform.(No difference in performance) combining with 3/4 shows some regression. Probably is due to prefetch or 128B cache line tuning specifics.

We checked it on Layerscape LS2088A platform. No difference for 1-2 core case. However observing ~2% regression for 4-8 cores.

Regards,

Hemant

#13 lore raw

Re: [dpdk-dev] [PATCH 4/4] examples/l3fwd: make data struct to be memory efficient

<snip>

Quoted 21 lines
Quoted 11 lines
Quoted 11 lines

There are some holes in data struct lcore_conf. The holes are due to alignment requirement.

For struct lcore_rx_queue, there is no need to make every element of this type to be cache line aligned, because the data is not shared between cores.

Member len of struct mbuf_table can be moved out. So data can be packed and there will be no need to load an extra cache line when mbuf table is empty.

Quoted 6 lines
Quoted 5 lines

The change showed slight performance improvement on N1SDP platform.

This change alone is OK in the octeontx2 platform.(No difference in

performance)

Quoted 2 lines

combining with 3/4 shows some regression. Probably is due to prefetch or 128B cache line tuning specifics.

Quoted 7 lines

We checked it on Layerscape LS2088A platform. No difference for 1-2 core case. However observing ~2% regression for 4-8 cores.

Regards,

Hemant

Hi Ruifeng,

l3fwd will no longer build with this patch as you have changed a struct used by the FIB lookup method. This patch will need to be updated to also update the FIB lookup method as you have done with EM and LPM.

Thanks,
Conor.

#14 lore raw

Re: [dpdk-dev] [PATCH 4/4] examples/l3fwd: make data struct to be memory efficient

Quoted 38 lines

-----Original Message-----
From: Walsh, Conor <[email protected]>
Sent: Monday, April 26, 2021 6:55 PM
To: [email protected]; Jerin Jacob <[email protected]>; Ruifeng Wang <[email protected]>
Cc: [email protected]; Yigit, Ferruh <[email protected]>; [email protected]; David Marchand <[email protected]>; dpdk-dev <[email protected]>; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>
Subject: RE: [dpdk-dev] [PATCH 4/4] examples/l3fwd: make data struct to be memory efficient

<snip>

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

There are some holes in data struct lcore_conf. The holes are due to alignment requirement.

For struct lcore_rx_queue, there is no need to make every element of this type to be cache line aligned, because the data is not shared between cores.

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

Member len of struct mbuf_table can be moved out. So data can be packed and there will be no need to load an extra cache line when mbuf table is empty.

The change showed slight performance improvement on N1SDP

platform.

Quoted 12 lines
Quoted 5 lines
Quoted 4 lines

This change alone is OK in the octeontx2 platform.(No difference in

performance)

Quoted 2 lines

combining with 3/4 shows some regression. Probably is due to prefetch or 128B cache line tuning specifics.

We checked it on Layerscape LS2088A platform. No difference for 1-2 core case. However observing ~2% regression for 4-8 cores.

Quoted 6 lines
Quoted 4 lines

Regards,

Hemant

Hi Ruifeng,

Hi Conor,

Quoted 5 lines

l3fwd will no longer build with this patch as you have changed a struct used by the FIB lookup method.
This patch will need to be updated to also update the FIB lookup method as you have done with EM and LPM.

Thanks for the comments.
I will provide v2 with updates. I'm considering to drop this one in v2 series.

Quoted 3 lines

Thanks,
Conor.

#15 lore raw
In reply to Ruifeng Wang

Hello,

This patch series targeted to improve L3fwd example. Performance gain was observed on N1SDP platform.

It would be good if you can run this series on your platforms and see if there is any performance impact.

Thanks,
Ruifeng

Quoted 28 lines

-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Thursday, March 18, 2021 6:26 PM
To: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>
Subject: [PATCH 0/4] l3fwd improvements

This series of patches include changes to l3fwd example application. Some improvements are made for better usage of CPU cycles and memory.

Ruifeng Wang (4):
examples/l3fwd: tune prefetch for better performance examples/l3fwd: eliminate unnecessary calculations examples/l3fwd: eliminate unnecessary reloads in loop examples/l3fwd: make data struct to be memory efficient

 examples/l3fwd/l3fwd.h          | 12 ++++++------
 examples/l3fwd/l3fwd_common.h   |  4 ++--
 examples/l3fwd/l3fwd_em.c       |  6 +++---
 examples/l3fwd/l3fwd_lpm.c      | 16 +++++++++-------
 examples/l3fwd/l3fwd_lpm_neon.h | 20 ++++++++++----------
 5 files changed, 30 insertions(+), 28 deletions(-)

--
2.25.1

#16 lore raw
Quoted 6 lines · Ruifeng Wang

Hello,

This patch series targeted to improve L3fwd example. Performance gain was observed on N1SDP platform.

It would be good if you can run this series on your platforms and see if there is any performance impact.

I will test and update and update in the respective patches.

Quoted 33 lines

Thanks,
Ruifeng

Quoted 28 lines

-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Thursday, March 18, 2021 6:26 PM
To: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>
Subject: [PATCH 0/4] l3fwd improvements

This series of patches include changes to l3fwd example application. Some improvements are made for better usage of CPU cycles and memory.

Ruifeng Wang (4):
examples/l3fwd: tune prefetch for better performance examples/l3fwd: eliminate unnecessary calculations examples/l3fwd: eliminate unnecessary reloads in loop examples/l3fwd: make data struct to be memory efficient

 examples/l3fwd/l3fwd.h          | 12 ++++++------
 examples/l3fwd/l3fwd_common.h   |  4 ++--
 examples/l3fwd/l3fwd_em.c       |  6 +++---
 examples/l3fwd/l3fwd_lpm.c      | 16 +++++++++-------
 examples/l3fwd/l3fwd_lpm_neon.h | 20 ++++++++++----------
 5 files changed, 30 insertions(+), 28 deletions(-)

--
2.25.1

Tested-by on the v1 series
#17 lore raw
In reply to Ruifeng Wang
Quoted 10 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ruifeng Wang
Sent: Thursday, March 18, 2021 06:26 PM
To: [email protected]; [email protected]; Yigit, Ferruh <[email protected]>; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Ruifeng Wang <[email protected]>
Subject: [dpdk-dev] [PATCH 0/4] l3fwd improvements

#18 lore raw

v2 cover l3fwd improvements

In reply to Ruifeng Wang

This series of patches include changes to l3fwd example application. Some improvements are made for better usage of CPU cycles and memory.

v2:
Dropped 1/4 prefetch to write change from v1.
Dropped 4/4 data struct change from v1.
Added 1/3 code reorganize.
Updated 3/3 to add 'const'. (Jerin)

Ruifeng Wang (3):
examples/l3fwd: reorganize code for better performance examples/l3fwd: eliminate unnecessary calculations examples/l3fwd: eliminate unnecessary reloads in loop

 examples/l3fwd/l3fwd_lpm.c      | 10 ++++++----
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 examples/l3fwd/l3fwd_neon.h     | 10 +++++-----
 3 files changed, 16 insertions(+), 14 deletions(-)
Signature · 1 line
2.25.1
#19 lore raw

PATCH v2 1/3 examples/l3fwd: reorganize code for better performance

Moved rfc1812 process prior to NEON registers store.
On N1SDP, this reorganization mitigates CPU frontend stall and backend stall when forwarding.

On N1SDP with MLX5 40G NIC, this change showed 10.2% performance gain in single port single core MRR test.
On ThunderX2, this changed showed no performance degradation.

---
 examples/l3fwd/l3fwd_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_neon.h +5 -5
index 86ac5971d7..ea7fe22d00 100644
@@ -43,11 +43,6 @@ processx4_step3(struct rte_mbuf *pkt[FWDSTEP], uint16_t dst_port[FWDSTEP])
 	ve[2] = vsetq_lane_u32(vgetq_lane_u32(te[2], 3), ve[2], 3);
 	ve[3] = vsetq_lane_u32(vgetq_lane_u32(te[3], 3), ve[3], 3);
 
-	vst1q_u32(p[0], ve[0]);
-	vst1q_u32(p[1], ve[1]);
-	vst1q_u32(p[2], ve[2]);
-	vst1q_u32(p[3], ve[3]);
-
 	rfc1812_process((struct rte_ipv4_hdr *)
 			((struct rte_ether_hdr *)p[0] + 1),
 			&dst_port[0], pkt[0]->packet_type);
@@ -60,6 +55,11 @@ processx4_step3(struct rte_mbuf *pkt[FWDSTEP], uint16_t dst_port[FWDSTEP])
 	rfc1812_process((struct rte_ipv4_hdr *)
 			((struct rte_ether_hdr *)p[3] + 1),
 			&dst_port[3], pkt[3]->packet_type);
+
+	vst1q_u32(p[0], ve[0]);
+	vst1q_u32(p[1], ve[1]);
+	vst1q_u32(p[2], ve[2]);
+	vst1q_u32(p[3], ve[3]);
 }
 
 /*
Signature · 1 line
2.25.1
#20 lore raw

Re: [dpdk-dev] [PATCH v2 1/3] examples/l3fwd: reorganize code for better performance

Quoted 7 lines · Ruifeng Wang

Moved rfc1812 process prior to NEON registers store.
On N1SDP, this reorganization mitigates CPU frontend stall and backend stall when forwarding.

On N1SDP with MLX5 40G NIC, this change showed 10.2% performance gain in single port single core MRR test.

I think, it may not have anything to do with N1SDP, It could be just the prefetch window timing
with MLX5 driver on Tx mbuf on touching with tx_burst() and L1 cache pressure timing.
I think, tuning the driver parameters can switch the window to some driver code.

On Octeontx2, this change has regression of -3.1% flow lookup miss case. so NACK.

Quoted 38 lines

On ThunderX2, this changed showed no performance degradation.

---
 examples/l3fwd/l3fwd_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_neon.h +5 -6
index 86ac5971d7..ea7fe22d00 100644
@@ -43,11 +43,6 @@ processx4_step3(struct rte_mbuf *pkt[FWDSTEP], uint16_t dst_port[FWDSTEP])
        ve[2] = vsetq_lane_u32(vgetq_lane_u32(te[2], 3), ve[2], 3);
        ve[3] = vsetq_lane_u32(vgetq_lane_u32(te[3], 3), ve[3], 3);
 
-       vst1q_u32(p[0], ve[0]);
-       vst1q_u32(p[1], ve[1]);
-       vst1q_u32(p[2], ve[2]);
-       vst1q_u32(p[3], ve[3]);
-
        rfc1812_process((struct rte_ipv4_hdr *)
                        ((struct rte_ether_hdr *)p[0] + 1),
                        &dst_port[0], pkt[0]->packet_type);
@@ -60,6 +55,11 @@ processx4_step3(struct rte_mbuf *pkt[FWDSTEP], uint16_t dst_port[FWDSTEP])
        rfc1812_process((struct rte_ipv4_hdr *)
                        ((struct rte_ether_hdr *)p[3] + 1),
                        &dst_port[3], pkt[3]->packet_type);
+
+       vst1q_u32(p[0], ve[0]);
+       vst1q_u32(p[1], ve[1]);
+       vst1q_u32(p[2], ve[2]);
+       vst1q_u32(p[3], ve[3]);
 }
 
 /*
--

2.25.1

#21 lore raw

PATCH v2 2/3 examples/l3fwd: eliminate unnecessary calculations

In reply to Ruifeng Wang

Both L2 and L3 headers will be used in forward processing. And these two headers are in the same cache line. It has the same effect for prefetching with L2 header address and prefetching with L3 header address.

Changed to use L2 header address for prefetching. The change showed no measurable performance improvement, but it definitely removed unnecessary instructions for address calculation.

---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_lpm_neon.h +5 -5
index d6c0ba64ab..78ee83b76c 100644
@@ -98,14 +98,14 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 	if (k) {
 		for (i = 0; i < FWDSTEP; i++) {
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
-						struct rte_ether_hdr *) + 1);
+							void *));
 		}
 
 		for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
 			for (i = 0; i < FWDSTEP; i++) {
 				rte_prefetch0(rte_pktmbuf_mtod(
 						pkts_burst[j + i + FWDSTEP],
-						struct rte_ether_hdr *) + 1);
+						void *));
 			}
 
 			processx4_step1(&pkts_burst[j], &dip, &ipv4_flag);
@@ -125,17 +125,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 		switch (m) {
 		case 3:
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 			/* fallthrough */
 		case 2:
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 			/* fallthrough */
 		case 1:
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 		}
 
Signature · 1 line
2.25.1
#22 lore raw

PATCH v2 3/3 examples/l3fwd: eliminate unnecessary reloads in loop

In reply to Ruifeng Wang

Number of rx queue and number of rx port in lcore config are constants during the period of l3 forward application running. But compiler has no this information.

Copied values from lcore config to local variables and used the local variables for iteration. Compiler can see that the local variables are not changed, so qconf reloads at each iteration can be eliminated.

The change showed 1.8% performance uplift in single core, single port, single queue test on N1SDP platform with MLX5 NIC.

---
 examples/l3fwd/l3fwd_lpm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
examples/l3fwd/l3fwd_lpm.c +6 -4
index 427c72b1d2..ff1c18a442 100644
@@ -154,14 +154,16 @@ lpm_main_loop(__rte_unused void *dummy)
 	lcore_id = rte_lcore_id();
 	qconf = &lcore_conf[lcore_id];
 
-	if (qconf->n_rx_queue == 0) {
+	const uint16_t n_rx_q = qconf->n_rx_queue;
+	const uint16_t n_tx_p = qconf->n_tx_port;
+	if (n_rx_q == 0) {
 		RTE_LOG(INFO, L3FWD, "lcore %u has nothing to do\n", lcore_id);
 		return 0;
 	}
 
 	RTE_LOG(INFO, L3FWD, "entering main loop on lcore %u\n", lcore_id);
 
-	for (i = 0; i < qconf->n_rx_queue; i++) {
+	for (i = 0; i < n_rx_q; i++) {
 
 		portid = qconf->rx_queue_list[i].port_id;
 		queueid = qconf->rx_queue_list[i].queue_id;
@@ -181,7 +183,7 @@ lpm_main_loop(__rte_unused void *dummy)
 		diff_tsc = cur_tsc - prev_tsc;
 		if (unlikely(diff_tsc > drain_tsc)) {
 
-			for (i = 0; i < qconf->n_tx_port; ++i) {
+			for (i = 0; i < n_tx_p; ++i) {
 				portid = qconf->tx_port_id[i];
 				if (qconf->tx_mbufs[portid].len == 0)
 					continue;
@@ -197,7 +199,7 @@ lpm_main_loop(__rte_unused void *dummy)
 		/*
 		 * Read packet from RX queues
 		 */
-		for (i = 0; i < qconf->n_rx_queue; ++i) {
+		for (i = 0; i < n_rx_q; ++i) {
 			portid = qconf->rx_queue_list[i].port_id;
 			queueid = qconf->rx_queue_list[i].queue_id;
 			nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
Signature · 1 line
2.25.1
Acked-by on v2 3/3
#23 lore raw

Re: [dpdk-dev] [PATCH v2 3/3] examples/l3fwd: eliminate unnecessary reloads in loop

Quoted 9 lines · Ruifeng Wang

Number of rx queue and number of rx port in lcore config are constants during the period of l3 forward application running. But compiler has no this information.

Copied values from lcore config to local variables and used the local variables for iteration. Compiler can see that the local variables are not changed, so qconf reloads at each iteration can be eliminated.

Quoted 6 lines

The change showed 1.8% performance uplift in single core, single port, single queue test on N1SDP platform with MLX5 NIC.

No performance regression with octeontx2.

Quoted 48 lines
---
 examples/l3fwd/l3fwd_lpm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
examples/l3fwd/l3fwd_lpm.c +6 -5
index 427c72b1d2..ff1c18a442 100644
@@ -154,14 +154,16 @@ lpm_main_loop(__rte_unused void *dummy)
        lcore_id = rte_lcore_id();
        qconf = &lcore_conf[lcore_id];
 
-       if (qconf->n_rx_queue == 0) {
+       const uint16_t n_rx_q = qconf->n_rx_queue;
+       const uint16_t n_tx_p = qconf->n_tx_port;
+       if (n_rx_q == 0) {
                RTE_LOG(INFO, L3FWD, "lcore %u has nothing to do\n", lcore_id);
                return 0;
        }
 
        RTE_LOG(INFO, L3FWD, "entering main loop on lcore %u\n", lcore_id);
 
-       for (i = 0; i < qconf->n_rx_queue; i++) {
+       for (i = 0; i < n_rx_q; i++) {
 
                portid = qconf->rx_queue_list[i].port_id;
                queueid = qconf->rx_queue_list[i].queue_id;
@@ -181,7 +183,7 @@ lpm_main_loop(__rte_unused void *dummy)
                diff_tsc = cur_tsc - prev_tsc;
                if (unlikely(diff_tsc > drain_tsc)) {
 
-                       for (i = 0; i < qconf->n_tx_port; ++i) {
+                       for (i = 0; i < n_tx_p; ++i) {
                                portid = qconf->tx_port_id[i];
                                if (qconf->tx_mbufs[portid].len == 0)
                                        continue;
@@ -197,7 +199,7 @@ lpm_main_loop(__rte_unused void *dummy)
                /*
                 * Read packet from RX queues
                 */
-               for (i = 0; i < qconf->n_rx_queue; ++i) {
+               for (i = 0; i < n_rx_q; ++i) {
                        portid = qconf->rx_queue_list[i].port_id;
                        queueid = qconf->rx_queue_list[i].queue_id;
                        nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
--

2.25.1

#24 lore raw

v3 cover l3fwd improvements

In reply to Ruifeng Wang

This series of patches include changes to l3fwd example application. Some improvements are made for better usage of CPU cycles and memory.

v3:
Dropped 1/3 code reorg from v2.

v2:
Dropped 1/4 prefetch to write change from v1.
Dropped 4/4 data struct change from v1.
Added 1/3 code reorganize.
Updated 3/3 to add 'const'. (Jerin)

Ruifeng Wang (2):
examples/l3fwd: eliminate unnecessary calculations examples/l3fwd: eliminate unnecessary reloads in loop

 examples/l3fwd/l3fwd_lpm.c      | 10 ++++++----
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 2 files changed, 11 insertions(+), 9 deletions(-)
Signature · 1 line
2.25.1
#25 lore raw

PATCH v3 1/2 examples/l3fwd: eliminate unnecessary calculations

Both L2 and L3 headers will be used in forward processing. And these two headers are in the same cache line. It has the same effect for prefetching with L2 header address and prefetching with L3 header address.

Changed to use L2 header address for prefetching. The change showed no measurable performance improvement, but it definitely removed unnecessary instructions for address calculation.

---
 examples/l3fwd/l3fwd_lpm_neon.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
examples/l3fwd/l3fwd_lpm_neon.h +5 -5
index d6c0ba64ab..78ee83b76c 100644
@@ -98,14 +98,14 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 	if (k) {
 		for (i = 0; i < FWDSTEP; i++) {
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i],
-						struct rte_ether_hdr *) + 1);
+							void *));
 		}
 
 		for (j = 0; j != k - FWDSTEP; j += FWDSTEP) {
 			for (i = 0; i < FWDSTEP; i++) {
 				rte_prefetch0(rte_pktmbuf_mtod(
 						pkts_burst[j + i + FWDSTEP],
-						struct rte_ether_hdr *) + 1);
+						void *));
 			}
 
 			processx4_step1(&pkts_burst[j], &dip, &ipv4_flag);
@@ -125,17 +125,17 @@ l3fwd_lpm_send_packets(int nb_rx, struct rte_mbuf **pkts_burst,
 		switch (m) {
 		case 3:
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 			/* fallthrough */
 		case 2:
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 			/* fallthrough */
 		case 1:
 			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[j],
-						struct rte_ether_hdr *) + 1);
+							void *));
 			j++;
 		}
 
Signature · 1 line
2.25.1
#26 lore raw

PATCH v3 2/2 examples/l3fwd: eliminate unnecessary reloads in loop

In reply to Ruifeng Wang

Number of rx queue and number of rx port in lcore config are constants during the period of l3 forward application running. But compiler has no this information.

Copied values from lcore config to local variables and used the local variables for iteration. Compiler can see that the local variables are not changed, so qconf reloads at each iteration can be eliminated.

The change showed 1.8% performance uplift in single core, single port, single queue test on N1SDP platform with MLX5 NIC.

---
 examples/l3fwd/l3fwd_lpm.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
examples/l3fwd/l3fwd_lpm.c +6 -4
index 427c72b1d2..ff1c18a442 100644
@@ -154,14 +154,16 @@ lpm_main_loop(__rte_unused void *dummy)
 	lcore_id = rte_lcore_id();
 	qconf = &lcore_conf[lcore_id];
 
-	if (qconf->n_rx_queue == 0) {
+	const uint16_t n_rx_q = qconf->n_rx_queue;
+	const uint16_t n_tx_p = qconf->n_tx_port;
+	if (n_rx_q == 0) {
 		RTE_LOG(INFO, L3FWD, "lcore %u has nothing to do\n", lcore_id);
 		return 0;
 	}
 
 	RTE_LOG(INFO, L3FWD, "entering main loop on lcore %u\n", lcore_id);
 
-	for (i = 0; i < qconf->n_rx_queue; i++) {
+	for (i = 0; i < n_rx_q; i++) {
 
 		portid = qconf->rx_queue_list[i].port_id;
 		queueid = qconf->rx_queue_list[i].queue_id;
@@ -181,7 +183,7 @@ lpm_main_loop(__rte_unused void *dummy)
 		diff_tsc = cur_tsc - prev_tsc;
 		if (unlikely(diff_tsc > drain_tsc)) {
 
-			for (i = 0; i < qconf->n_tx_port; ++i) {
+			for (i = 0; i < n_tx_p; ++i) {
 				portid = qconf->tx_port_id[i];
 				if (qconf->tx_mbufs[portid].len == 0)
 					continue;
@@ -197,7 +199,7 @@ lpm_main_loop(__rte_unused void *dummy)
 		/*
 		 * Read packet from RX queues
 		 */
-		for (i = 0; i < qconf->n_rx_queue; ++i) {
+		for (i = 0; i < n_rx_q; ++i) {
 			portid = qconf->rx_queue_list[i].port_id;
 			queueid = qconf->rx_queue_list[i].queue_id;
 			nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
Signature · 1 line
2.25.1
#27 lore raw
In reply to Ruifeng Wang
Quoted 4 lines · Ruifeng Wang

This series of patches include changes to l3fwd example application. Some improvements are made for better usage of CPU cycles and memory.

Quoted 9 lines

v3:
Dropped 1/3 code reorg from v2.

v2:
Dropped 1/4 prefetch to write change from v1.
Dropped 4/4 data struct change from v1.
Added 1/3 code reorganize.
Updated 3/3 to add 'const'. (Jerin)

Series applied, thanks.

Signature · 1 line
David Marchand