Hello Flavio, Great patch, I think you really did a lot to improve the code here and I think that's borne out by the consistent performance improvements across multiple tests.
Regarding the 4% regression that Intel detected, I found the following white paper to describe the "scatter" test: https://builders.intel.com/docs/networkbuilders/open-vswitch-optimized-deployment-benchmark-technology-guide.pdf This document calls out the following key points: The original test was summarized as: - 32 VMs with one million flows. - Test runs on four physical cores for OVS and 10 hyper-threaded cores for TestPMD - An Ixia pitches traffic at a sub 0.1% loss rate - The server catches traffic with a E810-C 100G - The traffic's profile is: Ether()/IP()/UDP()/VXLAN()/Ether()/IP() - On the outer IP, the source address changes incrementally across the 32 instances - The destination address remains the same on the outer IP. - The inner source IP remains - The inner destination address increments to create the one million flows for the test - EMC and SMC were disabled I could not reproduce this test exactly because I don't have access to the same hardware - notably the Intel NIC and an Ixia - and I didn't want to create an environment that wouldn't be reproduced in real world scenarios. I did pin VM and TXQs/RXQs threads to cores, but I didn't optimize the setup nearly to the extant that the white paper described. My test setup consisted of two Fedora 35 servers directly connected across Mellanox5E cards with Trex pitching traffic and TestPMD reflecting it. In my test I was still able to reproduce a similar performance penalty. I found that the key factors was the combination of VXLAN and a large number of flows. So once I had a setup that could reproduce close to the 4% penalty I stopped modifying my test framework and started searching for the slow code. I didn't see any obvious issues in the code that should cause a significant slowdown, in fact, most of the code is identical or slightly improved. So to help my analysis, I created several variations of your patch reverting small aspects of the change and benchmarked each variation. Because the difference in performance across each variation was so minor, I took a lot of samples. I pitched traffic over one million flows for 240 seconds and averaged out the throughput, I then repeated this process a total of five times for each patch. Finally, I repeated the whole process three times to produce 15 data points per patch. The best results came from the patch enclosed below, with the code from netdev_dpdk_common_send() protected by the splinlock, as it is in the pre-patch code. This yielded a 2.7% +/- 0.64 performance boost over the master branch. Cheers, Michael diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c index bc1633663..5db5d7e2a 100644 --- a/lib/netdev-dpdk.c +++ b/lib/netdev-dpdk.c @@ -2777,13 +2777,13 @@ netdev_dpdk_vhost_send(struct netdev *netdev, int qid, return 0; } - cnt = netdev_dpdk_common_send(netdev, batch, &stats); - if (OVS_UNLIKELY(!rte_spinlock_trylock(&dev->tx_q[qid].tx_lock))) { COVERAGE_INC(vhost_tx_contention); rte_spinlock_lock(&dev->tx_q[qid].tx_lock); } + cnt = netdev_dpdk_common_send(netdev, batch, &stats); + pkts = (struct rte_mbuf **) batch->packets; vhost_batch_cnt = cnt; retries = 0; @@ -2843,13 +2843,15 @@ netdev_dpdk_eth_send(struct netdev *netdev, int qid, return 0; } - cnt = netdev_dpdk_common_send(netdev, batch, &stats); - dropped = batch_cnt - cnt; if (OVS_UNLIKELY(concurrent_txq)) { qid = qid % dev->up.n_txq; rte_spinlock_lock(&dev->tx_q[qid].tx_lock); } + cnt = netdev_dpdk_common_send(netdev, batch, &stats); + + dropped = batch_cnt - cnt; + dropped += netdev_dpdk_eth_tx_burst(dev, qid, pkts, cnt); if (OVS_UNLIKELY(dropped)) { struct netdev_dpdk_sw_stats *sw_stats = dev->sw_stats; On Sun, 2021-01-10 at 00:05 -0300, Flavio Leitner wrote: > This patch split out the common code between vhost and > dpdk transmit paths to shared functions to simplify the > code and fix an issue. > > The issue is that the packet coming from non-DPDK device > and egressing on a DPDK device currently skips the hwol > preparation. > > This also have the side effect of leaving only the dpdk > transmit code under the txq lock. > > Signed-off-by: Flavio Leitner <[email protected]> > Reviewed-by: David Marchand <[email protected]> > --- > > V2: > - mentioned the tx lock change in the commit message. > - fixed packet leak when copy fails. > - moved pkt_cnt = cnt two lines up. > > I tested the following scenarios with iperf and iperf -R > and noticed no performance regression: > IPERF VM-Bridge 1.02x > IPERF VM-VM 1.04x > IPERF VM-ExtHost 1.00x > IPERF VM-NS 1.01x > IPERF VM-VLAN-Bridge 1.03x > IPERF VM-VLAN-VM 1.03x > IPERF VM-VLAN-ExtHost 1.01x > IPERF VM-VLAN-NS 1.02x > IPERF VM-V6-VM 1.03x > IPERF VM-V6-ExtHost 1.00x > IPERF VM-V6-NS 1.01x > IPERF-R VM-Bridge 1.01x > IPERF-R VM-VM 1.04x > IPERF-R VM-ExtHost 1.10x > IPERF-R VM-NS 1.01x > IPERF-R VM-VLAN-Bridge 1.03x > IPERF-R VM-VLAN-VM 1.02x > IPERF-R VM-VLAN-ExtHost 1.08x > IPERF-R VM-VLAN-NS 1.02x > IPERF-R VM-V6-VM 1.00x > IPERF-R VM-V6-ExtHost 1.11x > IPERF-R VM-V6-NS 1.00x > > Now using trex, 64byte packet, PVP: > Original: 3.6Mpps > avg. packets per output batch: 32.00 > idle cycles: 0 (0.00%) > avg cycles per packet: 304.92 (8331383020/27323150) > avg processing cycles per packet: 304.92 (8331383020/27323150) > > Patched: 3.6Mpps > avg. packets per output batch: 32.00 > idle cycles: 0 (0.00%) > avg cycles per packet: 304.08 (21875784116/71941516) > avg processing cycles per packet: 304.08 (21875784116/71941516) > > lib/netdev-dpdk.c | 335 +++++++++++++++++++------------------------- > -- > 1 file changed, 139 insertions(+), 196 deletions(-) > > diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c > index 2640a421a..a1437db4d 100644 > --- a/lib/netdev-dpdk.c > +++ b/lib/netdev-dpdk.c > @@ -2585,90 +2585,6 @@ netdev_dpdk_vhost_update_tx_counters(struct > netdev_dpdk *dev, > } > } > > -static void > -__netdev_dpdk_vhost_send(struct netdev *netdev, int qid, > - struct dp_packet **pkts, int cnt) > -{ > - struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); > - struct rte_mbuf **cur_pkts = (struct rte_mbuf **) pkts; > - struct netdev_dpdk_sw_stats sw_stats_add; > - unsigned int n_packets_to_free = cnt; > - unsigned int total_packets = cnt; > - int i, retries = 0; > - int max_retries = VHOST_ENQ_RETRY_MIN; > - int vid = netdev_dpdk_get_vid(dev); > - > - qid = dev->tx_q[qid % netdev->n_txq].map; > - > - if (OVS_UNLIKELY(vid < 0 || !dev->vhost_reconfigured || qid < 0 > - || !(dev->flags & NETDEV_UP))) { > - rte_spinlock_lock(&dev->stats_lock); > - dev->stats.tx_dropped+= cnt; > - rte_spinlock_unlock(&dev->stats_lock); > - goto out; > - } > - > - if (OVS_UNLIKELY(!rte_spinlock_trylock(&dev- > >tx_q[qid].tx_lock))) { > - COVERAGE_INC(vhost_tx_contention); > - rte_spinlock_lock(&dev->tx_q[qid].tx_lock); > - } > - > - sw_stats_add.tx_invalid_hwol_drops = cnt; > - if (userspace_tso_enabled()) { > - cnt = netdev_dpdk_prep_hwol_batch(dev, cur_pkts, cnt); > - } > - > - sw_stats_add.tx_invalid_hwol_drops -= cnt; > - sw_stats_add.tx_mtu_exceeded_drops = cnt; > - cnt = netdev_dpdk_filter_packet_len(dev, cur_pkts, cnt); > - sw_stats_add.tx_mtu_exceeded_drops -= cnt; > - > - /* Check has QoS has been configured for the netdev */ > - sw_stats_add.tx_qos_drops = cnt; > - cnt = netdev_dpdk_qos_run(dev, cur_pkts, cnt, true); > - sw_stats_add.tx_qos_drops -= cnt; > - > - n_packets_to_free = cnt; > - > - do { > - int vhost_qid = qid * VIRTIO_QNUM + VIRTIO_RXQ; > - unsigned int tx_pkts; > - > - tx_pkts = rte_vhost_enqueue_burst(vid, vhost_qid, cur_pkts, > cnt); > - if (OVS_LIKELY(tx_pkts)) { > - /* Packets have been sent.*/ > - cnt -= tx_pkts; > - /* Prepare for possible retry.*/ > - cur_pkts = &cur_pkts[tx_pkts]; > - if (OVS_UNLIKELY(cnt && !retries)) { > - /* > - * Read max retries as there are packets not sent > - * and no retries have already occurred. > - */ > - atomic_read_relaxed(&dev->vhost_tx_retries_max, > &max_retries); > - } > - } else { > - /* No packets sent - do not retry.*/ > - break; > - } > - } while (cnt && (retries++ < max_retries)); > - > - rte_spinlock_unlock(&dev->tx_q[qid].tx_lock); > - > - sw_stats_add.tx_failure_drops = cnt; > - sw_stats_add.tx_retries = MIN(retries, max_retries); > - > - rte_spinlock_lock(&dev->stats_lock); > - netdev_dpdk_vhost_update_tx_counters(dev, pkts, total_packets, > - &sw_stats_add); > - rte_spinlock_unlock(&dev->stats_lock); > - > -out: > - for (i = 0; i < n_packets_to_free; i++) { > - dp_packet_delete(pkts[i]); > - } > -} > - > static void > netdev_dpdk_extbuf_free(void *addr OVS_UNUSED, void *opaque) > { > @@ -2783,76 +2699,70 @@ dpdk_copy_dp_packet_to_mbuf(struct > rte_mempool *mp, struct dp_packet *pkt_orig) > return pkt_dest; > } > > -/* Tx function. Transmit packets indefinitely */ > -static void > -dpdk_do_tx_copy(struct netdev *netdev, int qid, struct > dp_packet_batch *batch) > +/* Replace packets in a 'batch' with their corresponding copies > using > + * DPDK memory. > + * > + * Returns the number of good packets in the batch. */ > +static size_t > +dpdk_copy_batch_to_mbuf(struct netdev *netdev, struct > dp_packet_batch *batch) > OVS_NO_THREAD_SAFETY_ANALYSIS > { > - const size_t batch_cnt = dp_packet_batch_size(batch); > -#if !defined(__CHECKER__) && !defined(_WIN32) > - const size_t PKT_ARRAY_SIZE = batch_cnt; > -#else > - /* Sparse or MSVC doesn't like variable length array. */ > - enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; > -#endif > struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); > - struct dp_packet *pkts[PKT_ARRAY_SIZE]; > - struct netdev_dpdk_sw_stats *sw_stats = dev->sw_stats; > - uint32_t cnt = batch_cnt; > - uint32_t dropped = 0; > - uint32_t tx_failure = 0; > - uint32_t mtu_drops = 0; > - uint32_t qos_drops = 0; > - > - if (dev->type != DPDK_DEV_VHOST) { > - /* Check if QoS has been configured for this netdev. */ > - cnt = netdev_dpdk_qos_run(dev, (struct rte_mbuf **) batch- > >packets, > - batch_cnt, false); > - qos_drops = batch_cnt - cnt; > - } > - > - uint32_t txcnt = 0; > - > - for (uint32_t i = 0; i < cnt; i++) { > - struct dp_packet *packet = batch->packets[i]; > - uint32_t size = dp_packet_size(packet); > - > - if (size > dev->max_packet_len > - && !(packet->mbuf.ol_flags & PKT_TX_TCP_SEG)) { > - VLOG_WARN_RL(&rl, "Too big size %u max_packet_len %d", > size, > - dev->max_packet_len); > - mtu_drops++; > - continue; > - } > + size_t i, size = dp_packet_batch_size(batch); > + struct dp_packet *packet; > > - pkts[txcnt] = dpdk_copy_dp_packet_to_mbuf(dev->dpdk_mp->mp, > packet); > - if (OVS_UNLIKELY(!pkts[txcnt])) { > - dropped = cnt - i; > - break; > - } > + DP_PACKET_BATCH_REFILL_FOR_EACH (i, size, packet, batch) { > + if (OVS_UNLIKELY(packet->source == DPBUF_DPDK)) { > + dp_packet_batch_refill(batch, packet, i); > + } else { > + struct dp_packet *pktcopy; > > - txcnt++; > - } > + pktcopy = dpdk_copy_dp_packet_to_mbuf(dev->dpdk_mp->mp, > packet); > + if (pktcopy) { > + dp_packet_batch_refill(batch, pktcopy, i); > + } > > - if (OVS_LIKELY(txcnt)) { > - if (dev->type == DPDK_DEV_VHOST) { > - __netdev_dpdk_vhost_send(netdev, qid, pkts, txcnt); > - } else { > - tx_failure += netdev_dpdk_eth_tx_burst(dev, qid, > - (struct rte_mbuf > **)pkts, > - txcnt); > + dp_packet_delete(packet); > } > } > > - dropped += qos_drops + mtu_drops + tx_failure; > - if (OVS_UNLIKELY(dropped)) { > - rte_spinlock_lock(&dev->stats_lock); > - dev->stats.tx_dropped += dropped; > - sw_stats->tx_failure_drops += tx_failure; > - sw_stats->tx_mtu_exceeded_drops += mtu_drops; > - sw_stats->tx_qos_drops += qos_drops; > - rte_spinlock_unlock(&dev->stats_lock); > + return dp_packet_batch_size(batch); > +} > + > +static size_t > +netdev_dpdk_common_send(struct netdev *netdev, struct > dp_packet_batch *batch, > + struct netdev_dpdk_sw_stats *stats) > +{ > + struct rte_mbuf **pkts = (struct rte_mbuf **) batch->packets; > + struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); > + size_t cnt, pkt_cnt = dp_packet_batch_size(batch); > + > + memset(stats, 0, sizeof *stats); > + > + /* Copy dp-packets to mbufs. */ > + if (OVS_UNLIKELY(batch->packets[0]->source != DPBUF_DPDK)) { > + cnt = dpdk_copy_batch_to_mbuf(netdev, batch); > + stats->tx_failure_drops += pkt_cnt - cnt; > + pkt_cnt = cnt; > } > + > + /* Drop oversized packets. */ > + cnt = netdev_dpdk_filter_packet_len(dev, pkts, pkt_cnt); > + stats->tx_mtu_exceeded_drops += pkt_cnt - cnt; > + pkt_cnt = cnt; > + > + /* Prepare each mbuf for hardware offloading. */ > + if (userspace_tso_enabled()) { > + cnt = netdev_dpdk_prep_hwol_batch(dev, pkts, pkt_cnt); > + stats->tx_invalid_hwol_drops += pkt_cnt - cnt; > + pkt_cnt = cnt; > + } > + > + /* Apply Quality of Service policy. */ > + cnt = netdev_dpdk_qos_run(dev, pkts, pkt_cnt, true); > + stats->tx_qos_drops += pkt_cnt - cnt; > + > + return cnt; > } > > static int > @@ -2860,82 +2770,115 @@ netdev_dpdk_vhost_send(struct netdev > *netdev, int qid, > struct dp_packet_batch *batch, > bool concurrent_txq OVS_UNUSED) > { > + struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); > + int max_retries = VHOST_ENQ_RETRY_MIN; > + int cnt, batch_cnt, vhost_batch_cnt; > + int vid = netdev_dpdk_get_vid(dev); > + struct netdev_dpdk_sw_stats stats; > + struct rte_mbuf **pkts; > + int retries; > > - if (OVS_UNLIKELY(batch->packets[0]->source != DPBUF_DPDK)) { > - dpdk_do_tx_copy(netdev, qid, batch); > + batch_cnt = cnt = dp_packet_batch_size(batch); > + qid = dev->tx_q[qid % netdev->n_txq].map; > + if (OVS_UNLIKELY(vid < 0 || !dev->vhost_reconfigured || qid < 0 > + || !(dev->flags & NETDEV_UP))) { > + rte_spinlock_lock(&dev->stats_lock); > + dev->stats.tx_dropped += cnt; > + rte_spinlock_unlock(&dev->stats_lock); > dp_packet_delete_batch(batch, true); > - } else { > - __netdev_dpdk_vhost_send(netdev, qid, batch->packets, > - dp_packet_batch_size(batch)); > + return 0; > + } > + > + cnt = netdev_dpdk_common_send(netdev, batch, &stats); > + > + if (OVS_UNLIKELY(!rte_spinlock_trylock(&dev- > >tx_q[qid].tx_lock))) { > + COVERAGE_INC(vhost_tx_contention); > + rte_spinlock_lock(&dev->tx_q[qid].tx_lock); > } > + > + pkts = (struct rte_mbuf **) batch->packets; > + vhost_batch_cnt = cnt; > + retries = 0; > + do { > + int vhost_qid = qid * VIRTIO_QNUM + VIRTIO_RXQ; > + int tx_pkts; > + > + tx_pkts = rte_vhost_enqueue_burst(vid, vhost_qid, pkts, > cnt); > + if (OVS_LIKELY(tx_pkts)) { > + /* Packets have been sent.*/ > + cnt -= tx_pkts; > + /* Prepare for possible retry.*/ > + pkts = &pkts[tx_pkts]; > + if (OVS_UNLIKELY(cnt && !retries)) { > + /* > + * Read max retries as there are packets not sent > + * and no retries have already occurred. > + */ > + atomic_read_relaxed(&dev->vhost_tx_retries_max, > &max_retries); > + } > + } else { > + /* No packets sent - do not retry.*/ > + break; > + } > + } while (cnt && (retries++ < max_retries)); > + > + rte_spinlock_unlock(&dev->tx_q[qid].tx_lock); > + > + stats.tx_failure_drops += cnt; > + stats.tx_retries = MIN(retries, max_retries); > + > + rte_spinlock_lock(&dev->stats_lock); > + netdev_dpdk_vhost_update_tx_counters(dev, batch->packets, > batch_cnt, > + &stats); > + rte_spinlock_unlock(&dev->stats_lock); > + > + pkts = (struct rte_mbuf **) batch->packets; > + for (int i = 0; i < vhost_batch_cnt; i++) { > + rte_pktmbuf_free(pkts[i]); > + } > + > return 0; > } > > -static inline void > -netdev_dpdk_send__(struct netdev_dpdk *dev, int qid, > - struct dp_packet_batch *batch, > - bool concurrent_txq) > +static int > +netdev_dpdk_eth_send(struct netdev *netdev, int qid, > + struct dp_packet_batch *batch, bool > concurrent_txq) > { > + struct rte_mbuf **pkts = (struct rte_mbuf **) batch->packets; > + struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); > + int batch_cnt = dp_packet_batch_size(batch); > + struct netdev_dpdk_sw_stats stats; > + int cnt, dropped; > + > if (OVS_UNLIKELY(!(dev->flags & NETDEV_UP))) { > dp_packet_delete_batch(batch, true); > - return; > + return 0; > } > > + cnt = netdev_dpdk_common_send(netdev, batch, &stats); > + dropped = batch_cnt - cnt; > if (OVS_UNLIKELY(concurrent_txq)) { > qid = qid % dev->up.n_txq; > rte_spinlock_lock(&dev->tx_q[qid].tx_lock); > } > > - if (OVS_UNLIKELY(batch->packets[0]->source != DPBUF_DPDK)) { > - struct netdev *netdev = &dev->up; > - > - dpdk_do_tx_copy(netdev, qid, batch); > - dp_packet_delete_batch(batch, true); > - } else { > + dropped += netdev_dpdk_eth_tx_burst(dev, qid, pkts, cnt); > + if (OVS_UNLIKELY(dropped)) { > struct netdev_dpdk_sw_stats *sw_stats = dev->sw_stats; > - int dropped; > - int tx_failure, mtu_drops, qos_drops, hwol_drops; > - int batch_cnt = dp_packet_batch_size(batch); > - struct rte_mbuf **pkts = (struct rte_mbuf **) batch- > >packets; > > - hwol_drops = batch_cnt; > - if (userspace_tso_enabled()) { > - batch_cnt = netdev_dpdk_prep_hwol_batch(dev, pkts, > batch_cnt); > - } > - hwol_drops -= batch_cnt; > - mtu_drops = batch_cnt; > - batch_cnt = netdev_dpdk_filter_packet_len(dev, pkts, > batch_cnt); > - mtu_drops -= batch_cnt; > - qos_drops = batch_cnt; > - batch_cnt = netdev_dpdk_qos_run(dev, pkts, batch_cnt, true); > - qos_drops -= batch_cnt; > - > - tx_failure = netdev_dpdk_eth_tx_burst(dev, qid, pkts, > batch_cnt); > - > - dropped = tx_failure + mtu_drops + qos_drops + hwol_drops; > - if (OVS_UNLIKELY(dropped)) { > - rte_spinlock_lock(&dev->stats_lock); > - dev->stats.tx_dropped += dropped; > - sw_stats->tx_failure_drops += tx_failure; > - sw_stats->tx_mtu_exceeded_drops += mtu_drops; > - sw_stats->tx_qos_drops += qos_drops; > - sw_stats->tx_invalid_hwol_drops += hwol_drops; > - rte_spinlock_unlock(&dev->stats_lock); > - } > + rte_spinlock_lock(&dev->stats_lock); > + dev->stats.tx_dropped += dropped; > + sw_stats->tx_failure_drops += stats.tx_failure_drops; > + sw_stats->tx_mtu_exceeded_drops += > stats.tx_mtu_exceeded_drops; > + sw_stats->tx_qos_drops += stats.tx_qos_drops; > + sw_stats->tx_invalid_hwol_drops += > stats.tx_invalid_hwol_drops; > + rte_spinlock_unlock(&dev->stats_lock); > } > > if (OVS_UNLIKELY(concurrent_txq)) { > rte_spinlock_unlock(&dev->tx_q[qid].tx_lock); > } > -} > - > -static int > -netdev_dpdk_eth_send(struct netdev *netdev, int qid, > - struct dp_packet_batch *batch, bool > concurrent_txq) > -{ > - struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); > > - netdev_dpdk_send__(dev, qid, batch, concurrent_txq); > return 0; > } > _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
