Re: [ovs-dev] [PATCH v4] dpif-netdev: Avoid reading RSS hash when EMC is disabled
On 8/25/17, 3:21 PM, "Fischetti, Antonio" <antonio.fische...@intel.com> wrote: > -Original Message- > From: Darrell Ball [mailto:db...@vmware.com] > Sent: Friday, August 25, 2017 6:19 PM > To: Fischetti, Antonio <antonio.fische...@intel.com>; d...@openvswitch.org > Subject: Re: [ovs-dev] [PATCH v4] dpif-netdev: Avoid reading RSS hash when EMC > is disabled > > Hi Antonio > > On 8/25/17, 6:56 AM, "ovs-dev-boun...@openvswitch.org on behalf of > antonio.fische...@intel.com" <ovs-dev-boun...@openvswitch.org on behalf of > antonio.fische...@intel.com> wrote: > > When EMC is disabled the reading of RSS hash is skipped. > Also, for packets that are not recirculated it retrieves > the hash value without considering the recirc id. > > Signed-off-by: Antonio Fischetti <antonio.fische...@intel.com> > --- > > V4 > - reworked to remove dependencies from other patches in > patchset "Skip EMC for recirc pkts and other optimizations." > https://urldefense.proofpoint.com/v2/url?u=https- > 3A__mail.openvswitch.org_pipermail_ovs-2Ddev_2017- > 2DAugust_337320.html=DwICAg=uilaK90D4TOVoH58JNXRgQ=BVhFA09CGX7JQ5Ih- > uZnsw=7UGZunJeutRugFmidRAykhZFRYscZ2rUUmhF7Ys9IM8=hzVGT1bI3zOj_x9J5cu18Q2D5 > WLHdEC1qhQ-Gu--TEE= > > - measurements were repeated with the latest head of master. > > Port-to-Port Test > = > Software built with "-O2 -march=native -g". > > I measured the Rx rate regardless of pkt loss by sending 1 UDP > flow, > 64B packets, at the line-rate. > > 2 PMDs with 3 Tx queues. > > Flows setup: > in_port=dpdk0 actions=output:dpdk1 > in_port=dpdk1 actions=output:dpdk0 > > Results > --- > Values are for the Rx rate in Mpps, regardless of packet loss. > > RSS column: >Yes: RSS hash is provided by the NIC >No: RSS is disabled and the 5-tuple hash must be >computed in software. > >Note: to simulate RSS disabled I added the line >dp_packet_rss_valid(struct dp_packet *p) >{ >+return false; >#ifdef DPDK_NETDEV > > EMC column: >Yes: default probability insertion, >No: EMC disabled. > > Orig means Commit ID: > 75f9e007e7f7eb91461e238f882d1c539c56bb8d > > [Darrell] > It looks like the main benefit (+6% throughput) occurs > when RSS is not calculated by the NIC and emc is disabled. What are the common > cases when RSS is not generated by the NIC ? [Antonio] A common case could be when packets are coming in from a VM. Well, there is no pnic in this case. I am thinking gateway usage is predominate in importance, but the RSS cases are about equal based on the sampling below, so we could almost ‘remove them from the equation’, as it were. For non-RSS: Using emc, there is about 0.7 % degradation, but it is more common to use emc So overall, it is not clear; what do you think ? Also when the NIC does not provide a hash value, but this shouldn't be considered a common case I guess. I agree. > > > > +--++ > +-+-+Orig | Orig + this| > | RSS | EMC | |patch | > +-+-+--++ > | Yes | Yes | 12.02, 11.38 | 12.18, 11.32 | > | Yes | No | 8.35, 8.36 | 8.36, 8.38 | > +-+-+--++ > | No | Yes | 9.91, 11.15 | 9.84, 11.08 | > | No | No | 7.83, 7.87 | 8.33, 8.35 | > +-+-+--++ > > --- > lib/dpif-netdev.c | 32 > 1 file changed, 28 insertions(+), 4 deletions(-) > > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c > index e2cd931..1157998 100644 > --- a/lib/dpif-netdev.c > +++ b/lib/dpif-netdev.c > @@ -4567,6 +4567,22 @@ dp_netdev_upcall(struct d
Re: [ovs-dev] [PATCH v4] dpif-netdev: Avoid reading RSS hash when EMC is disabled
> -Original Message- > From: Darrell Ball [mailto:db...@vmware.com] > Sent: Friday, August 25, 2017 6:19 PM > To: Fischetti, Antonio <antonio.fische...@intel.com>; d...@openvswitch.org > Subject: Re: [ovs-dev] [PATCH v4] dpif-netdev: Avoid reading RSS hash when EMC > is disabled > > Hi Antonio > > On 8/25/17, 6:56 AM, "ovs-dev-boun...@openvswitch.org on behalf of > antonio.fische...@intel.com" <ovs-dev-boun...@openvswitch.org on behalf of > antonio.fische...@intel.com> wrote: > > When EMC is disabled the reading of RSS hash is skipped. > Also, for packets that are not recirculated it retrieves > the hash value without considering the recirc id. > > Signed-off-by: Antonio Fischetti <antonio.fische...@intel.com> > --- > > V4 > - reworked to remove dependencies from other patches in > patchset "Skip EMC for recirc pkts and other optimizations." > https://urldefense.proofpoint.com/v2/url?u=https- > 3A__mail.openvswitch.org_pipermail_ovs-2Ddev_2017- > 2DAugust_337320.html=DwICAg=uilaK90D4TOVoH58JNXRgQ=BVhFA09CGX7JQ5Ih- > uZnsw=7UGZunJeutRugFmidRAykhZFRYscZ2rUUmhF7Ys9IM8=hzVGT1bI3zOj_x9J5cu18Q2D5 > WLHdEC1qhQ-Gu--TEE= > > - measurements were repeated with the latest head of master. > > Port-to-Port Test > = > Software built with "-O2 -march=native -g". > > I measured the Rx rate regardless of pkt loss by sending 1 UDP > flow, > 64B packets, at the line-rate. > > 2 PMDs with 3 Tx queues. > > Flows setup: > in_port=dpdk0 actions=output:dpdk1 > in_port=dpdk1 actions=output:dpdk0 > > Results > --- > Values are for the Rx rate in Mpps, regardless of packet loss. > > RSS column: >Yes: RSS hash is provided by the NIC >No: RSS is disabled and the 5-tuple hash must be >computed in software. > >Note: to simulate RSS disabled I added the line >dp_packet_rss_valid(struct dp_packet *p) >{ >+return false; >#ifdef DPDK_NETDEV > > EMC column: >Yes: default probability insertion, >No: EMC disabled. > > Orig means Commit ID: > 75f9e007e7f7eb91461e238f882d1c539c56bb8d > > [Darrell] > It looks like the main benefit (+6% throughput) occurs > when RSS is not calculated by the NIC and emc is disabled. What are the common > cases when RSS is not generated by the NIC ? [Antonio] A common case could be when packets are coming in from a VM. Also when the NIC does not provide a hash value, but this shouldn't be considered a common case I guess. > > > > +--++ > +-+-+Orig | Orig + this| > | RSS | EMC | |patch | > +-+-+--++ > | Yes | Yes | 12.02, 11.38 | 12.18, 11.32 | > | Yes | No | 8.35, 8.36 | 8.36, 8.38 | > +-+-+--++ > | No | Yes | 9.91, 11.15 | 9.84, 11.08 | > | No | No | 7.83, 7.87 | 8.33, 8.35 | > +-+-+--++ > > --- > lib/dpif-netdev.c | 32 > 1 file changed, 28 insertions(+), 4 deletions(-) > > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c > index e2cd931..1157998 100644 > --- a/lib/dpif-netdev.c > +++ b/lib/dpif-netdev.c > @@ -4567,6 +4567,22 @@ dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, > struct dp_packet *packet_, > } > > static inline uint32_t > +dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet *packet, > +const struct miniflow *mf) > +{ > +uint32_t hash; > + > +if (OVS_LIKELY(dp_packet_rss_valid(packet))) { > +hash = dp_packet_get_rss_hash(packet); > +} else { > +hash = miniflow_hash_5tuple(mf, 0); > +dp_packet_set_rss_hash(packet, hash); > +} > + > +return hash; > +} > + > +static inline uint32_t > dpif_netdev_packet_get_rss_hash(struct dp_packet *packet, > const struct miniflow *mf) > { > @@ -4701,10 +4717,18 @@ emc_processing(struct dp_netdev_
Re: [ovs-dev] [PATCH v4] dpif-netdev: Avoid reading RSS hash when EMC is disabled
Hi Antonio On 8/25/17, 6:56 AM, "ovs-dev-boun...@openvswitch.org on behalf of antonio.fische...@intel.com"wrote: When EMC is disabled the reading of RSS hash is skipped. Also, for packets that are not recirculated it retrieves the hash value without considering the recirc id. Signed-off-by: Antonio Fischetti --- V4 - reworked to remove dependencies from other patches in patchset "Skip EMC for recirc pkts and other optimizations." https://urldefense.proofpoint.com/v2/url?u=https-3A__mail.openvswitch.org_pipermail_ovs-2Ddev_2017-2DAugust_337320.html=DwICAg=uilaK90D4TOVoH58JNXRgQ=BVhFA09CGX7JQ5Ih-uZnsw=7UGZunJeutRugFmidRAykhZFRYscZ2rUUmhF7Ys9IM8=hzVGT1bI3zOj_x9J5cu18Q2D5WLHdEC1qhQ-Gu--TEE= - measurements were repeated with the latest head of master. Port-to-Port Test = Software built with "-O2 -march=native -g". I measured the Rx rate regardless of pkt loss by sending 1 UDP flow, 64B packets, at the line-rate. 2 PMDs with 3 Tx queues. Flows setup: in_port=dpdk0 actions=output:dpdk1 in_port=dpdk1 actions=output:dpdk0 Results --- Values are for the Rx rate in Mpps, regardless of packet loss. RSS column: Yes: RSS hash is provided by the NIC No: RSS is disabled and the 5-tuple hash must be computed in software. Note: to simulate RSS disabled I added the line dp_packet_rss_valid(struct dp_packet *p) { +return false; #ifdef DPDK_NETDEV EMC column: Yes: default probability insertion, No: EMC disabled. Orig means Commit ID: 75f9e007e7f7eb91461e238f882d1c539c56bb8d [Darrell] It looks like the main benefit (+6% throughput) occurs when RSS is not calculated by the NIC and emc is disabled. What are the common cases when RSS is not generated by the NIC ? +--++ +-+-+Orig | Orig + this| | RSS | EMC | |patch | +-+-+--++ | Yes | Yes | 12.02, 11.38 | 12.18, 11.32 | | Yes | No | 8.35, 8.36 | 8.36, 8.38 | +-+-+--++ | No | Yes | 9.91, 11.15 | 9.84, 11.08 | | No | No | 7.83, 7.87 | 8.33, 8.35 | +-+-+--++ --- lib/dpif-netdev.c | 32 1 file changed, 28 insertions(+), 4 deletions(-) diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index e2cd931..1157998 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -4567,6 +4567,22 @@ dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_, } static inline uint32_t +dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet *packet, +const struct miniflow *mf) +{ +uint32_t hash; + +if (OVS_LIKELY(dp_packet_rss_valid(packet))) { +hash = dp_packet_get_rss_hash(packet); +} else { +hash = miniflow_hash_5tuple(mf, 0); +dp_packet_set_rss_hash(packet, hash); +} + +return hash; +} + +static inline uint32_t dpif_netdev_packet_get_rss_hash(struct dp_packet *packet, const struct miniflow *mf) { @@ -4701,10 +4717,18 @@ emc_processing(struct dp_netdev_pmd_thread *pmd, } miniflow_extract(packet, >mf); key->len = 0; /* Not computed yet. */ -key->hash = dpif_netdev_packet_get_rss_hash(packet, >mf); - -/* If EMC is disabled skip emc_lookup */ -flow = (cur_min == 0) ? NULL: emc_lookup(flow_cache, key); +/* If EMC is disabled skip hash computation and emc_lookup */ +if (OVS_LIKELY(cur_min)) { +if (!md_is_valid) { +key->hash = dpif_netdev_packet_get_rss_hash_orig_pkt(packet, +>mf); +} else { +key->hash = dpif_netdev_packet_get_rss_hash(packet, >mf); +} +flow = emc_lookup(flow_cache, key); +} else { +flow = NULL; +} if (OVS_LIKELY(flow)) { dp_netdev_queue_batches(packet, flow, >mf, batches, n_batches); -- 2.4.11
[ovs-dev] [PATCH v4] dpif-netdev: Avoid reading RSS hash when EMC is disabled
When EMC is disabled the reading of RSS hash is skipped. Also, for packets that are not recirculated it retrieves the hash value without considering the recirc id. Signed-off-by: Antonio Fischetti--- V4 - reworked to remove dependencies from other patches in patchset "Skip EMC for recirc pkts and other optimizations." https://mail.openvswitch.org/pipermail/ovs-dev/2017-August/337320.html - measurements were repeated with the latest head of master. Port-to-Port Test = Software built with "-O2 -march=native -g". I measured the Rx rate regardless of pkt loss by sending 1 UDP flow, 64B packets, at the line-rate. 2 PMDs with 3 Tx queues. Flows setup: in_port=dpdk0 actions=output:dpdk1 in_port=dpdk1 actions=output:dpdk0 Results --- Values are for the Rx rate in Mpps, regardless of packet loss. RSS column: Yes: RSS hash is provided by the NIC No: RSS is disabled and the 5-tuple hash must be computed in software. Note: to simulate RSS disabled I added the line dp_packet_rss_valid(struct dp_packet *p) { +return false; #ifdef DPDK_NETDEV EMC column: Yes: default probability insertion, No: EMC disabled. Orig means Commit ID: 75f9e007e7f7eb91461e238f882d1c539c56bb8d +--++ +-+-+Orig | Orig + this| | RSS | EMC | |patch | +-+-+--++ | Yes | Yes | 12.02, 11.38 | 12.18, 11.32 | | Yes | No | 8.35, 8.36 | 8.36, 8.38 | +-+-+--++ | No | Yes | 9.91, 11.15 | 9.84, 11.08 | | No | No | 7.83, 7.87 | 8.33, 8.35 | +-+-+--++ --- lib/dpif-netdev.c | 32 1 file changed, 28 insertions(+), 4 deletions(-) diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index e2cd931..1157998 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -4567,6 +4567,22 @@ dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_, } static inline uint32_t +dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet *packet, +const struct miniflow *mf) +{ +uint32_t hash; + +if (OVS_LIKELY(dp_packet_rss_valid(packet))) { +hash = dp_packet_get_rss_hash(packet); +} else { +hash = miniflow_hash_5tuple(mf, 0); +dp_packet_set_rss_hash(packet, hash); +} + +return hash; +} + +static inline uint32_t dpif_netdev_packet_get_rss_hash(struct dp_packet *packet, const struct miniflow *mf) { @@ -4701,10 +4717,18 @@ emc_processing(struct dp_netdev_pmd_thread *pmd, } miniflow_extract(packet, >mf); key->len = 0; /* Not computed yet. */ -key->hash = dpif_netdev_packet_get_rss_hash(packet, >mf); - -/* If EMC is disabled skip emc_lookup */ -flow = (cur_min == 0) ? NULL: emc_lookup(flow_cache, key); +/* If EMC is disabled skip hash computation and emc_lookup */ +if (OVS_LIKELY(cur_min)) { +if (!md_is_valid) { +key->hash = dpif_netdev_packet_get_rss_hash_orig_pkt(packet, +>mf); +} else { +key->hash = dpif_netdev_packet_get_rss_hash(packet, >mf); +} +flow = emc_lookup(flow_cache, key); +} else { +flow = NULL; +} if (OVS_LIKELY(flow)) { dp_netdev_queue_batches(packet, flow, >mf, batches, n_batches); -- 2.4.11 ___ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev