| Message ID | 1568108770-7918-1-git-send-email-vishal.deep.ajmera@ericsson.com |
|---|---|
| Headers | show |
| Series | Balance-tcp bond mode optimization | expand |
On Tue, Sep 10, 2019 at 11:53 AM Vishal Deep Ajmera <vishal.deep.ajmera@ericsson.com> wrote: > > v5->v6: > Addressed comments from Ilya Maximets. > https://mail.openvswitch.org/pipermail/ovs-dev/2019-August/362001.html > Rebased to OVS master. > > v4->v5: > Support for stats per hash bucket. > Support for dynamic load balancing. > Rebased to OVS Master. > > v3->v4: > Addressed Ilya Maximets comments. > https://mail.openvswitch.org/pipermail/ovs-dev/2019-July/360452.html > > v2->v3: > Rebased to OVS master. > Fixed git merge issue. > > v1->v2: > Updated datapath action to hash + lb-output. > Updated throughput test observations. > Rebased to OVS master. > > Vishal Deep Ajmera (1): > Avoid dp_hash recirculation for balance-tcp bond selection mode > > datapath/linux/compat/include/linux/openvswitch.h | 2 + > lib/dpif-netdev.c | 515 ++++++++++++++++++++-- > lib/dpif-netlink.c | 3 + > lib/dpif-provider.h | 8 + > lib/dpif.c | 48 ++ > lib/dpif.h | 7 + > lib/odp-execute.c | 2 + > lib/odp-util.c | 4 + > ofproto/bond.c | 52 ++- > ofproto/bond.h | 9 + > ofproto/ofproto-dpif-ipfix.c | 1 + > ofproto/ofproto-dpif-sflow.c | 1 + > ofproto/ofproto-dpif-xlate.c | 39 +- > ofproto/ofproto-dpif.c | 32 ++ > ofproto/ofproto-dpif.h | 12 +- > tests/lacp.at | 9 + > vswitchd/bridge.c | 4 + > vswitchd/vswitch.xml | 10 + > 18 files changed, 698 insertions(+), 60 deletions(-) > > -- > 1.9.1 > Hi, I confirm a decent performance improvement with DPDK and balance-tcp bonding: lb-output-action=false rx: 740 Mbps 1446 kpps lb-output-action=true rx: 860 Mbps 1680 kpps I'm running a very simple test with a tweaked version of testpmd which generates 256 L4 flows, I guess that with much flows the improvement is way higher. Tested-by: Matteo Croce <mcroce@redhat.com> -- Matteo Croce per aspera ad upstream
On Wed, Sep 11, 2019 at 8:37 PM Matteo Croce <mcroce@redhat.com> wrote: > > On Tue, Sep 10, 2019 at 11:53 AM Vishal Deep Ajmera > <vishal.deep.ajmera@ericsson.com> wrote: > > > > v5->v6: > > Addressed comments from Ilya Maximets. > > https://mail.openvswitch.org/pipermail/ovs-dev/2019-August/362001.html > > Rebased to OVS master. > > > > v4->v5: > > Support for stats per hash bucket. > > Support for dynamic load balancing. > > Rebased to OVS Master. > > > > v3->v4: > > Addressed Ilya Maximets comments. > > https://mail.openvswitch.org/pipermail/ovs-dev/2019-July/360452.html > > > > v2->v3: > > Rebased to OVS master. > > Fixed git merge issue. > > > > v1->v2: > > Updated datapath action to hash + lb-output. > > Updated throughput test observations. > > Rebased to OVS master. > > > > Vishal Deep Ajmera (1): > > Avoid dp_hash recirculation for balance-tcp bond selection mode > > > > datapath/linux/compat/include/linux/openvswitch.h | 2 + > > lib/dpif-netdev.c | 515 ++++++++++++++++++++-- > > lib/dpif-netlink.c | 3 + > > lib/dpif-provider.h | 8 + > > lib/dpif.c | 48 ++ > > lib/dpif.h | 7 + > > lib/odp-execute.c | 2 + > > lib/odp-util.c | 4 + > > ofproto/bond.c | 52 ++- > > ofproto/bond.h | 9 + > > ofproto/ofproto-dpif-ipfix.c | 1 + > > ofproto/ofproto-dpif-sflow.c | 1 + > > ofproto/ofproto-dpif-xlate.c | 39 +- > > ofproto/ofproto-dpif.c | 32 ++ > > ofproto/ofproto-dpif.h | 12 +- > > tests/lacp.at | 9 + > > vswitchd/bridge.c | 4 + > > vswitchd/vswitch.xml | 10 + > > 18 files changed, 698 insertions(+), 60 deletions(-) > > > > -- > > 1.9.1 > > > > Hi, > > I confirm a decent performance improvement with DPDK and balance-tcp bonding: > > lb-output-action=false > > rx: 740 Mbps 1446 kpps > > lb-output-action=true > > rx: 860 Mbps 1680 kpps > > I'm running a very simple test with a tweaked version of testpmd which > generates 256 L4 flows, I guess that with much flows the improvement > is way higher. > > Tested-by: Matteo Croce <mcroce@redhat.com> > > -- > Matteo Croce > per aspera ad upstream Hi, I found an issue with the patch. It's not 100% reproducible, but sometimes the option gets enabled regardless of the bonding type and the configuration. This breaks the unit tests, as the bond/show output is wrong: # ovs-vsctl add-br br0 [63129.589500] device ovs-system entered promiscuous mode [63129.591802] device br0 entered promiscuous mode # ovs-vsctl add-bond br0 bond dummy0 dummy1 -- set Port bond lacp=active bond-mode=active-backup [63150.700542] device dummy1 entered promiscuous mode [63150.700892] device dummy0 entered promiscuous mode # ovs-vsctl get port bond other_config {} # ovs-appctl bond/show ---- bond ---- bond_mode: active-backup bond may use recirculation: no, Recirc-ID : -1 bond-hash-basis: 0 lb-output-action: enabled updelay: 0 ms downdelay: 0 ms lacp_status: negotiated lacp_fallback_ab: false active slave mac: 00:00:00:00:00:00(none) slave dummy0: disabled may_enable: false slave dummy1: disabled may_enable: false Regards,
> > > > I confirm a decent performance improvement with DPDK and balance-tcp > bonding: > > > > lb-output-action=false > > > > rx: 740 Mbps 1446 kpps > > > > lb-output-action=true > > > > rx: 860 Mbps 1680 kpps > > > > I'm running a very simple test with a tweaked version of testpmd which > > generates 256 L4 flows, I guess that with much flows the improvement > > is way higher. > > > > Tested-by: Matteo Croce <mcroce@redhat.com> > > Thank you Matteo for testing this patch. > I found an issue with the patch. It's not 100% reproducible, but sometimes > the > option gets enabled regardless of the bonding type and the configuration. > This breaks the unit tests, as the bond/show output is wrong: > > # ovs-vsctl add-br br0 > [63129.589500] device ovs-system entered promiscuous mode [63129.591802] > device br0 entered promiscuous mode > > # ovs-vsctl add-bond br0 bond dummy0 dummy1 -- set Port bond lacp=active > bond-mode=active-backup [63150.700542] device dummy1 entered > promiscuous mode [63150.700892] device dummy0 entered promiscuous mode > Let me check this in my setup. I always used 'netdev' bridges for testing my patch. May be I need to be check for data path support in the display function as well.
> > > Let me check this in my setup. I always used 'netdev' bridges for testing my > patch. > May be I need to be check for data path support in the display function as > well. Hi, I have sent v7 version of patch fixing this issue. Warm Regards, Vishal Ajmera