[ovs-dev] OVS DPDK NUMA pmd assignment question for physical port

[O Mahony, Billy]

Hi Wang,

Thanks for the figures. Unexpected results as you say. Two things come to mind:

I’m not sure what code you are using but the cycles per packet statistic was broken for a while recently. Ilya posted a patch to fix it so make sure you have that patch included.

Also remember to reset the pmd stats after you start your traffic and then measure after a short duration.

Regards,
Billy.



From: 王志克 [mailto:wangzhike at jd.com]
Sent: Friday, September 8, 2017 8:01 AM
To: Jan Scheurich <jan.scheurich at ericsson.com>; O Mahony, Billy <billy.o.mahony at intel.com>; Darrell Ball <dball at vmware.com>; ovs-discuss at openvswitch.org; ovs-dev at openvswitch.org; Kevin Traynor <ktraynor at redhat.com>
Subject: RE: [ovs-dev] OVS DPDK NUMA pmd assignment question for physical port


Hi All,



I tested below cases, and get some performance data. The data shows there is little impact for cross NUMA communication, which is different from my expectation. (Previously I mentioned that cross NUMA would add 60% cycles, but I can NOT reproduce it any more).



@Jan,

You mentioned cross NUMA communication would cost lots more cycles. Can you share your data? I am not sure whether I made some mistake or not.



@All,

Welcome your data if you have data for similar cases. Thanks.



Case1: VM0->PMD0->NIC0

Case2:VM1->PMD1->NIC0

Case3:VM1->PMD0->NIC0

Case4:NIC0->PMD0->VM0

Case5:NIC0->PMD1->VM1

Case6:NIC0->PMD0->VM1



　     VM Tx Mpps  Host Tx Mpps  avg cycles per packet       avg processing cycles per packet

Case1     1.4           1.4                 512                             415

Case2     1.3           1.3                 537                             436

Case3     1.35        1.35               514                             390



　  VM Rx Mpps    Host Rx Mpps  avg cycles per packet       avg processing cycles per packet

Case4     1.3       1.3                     549                             533

Case5     1.3       1.3                     559                             540

Case6     1.28     1.28                  568                             551



Br,

Wang Zhike



-----Original Message-----
From: Jan Scheurich [mailto:jan.scheurich at ericsson.com]
Sent: Wednesday, September 06, 2017 9:33 PM
To: O Mahony, Billy; 王志克; Darrell Ball; ovs-discuss at openvswitch.org<mailto:ovs-discuss at openvswitch.org>; ovs-dev at openvswitch.org<mailto:ovs-dev at openvswitch.org>; Kevin Traynor
Subject: RE: [ovs-dev] OVS DPDK NUMA pmd assignment question for physical port



Hi Billy,



> You are going to have to take the hit crossing the NUMA boundary at some point if your NIC and VM are on different NUMAs.

>

> So are you saying that it is more expensive to cross the NUMA boundary from the pmd to the VM that to cross it from the NIC to the

> PMD?



Indeed, that is the case: If the NIC crosses the QPI bus when storing packets in the remote NUMA there is no cost involved for the PMD. (The QPI bandwidth is typically not a bottleneck.) The PMD only performs local memory access.



On the other hand, if the PMD crosses the QPI when copying packets into a remote VM, there is a huge latency penalty involved, consuming lots of PMD cycles that cannot be spent on processing packets. We at Ericsson have observed exactly this behavior.



This latency penalty becomes even worse when the LLC cache hit rate is degraded due to LLC cache contention with real VNFs and/or unfavorable packet buffer re-use patterns as exhibited by real VNFs compared to typical synthetic benchmark apps like DPDK testpmd.



>

> If so then in that case you'd like to have two (for example) PMDs polling 2 queues on the same NIC. With the PMDs on each of the

> NUMA nodes forwarding to the VMs local to that NUMA?

>

> Of course your NIC would then also need to be able know which VM (or at least which NUMA the VM is on) in order to send the frame

> to the correct rxq.



That would indeed be optimal but hard to realize in the general case (e.g. with VXLAN encapsulation) as the actual destination is only known after tunnel pop. Here perhaps some probabilistic steering of RSS hash values based on measured distribution of final destinations might help in the future.



But even without that in place, we need PMDs on both NUMAs anyhow (for NUMA-aware polling of vhostuser ports), so why not use them to also poll remote eth ports. We can achieve better average performance with fewer PMDs than with the current limitation to NUMA-local polling.



BR, Jan