> -----Original Message----- > From: netdev-ow...@vger.kernel.org [mailto:netdev-ow...@vger.kernel.org] > Subject: RE: [RFC net-next 9/9] xfrm: add a small xdst pcpu cache > > > -----Original Message----- > > From: netdev-ow...@vger.kernel.org [mailto:netdev-ow...@vger.kernel.org] > > Subject: [RFC net-next 9/9] xfrm: add a small xdst pcpu cache > > > > retain last used xfrm_dst in a pcpu cache. > > On next request, reuse this dst if the policies are the same. > > > > The cache does'nt help at all with strictly-RR workloads as > > we never have a hit. > > > > Also, the cache adds cost of this_cpu_xchg() in packet path. > > It would be better to use plain this_cpu_read/write, however, > > a netdev notifier can run in parallel on other cpu and write same > > pcpu value so the xchg is needed to avoid race. > > > > The notifier is needed so we do not add long hangs when a device > > is dismantled but some pcpu xdst still holds a reference. > > > > Test results using 4 network namespaces and null encryption: > > > > ns1 ns2 -> ns3 -> ns4 > > netperf -> xfrm/null enc -> xfrm/null dec -> netserver > > > > what TCP_STREAM UDP_STREAM UDP_RR > > Flow cache: 14804.4 279.738 326213.0 > > No flow cache: 14158.3 257.458 228486.8 > > Pcpu cache: 14766.4 286.958 239433.5 > > > > UDP tests used 64byte packets, tests ran for one minute each, > > value is average over ten iterations. > > Hi Florian, > > I want to give this a go with hw-offload and see the impact on > performance. > It may take us a few days to do that.
Hi Florian, We tested with and without your patchset, using single SA with hw-crypto offload (RFC4106) IPv4 ESP tunnel mode, and a single netperf TCP_STREAM with a few different messages Sizes. We didn't separate the pcpu cache patch from the rest of the patchset. Here are the findings: What 64-byte 512-byte 1024-byte 1500-byte Flow cache 1602.89 11004.97 14634.46 14577.60 Pcpu cache 1513.38 10862.55 14246.94 14231.07 The overall degradation seems a bit more than what you measured with null-crypto. We used two machines and no namespaces. Ilan.
