> -----Original Message----- > From: Thomas Monjalon <tho...@monjalon.net> > Sent: Thursday, March 25, 2021 1:23 AM > To: Jerin Jacob <jerinjac...@gmail.com>; David Marchand > <david.march...@redhat.com>; Rong, Leyi <leyi.r...@intel.com> > Cc: dev@dpdk.org; Zhang, Qi Z <qi.z.zh...@intel.com>; Richardson, Bruce > <bruce.richard...@intel.com>; Ananyev, Konstantin > <konstantin.anan...@intel.com> > Subject: Re: [dpdk-dev] [PATCH] examples/l3fwd: enable multiple Tx queues on > a lcore > > 05/11/2020 10:24, Rong, Leyi: > > From: Jerin Jacob <jerinjac...@gmail.com> > > > On Wed, Nov 4, 2020 at 2:34 PM Rong, Leyi <leyi.r...@intel.com> wrote: > > > > From: David Marchand <david.march...@redhat.com> > > > > > On Wed, Nov 4, 2020 at 9:34 AM Rong, Leyi <leyi.r...@intel.com> > wrote: > > > > > > As there always has thoughput limit for per queue, on some > > > > > > performance test case by using l3fwd, the result will limited > > > > > > by the per queue thoughput limit. With multiple Tx queue > > > > > > enabled, the per queue thoughput limit can be eliminated if > > > > > > the CPU core is not the bottleneck. > > > > > > > > > > Ah interesting. > > > > > Which nic has such limitations? > > > > > How much of an improvement can be expected from this? > > > > > > > > The initial found was on XXV710 25Gb NIC, but suppose such issue > > > > can happen on more NICs as the high-end CPU per core boundary is > > > > higher than many NICs(except 100Gb and above) per queue performance > boundary. > > > > The improvement can be about 1.8X with that case@1t2q. > > > > > > As far as I understand, the Current l3fwd Tx queue creation is like this: > > > If the app has N cores and M ports then l3fwd creates, N x M Tx > > > queues in total, What will be new values based on this patch? > > Thank you Jerin for providing some info missing in the description of the > patch. > > > Hi Jacob, > > > > Total queues number equals to queues per port multiply port number. > > Just take #l3fwd -l 5,6 -n 6 -- -p 0x3 --config > > '(0,0,5),(0,1,5),(1,0,6),(1,1,6)' as example, With this patch appied, > > totally 2x2=4 tx queues can be polled, while only > > 1x2=2 tx queues can be used before. > > It does not reply above question with N x M. > > > > Does this patch has any regression in case the NIC queues able to > > > cope up with the throughput limit from CPU. > > > > Regression test relevant with l3fwd passed with this patch, no obvious > > result drop on other cases. > > It does not reply the general question for all drivers you did not test. > > As you probably noticed, this patch is blocked for months because it is not > properly explained. >
Hi Thomas, This patch can be abandoned after synced with Konstantin months ago. And update the state to superseded on patchwork, Thanks!
