Yeah you completely got the point. Thanks for the hint ;)
Alessio On Oct 4, 2017 12:11, "Damjan Marion" <dmarion.li...@gmail.com> wrote: > > > On 2 Oct 2017, at 18:14, Alessio Silvestro <ale.silver...@gmail.com> > wrote: > > Dear all, > > I am running VPP on a CPU with 2 sockets and 4 virtual cores. The startup > configuration is the following: > > unix { >> interactive >> nodaemon >> } >> >> cpu { >> main-core 0 >> corelist-workers 2-3 >> *workers 2* >> } >> >> dpdk { >> dev 0000:07:00.0 > > { >> num-rx-queues 2 >> } >> dev 0000:07:00.1 > > socket-mem 1024,1024 >> } > > > The thread placement is the following: > >> Thread 1 (vpp_wk_0 at lcore 2): >> *TenGigabitEthernet7/0/0 queue 0* >> TenGigabitEthernet7/0/1 queue 0 >> Thread 2 (vpp_wk_1 at lcore 3): >> * TenGigabitEthernet7/0/0 queue 1* > > > So, I can see that RSS is working because the first queue of the first > interface is on lcore2 whereas the second queue is on lcore3. > > I am running a simple L2-xconnect with the following command: > >> set int l2 xconnect TenGigabitEthernet7/0/0 TenGigabitEthernet7/0/1 > > > > I am expecting the second thread actually working when the first cannot > handle all the traffic. > > Therefore, I am saturating the RX bandwidth of TenGigabitEthernet7/0/0 sending > 14 Mpps (packets of 64B). > I am receiving on TenGigabitEthernet7/0/1 only ~13Mpps which means that I > am not able to process ~ 1Mpps. > > However, when I perform "show run", the second thread does not work at > all! > > Am I missing something in the configuration of VPP or what do you think > can be the cause? > > Thanks for the help :) > > > RSS is per-flow so likely your NIC thinks that all your traffic belongs to > the same flow and sends it to the same queue. > You need to make your traffic more diverse, i.e. by randomizing address or > port values… > > > > >
_______________________________________________ vpp-dev mailing list vpp-dev@lists.fd.io https://lists.fd.io/mailman/listinfo/vpp-dev