Please find the comment inline below. *Regards*, Mrityunjay Kumar. Mobile: +91 - 9731528504
On Fri, Sep 3, 2021 at 9:52 PM Damjan Marion <dmar...@me.com> wrote: > > > On 03.09.2021., at 12:05, Mrityunjay Kumar <kumarn...@gmail.com> wrote: > > Damjan Hi, > > > I’m so sorry for pointing but I’d like to make sure I understood you > correctly. Since I don’t have specific case of tx-placement but please help > vpp-dev <https://lists.fd.io/g/vpp-dev/message/19192> mail readers. > · tx queues are statically mapped by vpp. > · main thread always maps to queue 0. of each interface in vpp. > · For the dpdk interfaces by default, number of vlib_mains [main > thread + workers] is equal to the number of tx queues if the dpdk driver > support such limit. > · Tx queues limit can be controlled by startup.conf section dpdk { > num-tx-queues > #abc }. But it might leads to spinlock on workers threads, refer the code. > if (xd->tx_q_used < tm->n_vlib_mains) > clib_spinlock_init (&vec_elt > (xd->tx_queues, j).lock); > > > So I think, we can’t generalised workers to tx queue mapping as because, > different vpp interface can have different number of tx queues. > > > If VPP have more worker threads than number of TX queues on the specific > interace, then VPP will share 1 queue between multiple workers for that > interface. > > With the new tx queue infra (not yet enabled for dpdk). we allow dynamic > mapping of tx queues and also sharing single queue between multiple > workers. I.e. you can have 8 workers sharing 4 queues (2:1 mapping). > [MJ]: Why should we implement some features to allow an user to configure less TX queues than worker threads? In case of hardware limitation, that's ok. If the number of threads is more than TX Queue , it leads to lock on thread. Being a decade old user of dpdk, we recommend a lock-less mechanism. [MJ]: Why do we think of implementing tx-placemats in VPP? If someone shares a use case, it will definitely help me to improve my knowledge. //MJ > — > Damjan > > > //MJ > > > *Regards*, > Mrityunjay Kumar. > Mobile: +91 - 9731528504 > > > > On Wed, Apr 14, 2021 at 4:54 PM Damjan Marion via lists.fd.io <dmarion= > me....@lists.fd.io> wrote: > >> >> >> On 14.04.2021., at 13:21, Marcos - Mgiga <mar...@mgiga.com.br> wrote: >> >> Hello, >> >> I increased VPP rx/tx queues in order to enable RSS on VPP instance. >> Since VPP is running on a NUMA system with two threads with 8 cores each, I >> would like to pin TX / RX queue to proper NUMA nodes. >> >> Using set interface rx-placement I was able to associate rx queue to >> desired cores, so I would like to know if is there any possibility to pin >> tx queue to a certain workers as well. >> >> >> >> Not at the moment. tx queues are statically mapped (0 to main thread, 1 >> to worker 0, 2 to worker 1, etc.). >> There are some plans to implement such capability….. >> >> >> >> >> >> > > > >
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#20068): https://lists.fd.io/g/vpp-dev/message/20068 Mute This Topic: https://lists.fd.io/mt/82088483/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
