Dear Gopa,
> On 20 Nov 2019, at 04:35, Gopakumar Choorakkot Edakkunni > <gopakumar....@gmail.com> wrote: > > Hi, > > I understand from your architecture that threads have a complete copy of the > entire graph, to avoid locking etc.. Not sure what do you exactly mean by copy of the entire graph. We run independent instance of graph scheduler on each thread and each graph node code is executed on demand basis, when there is some real work for that node to do. Each graph node can maintain per-thread data structures to avoid locking or it can use locks if they can not be avoided. It is all about node developer judgement what is the right solution to use. > But these days with more and more high end networking boxes having offload > engines like QAT, I am interested to know wont that be a bottleneck ? And all those offload engines have multiple queues, like QAT so that actually should not be the problem, right? > The QAT hardware assist node would obviously have to be a single node shared > across threads, so wont that reduce the utility of having the entire graph > being a copy across threads (as opposed to the graph nodes being just shared > across threads) ? What do you mean by "node shared across threads"? Sure, node function code is shared between the threads. Data structures typically not... -- Damjan
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#14683): https://lists.fd.io/g/vpp-dev/message/14683 Mute This Topic: https://lists.fd.io/mt/61874479/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
