[ovs-dev] [PATCH v10 2/5] dpif-netdev: Count cycles on per-rxq basis.

Jan Scheurich jan.scheurich at ericsson.com
Mon Jan 15 08:08:33 UTC 2018


> >>> +    /* Update send cycles for all the rx queues evenly. */
> >>
> >> Just a query, is it right that this is distributed evenly?
> >> If there are more packets from one rx queue than another will it make a difference or will the cycles spent sending that batch be the
> same as a batch of a small or larger size?
> >>
> >
> > Maybe the "evenly" comment is a misleading? The send cycles for each rxq
> > is updated proportionally for how many packets that rxq had in the
> > batch. I think this is about the best that can be done.
> >
> 
> Yes, it means what Kevin said.
> I'm not sure about this comment. Should I change it?
> I'm also not sure if we need to describe distribution algorithm in comments.
> It's quiet obvious for me, but I, as an author, can not be objective.
> 
> How about something neutral like:
> /* Distribute the send cycles between the rx queues. */ ?

How about this:
/* Distribute send cycles evenly among transmitted packets and assign to their respective rx queues. */

This approximation is actually not too bad as all packets in the batch went to the same tx queue. The only error we make is to ignore the size of the packets (which may still be substantial, though, given that the packet copying cost for large packets is significant in real-world scenarios with lots of L3 cache misses and especially across the QPI bus).

BR, Jan


More information about the dev mailing list