On 08/17/2017 11:46 PM, Jan Scheurich wrote: > Hi Kevin, > >> @Jan, can you comment on length of time to take into consideration when >> re-balancing - when we talked before you had some views on it. I will post >> a v5 when we get consensus around that. > > When we originally discussed time scale, I was thinking about the automatic > rebalancing of rx queues in the event of sudden PMD congestion. In that case > a reaction time of 1 second seemed appropriate to minimize the loss of > packets and to base the new distribution on the actual load rather than some > outdated long-term average. > > Now that we are looking at a manual procedure (or a procedure triggered at > reconfiguration time), it would appear more suitable to look at an interval > of, say, 60 seconds. I wouldn't go beyond that. >
Thanks. I will change to 60 secs for this patchset, do some additional tests and re-submit. > If the overall load level is not too high, it doesn’t matter for re-balancing > if the queue load levels are not very accurate (we have been living with > assuming equal load levels in the past.) But if I manually trigger a > re-balance because one PMD is (close to being) congested, I want the current > queue load used for the procedure. Otherwise I risk that the new queue > distribution is worse than before. > > In the future, when we tackle the automatic rebalancing in case of overload > of individual PMDs, we can reduce the measurement time as needed. > > BR, Jan > _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
