> From: RYAN D. MOATS <rmoats at us.ibm.com> > > This series of patches changes lflow_run from processing the > entire southbound logical flow table to doing incremental > processing during each run. > > To make this work, several pre-requisites are necessary, and > these are contained in the initial patches of the series.
This follow-up discusses scale testing results with this patch set. The scale test was to stamp out 2000 copies of the following openstack pattern: n1 --- r1 --- n2 and see how ovn-controller behaves with this patch set. The good news is that the patch set acts as advertised: when there are no new updates or if updates do not cause processing to be reset, then time spent in lflow_run is of the order of 4-5 msec per main control loop The bad news is two-fold: 1. this isn't enough as physical_run ends up spending multiple seconds each main control loop doing a full processing of bindings and multicast groups. 2. When examining processing time, it initially grows by something more than O(n) until ovn-northd saturates, at which point processing time becomes a stair step, as ovn-northd makes changes to the SB database. I plan on looking at the binding/multicast loops to see if they can benefit from incremental processing and instrumenting ovn-northd to see if there is some low hanging items there that would improve its performance. Ryan (regXboi) _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
