On Wed, Aug 09, 2023 at 06:02:52PM +0300, gudkov.and...@huawei.com wrote: > > Not for the patch, I'm just curious about how the predication > > decides the network throughput, I mean QEMU predicts > > if migration will converge based on how fast it sends the data, > > not the actual bandwidth of the interface, which one the > > prediction uses? > > > Currently I use network nominal bandwidth, e.g. 1gbps. It would > be nice, of course, to use measured throughput since it can take > into account network headers overhead (as Wang Lei previously > mentioned), compression, etc., but it looks too complicated to > implement outside of migration process.
Using line speed is definitely wise, but qemu may be stupid enough to do wrong calculations and when it happens migration may not switchover.. See this: https://lore.kernel.org/r/20230803155344.11450-3-pet...@redhat.com Probably should be used together with your ways of prediction to be even more accurate on some workloads where mbps reports insane numbers. Thanks, -- Peter Xu
