On Thu, Jun 01, 2023 at 11:46:01AM -0400, Peter Xu wrote: > On Wed, May 31, 2023 at 02:24:00PM +0100, Daniel P. Berrangé wrote: > > The migration test cases that actually exercise live migration want to > > ensure there is a minimum of two iterations of pre-copy, in order to > > exercise the dirty tracking code. > > > > Historically we've queried the migration status, looking for the > > 'dirty-sync-count' value to increment to track iterations. This was > > not entirely reliable because often all the data would get transferred > > quickly enough that the migration would finish before we wanted it > > to. So we massively dropped the bandwidth and max downtime to > > guarantee non-convergance. This had the unfortunate side effect > > that every migration took at least 30 seconds to run (100 MB of > > dirty pages / 3 MB/sec). > > > > This optimization takes a different approach to ensuring that a > > mimimum of two iterations. Rather than waiting for dirty-sync-count > > to increment, directly look for an indication that the source VM > > has dirtied RAM that has already been transferred. > > > > On the source VM a magic marker is written just after the 3 MB > > offset. The destination VM is now montiored to detect when the > > magic marker is transferred. This gives a guarantee that the > > first 3 MB of memory have been transferred. Now the source VM > > memory is monitored at exactly the 3MB offset until we observe > > a flip in its value. This gives us a guaranteed that the guest > > workload has dirtied a byte that has already been transferred. > > > > Since we're looking at a place that is only 3 MB from the start > > of memory, with the 3 MB/sec bandwidth, this test should complete > > in 1 second, instead of 30 seconds. > > > > Once we've proved there is some dirty memory, migration can be > > set back to full speed for the remainder of the 1st iteration, > > and the entire of the second iteration at which point migration > > should be complete. > > > > On a test machine this further reduces the migration test time > > from 8 minutes to 1 minute 40. > > The outcome is definitely nice, but it does looks slightly hacky to me and > make the test slightly more complicated. > > If it's all about making sure we finish the 1st iteration, can we simply > add a src qemu parameter "switchover-hold"? If it's set, src never > switchover to dst but keeps the iterations.
For *most* of the tests, we want to ensure there are a minimum of 2 iterations. For the XBZRLE test we want to ensure there are a minimum of 3 iterations, so the cache gets workout. > Then migrate_ensure_non_converge() will be as simple as setting > switchover-hold to true. > > I am even thinking whether there can even be real-life use case for that, > e.g., where a user might want to have a pre-heat of a migration of some VM, > and trigger it immediately when the admin really wants (the pre-heats moved > most of the pages and keep doing so). > > It'll be also similar to what Avihai proposed here on switchover-ack, just > an ack mechanism on the src side: > > https://lore.kernel.org/r/20230530144821.1557-3-avih...@nvidia.com In general I strongly wanted to avoid adding special logic to the migration code that makes it directly synchronize with the test suite, because once we do that I don't think the test suite is a providing coverage of the real world usage scenario. IOW, if we add a switchover-ack feature, we should certainly have *a* test that uses it, but we should not modify other tests because we want to exercise the logic as it would run with apps that don't rely on switchover-ack. Also, this slow migration test is incredibly painful for people right now, so I'd like to see us get a speed up committed to git quickly. I don't want to block it on a feature proposal that might yet take months to merge. With regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
