On 11/22/2016 03:23 AM, Alberto Garcia wrote: > On Mon 21 Nov 2016 09:08:06 PM CET, Eric Blake <ebl...@redhat.com> wrote: > >> Back to my comments earlier in the series - We may want to think about >> a parallel FIFO mode, where we kick off all reads, but then are >> prepared to cancel reads on later children once we have a positive >> answer on an earlier child. > > And what happens if the positive answer comes first from a later child? > Does COLO rely on the assumption that the first child is always read > first?
In parallel FIFO mode, it wouldn't matter if the second child has an answer first, you STILL have to wait for the first child's answer. It's just that if the first child fails, you have a headstart (or even an answer already) on the second child, rather than starting the second child from scratch. In other words, parallel mode wastes a lot of work when the first child is successful, but saves time if the first child fails. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature