On Wed, May 11, 2022 at 12:03:24PM +0200, David Hildenbrand wrote: > On 11.05.22 11:34, Daniel P. Berrangé wrote: > > On Wed, May 11, 2022 at 11:31:23AM +0200, David Hildenbrand wrote: > >>>> Long story short, management application has no way of learning > >>>> TIDs of allocator threads so it can't make them run NUMA aware. > >>> > >>> This feels like the key issue. The preallocation threads are > >>> invisible to libvirt, regardless of whether we're doing coldplug > >>> or hotplug of memory-backends. Indeed the threads are invisible > >>> to all of QEMU, except the memory backend code. > >>> > >>> Conceptually we need 1 or more explicit worker threads, that we > >>> can assign CPU affinity to, and then QEMU can place jobs on them. > >>> I/O threads serve this role, but limited to blockdev work. We > >>> need a generalization of I/O threads, for arbitrary jobs that > >>> QEMU might want to farm out to specific numa nodes. > >> > >> At least the "-object iothread" thingy can already be used for actions > >> outside of blockdev. virtio-balloon uses one for free page hinting. > > > > Ah that's good to know, so my idea probably isn't so much work as > > I thought it might be. > > I guess we'd have to create a bunch of iothreads on the command line and > then feed them as an array to the memory backend we want to create. We > could then forward the threads to a new variant of os_mem_prealloc(). > > We could > > a) Allocate new iothreads for each memory backend we create. Hm, that > might be suboptimal, we could end up with many iothreads. > > b) Reuse iothreads and have separate sets of iothreads per NUMA node. > Assign them to a node once. > > c) Reuse iothreads and reassign them to NUMA nodes on demand.
If all we needs is NUMA affinity, not CPU affinity, then it would be sufficient to create 1 I/O thread per host NUMA node that the VM needs to use. The job running in the I/O can spawn further threads and inherit the NUMA affinity. This might be more clever than it is needed though. I expect creating/deleting I/O threads is cheap in comparison to the work done for preallocation. If libvirt is using -preconfig and object-add to create the memory backend, then we could have option of creating the I/O threads dynamically in -preconfig mode, create the memory backend, and then delete the I/O threads again. > However, I'm not sure what the semantics are when having multiple > backends referencing the iothreads ... Yep, we don't especially need an "ownership" relationship for what we want todo with preallocatino, specially because it is a one off point-in-time usage, not continuous usage as with block devices 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 :|
