On Fri, Apr 24, 2020 at 8:34 AM David Hildenbrand <da...@redhat.com> wrote: > > > >> Also, I do wonder if we want to default-enable it. It can still have a > >> negative performance impact and some people might not want that. > > > > The negative performance impact should be minimal. At this point the > > hinting process ends up being very light since it only processes a > > small chunk of the memory before it shuts down for a couple seconds. > > In my original data the only time any significant performance > > regression was seen was with page shuffling enabled, and that was > > before I put limits on how many pages we could process per pass and > > how frequently we would make a pass through memory. My thought is that > > we are much better having it enabled by default rather than having it > > disabled by default. In the worst case scenario we have a little bit > > of extra thread noise from it popping up running for a few > > milliseconds and then sleeping for about two seconds, but the benefit > > side is that the VMs will do us a favor and restrict themselves to a > > much smaller idle footprint until such time as the memory is actually > > needed in the guest. > > While I agree that the impact is small, there *is* a noticeable > performance impact. One of the main users is memory overcommit. > > Also, imagine the following scenarios: > - RT workload in the guest. Just because you have a virtio-balloon > device defined would mean something is suddenly active and > discarding/trying to discard pages. > - vfio: free page reporting is completely useless right now and > *only* overhead. > - prealloc does not expect that pages get suddenly discarded > - s390x, which has CMM and might not benefit at all (except when using > huge pages as backing storage) > > No, I don't think it's acceptable to enable this as default. I remember > that it is quite common to just define a balloon device but never use it. > > E.g., "A virtual memory balloon device is added to all Xen and KVM/QEMU > guest virtual machines. It will be seen as <memballoon> element" [1]. > > I think we should let upper layers decide (just as we do for free page > hinting, for example).
Okay. I guess there are a number of other cases I hand't thought of. I will switch the default to false. Thanks. - Alex
