On 23.07.21 14:51, Peter Xu wrote:
On Fri, Jul 23, 2021 at 09:50:18AM +0200, David Hildenbrand wrote:
On 22.07.21 19:41, Peter Xu wrote:
On Thu, Jul 22, 2021 at 04:51:48PM +0200, David Hildenbrand wrote:
I'll give it a churn.
Thanks, David.
Migration of a 8 GiB VM
* within the same host
* after Linux is up and idle
* free page hinting enabled
* after dirtying most VM memory using memhog
* keeping bandwidth set to QEMU defaults
* On my 16 GiB notebook with other stuff running
Current upstream with 63268c4970a, without this patch:
total time: 28606 ms
downtime: 33 ms
setup: 3 ms
transferred ram: 3722913 kbytes
throughput: 1066.37 mbps
remaining ram: 0 kbytes
total ram: 8389384 kbytes
duplicate: 21674 pages
skipped: 0 pages
normal: 928866 pages
normal bytes: 3715464 kbytes
dirty sync count: 5
pages-per-second: 32710
Current upstream without 63268c4970a, without this patch:
total time: 28530 ms
downtime: 277 ms
setup: 4 ms
transferred ram: 3726266 kbytes
throughput: 1070.21 mbps
remaining ram: 0 kbytes
total ram: 8389384 kbytes
duplicate: 21890 pages
skipped: 0 pages
normal: 929702 pages
normal bytes: 3718808 kbytes
dirty sync count: 5
pages-per-second: 32710
Current upstream without 63268c4970a, with this patch:
total time: 5115 ms
downtime: 37 ms
setup: 5 ms
transferred ram: 659532 kbytes
throughput: 1057.94 mbps
remaining ram: 0 kbytes
total ram: 8389384 kbytes
duplicate: 20748 pages
skipped: 0 pages
normal: 164516 pages
normal bytes: 658064 kbytes
dirty sync count: 4
pages-per-second: 32710
Current upstream with 63268c4970a, with this patch:
total time: 5205 ms
downtime: 45 ms
setup: 3 ms
transferred ram: 659636 kbytes
throughput: 1039.39 mbps
remaining ram: 0 kbytes
total ram: 8389384 kbytes
duplicate: 20264 pages
skipped: 0 pages
normal: 164543 pages
normal bytes: 658172 kbytes
dirty sync count: 4
pages-per-second: 32710
I repeated the last two measurements two times and took the "better"
results.
Looks like this patch does it job and that 63268c4970a doesn't seem to
degrade migration in this combination/setup significantly (if at all, we
would have to do more measurements).
Thanks again for helping!
Just to double check: the loop in qemu_guest_free_page_hint() won't run for a
lot of iterations, right? Looks like that only happens when over ramblock
boundaries. Otherwise we may also want to move that mutex out of the loop at
some point because atomic looks indeed expensive on huge hosts.
I'd expect it never ever happens.
--
Thanks,
David / dhildenb
