Alex Williamson wrote:
On Thu, 2010-06-24 at 09:04 -0600, Alex Williamson wrote:
On Thu, 2010-06-24 at 15:02 +0900, Yoshiaki Tamura wrote:
Hi Alex,
Is there additional overhead to save rams introduce by this series?
If so, how much?
Yes, there is overhead, but it's typically quite small. If I migrate a
1G VM immediately after I boot to a login prompt (lots of zero pages), I
get an overhead of 0.000076%. That's only 226 extra bytes over the
297164995 bytes otherwise transferred. If I build a kernel on the guest
and migrate during the compilation, the overhead is 0.000019%. The
overhead is tiny largely due to patch 12/15, which avoids sending the
block name if we're working within the same block as sent previously.
If I disable this optimization, the overhead goes up to 0.93% after boot
and 0.26% during a kernel compile.
Thank your for the detailed numbers and analysis!
If the overhead is at this level, I think it's worth introducing to support
migration with hot plug devices.
Note that an x86 VM does a separate qemu_ram_alloc for memory above 4G,
which means in bigger VMs we may end up needing to resend the ramblock
name once in a while as we bounce between above and below 4G. Worst
case for this could match the 0.26% above, but in my testing during a
kernel compile, this seems to increase the overhead to 0.000026% on a 6G
VM.
If we run a program which bounce between the region intentionally, we should get
a number close to 0.26%, but the overhead is still low enough, and it shouldn't
be a big problem.
I don't see any reason why we couldn't allocate all the ram in a
single qemu_ram_alloc call, so I'll add another patch to make that
change (which will also shorten the name to "pc.ram" for even less
overhead ;). Thanks,
Hmm. I didn't know about the qemu_ram_alloc over 4G issue.
If there isn't any reason, how about submitting a patch to fix it besides this
series?
FWIW, with this change, my migration during kernel compile on the 6G VM
seems to be running just at 0.000019%-0.000020%, so that eliminates the
penalty for bigger memory VMs. Thanks,
It make sense:-)
Thanks,
Yoshi
Alex
