[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] block: regression: savevm/delvm too slow
From: |
Luiz Capitulino |
Subject: |
Re: [Qemu-devel] block: regression: savevm/delvm too slow |
Date: |
Wed, 22 Jun 2011 09:44:19 -0300 |
On Wed, 22 Jun 2011 10:56:07 +0200
Kevin Wolf <address@hidden> wrote:
> Am 21.06.2011 19:29, schrieb Luiz Capitulino:
> > I'm getting the following times when doing a savevm and delvm in current
> > HEAD eb47d7c5d (time in minutes, each time corresponds to a savevm/delvm
> > run):
> >
> > savevm: 5:28m, 11:00m, 11:10m
> > delvm: 4:30m, 4:40m, > 15m
> >
> > Now, trying with qemu 0.13.0 I get:
> >
> > savevm: < 1:00m, 4:00m, 4:34m
> > delvm: few seconds for all cases
> >
> > Yes, you read it correctly, I tried with 0.13.0 because 0.14.0 also has the
> > bug. This is the pattern I see when I run strace against HEAD while running
> > the savevm command:
> >
> > pwrite(7,
> > "\0\1\0\1\0\1\0\1\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2"...,
> > 65536, 196608) = 65536
> > fdatasync(7) = 0
> >
> > Ie. a fdatasync() follows every single pwrite(). Something similar also
> > happens
> > with delvm. I don't see this pattern with 0.13.0.
> >
> > The good news is that I've tracked it down and Mr. git bisect says that:
> >
> > 29c1a7301af752de6721e031d31faa48887204bd is the first bad commit
> > commit 29c1a7301af752de6721e031d31faa48887204bd
> > Author: Kevin Wolf <address@hidden>
> > Date: Mon Jan 10 17:17:28 2011 +0100
> >
> > qcow2: Use QcowCache
> >
> > Use the new functions of qcow2-cache.c for everything that works on
> > refcount
> > block and L2 tables.
> >
> > Signed-off-by: Kevin Wolf <address@hidden>
> >
> > :040000 040000 83e364185d37845bb27f1dccd1249d14cc7a9a1e
> > 0c91964a52b5869333d4fb2cb0fa83104151359e M block
>
> Let me guess... You're using cache=writethrough?
I'm using default options. To do the tests I started qemu with:
# qemu -hda disks/test2.img -enable-kvm -m 1G -monitor stdio
>
> Previously, qcow2_update_snapshot_refcount() had implemented its own
> kind of writeback cache that would be used even with cache=writethrough.
> Now we're using the generic Qcow2Cache, which implements a write-through
> behaviour for cache=writethrough.
>
> We could fix this by temporarily switching the cache to writeback mode.
>
> Kevin
>