[Qemu-devel] block: regression: savevm/delvm too slow

[Qemu-devel] block: regression: savevm/delvm too slow

[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 <kwolf@redhat.com>
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 <kwolf@redhat.com>

:040000 040000 83e364185d37845bb27f1dccd1249d14cc7a9a1e 0c91964a52b5869333d4fb2cb0fa83104151359e M	block

Re: [Qemu-devel] block: regression: savevm/delvm too slow

[Kevin Wolf]

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 <kwolf@redhat.com>
> 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 <kwolf@redhat.com>
> 
> :040000 040000 83e364185d37845bb27f1dccd1249d14cc7a9a1e 0c91964a52b5869333d4fb2cb0fa83104151359e M	block

Let me guess... You're using cache=writethrough?

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

Re: [Qemu-devel] block: regression: savevm/delvm too slow

[Luiz Capitulino]

On Wed, 22 Jun 2011 10:56:07 +0200
Kevin Wolf <kwolf@redhat.com> 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 <kwolf@redhat.com>
> > 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 <kwolf@redhat.com>
> > 
> > :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
>