Re: [Qemu-devel] [PATCH 1.8 5/6] qemu-img: add option to align writes to cluster_sectors during convert

[Peter Lieven <pl@kamp.de>]

On 25.11.2013 16:50, Paolo Bonzini wrote:
> Il 25/11/2013 16:32, Peter Lieven ha scritto:
>>> Also, a couple of ideas for separate patches.  Perhaps the default value
>>> of "-S" could be cluster_size if specified?  This would avoid making raw
>>> images too fragmented, and compounding filesystem-level fragmentation
>>> with qcow2-level fragmentation.  And 4K is too small a default in my
>>> opinion; it could be easily changed to 64K, though 4K was of course an
>>> improvement compared to 512 before commit a22f123 (qemu-img: Require
>>> larger zero areas for sparse handling, 2011-08-26).
>> I would vote for 64K or 256K, we already use the first for some time.
>> However, it turned out
>> that (much) bigger values decrease performance. Setting it
>> to cluster_size can be dangerous. As described in my case its 15MB and
>> I think for vhd its 1MB. This can be a lot of zeros that have to be
>> written.
> What about max(4096, min(bdi->cluster_size, 1048576))?
chaning sparse_size from 65536 to 1048576 about 5% performance decrease...

lieven@lieven-pc:~/git/qemu$ time ./qemu-img convert -pp -m 15728640 -S 1048576 /tmp/VC-Ubuntu-LTS-12.04.2-64bit.qcow2 iscsi://172.21.200.45/iqn.2001-05.com.equallogic:0-8a0906-9d95c510a-344001d54795289f-2012-r2-1-7-0/0
40980480 of 40980480 sectors converted.

real    0m29.263s
user    0m7.544s
sys    0m1.636s
lieven@lieven-pc:~/git/qemu$ time ./qemu-img convert -pp -m 15728640 -S 4096 /tmp/VC-Ubuntu-LTS-12.04.2-64bit.qcow2 iscsi://172.21.200.45/iqn.2001-05.com.equallogic:0-8a0906-9d95c510a-344001d54795289f-2012-r2-1-7-0/0
40980480 of 40980480 sectors converted.

real    0m28.169s
user    0m7.792s
sys    0m1.516s
lieven@lieven-pc:~/git/qemu$ time ./qemu-img convert -pp -m 15728640 -S 65536 /tmp/VC-Ubuntu-LTS-12.04.2-64bit.qcow2 iscsi://172.21.200.45/iqn.2001-05.com.equallogic:0-8a0906-9d95c510a-344001d54795289f-2012-r2-1-7-0/0
40980480 of 40980480 sectors converted.

real    0m27.643s
user    0m7.644s
sys    0m1.520s

i wouldn't go over 64k until we fully understand which impact it has.

Peter