Re: [Qemu-devel] [patch 0/5][v2] qcow2: improve I/O performance with cache=off

[Anthony Liguori <anthony@codemonkey.ws>]

Laurent Vivier wrote:
> These patches improve qcow2 performance when used with cache=off.
>
> They modify block-qcow2.c to read/write as many clusters as
> possible per bdrv_aio_[read|write]().
>   

This patch series looks like a pretty good clean up of the code.  The 
perf improvement is a nice side effect too.

I'm a little concerned about how much code this touches though.  How 
much testing have you done of these changes?  Have you tested all of the 
corner cases (backing files, filling up a disk image, etc.)?

Regards,

Anthony Liguori

> Some benchmarks:
>
> * mkfs on 500 MB qcow2 disk
>
> mkfs                            WITHOUT         WITH
>
> ide, cache=off,snapshot=off     41 s            5 s     8x faster
> ide, cache=off,snapshot=on      40 s            5 s     8x faster
> ide, cache=on, snapshot=off      3 s            3 s
> ide, cache=on, snapshot=on       4 s            3 s
>
> * tar jxvf linux-2.6.25.7.tar.bz2
>
> ide, cache=off,snapshot=off     847 s           379 s   2.2x faster
> ide, cache=off,snapshot=on      801 s           364 s   2.2x faster
> ide, cache=on, snapshot=off     238 s           236 s
> ide, cache=on, snapshot=on      236 s           237 s
>
> * dd if=/dev/zero of=file
>
> dd if=/dev/null                 WITHOUT         WITH
>
> ide, cache=off,snapshot=off     333 kB/s        3.7 MB/s  11.3x faster
> ide, cache=off,snapshot=on      337 kB/s        3.6 MB/s  10.9x faster
> ide, cache=on, snapshot=off    9.06 MB/s        9.23 MB/s
> ide, cache=on, snapshot=on     8,89 MB/s        8.89 MB/s
>
> * dbench 1
>
> dbench                          WITHOUT          WITH
>
> ide, cache=off,snapshot=off     20.8494 MB/sec  24.8521 MB/sec          +19,2%
> ide, cache=off,snapshot=on      20.9349 MB/sec  24.2296 MB/sec          +15,7%
> ide, cache=on, snapshot=off     23.6612 MB/sec  25.4724 MB/sec          + 7,6%
> ide, cache=on, snapshot=on      24.1836 MB/sec  24.8169 MB/sec
>
> [PATCH 1/5] extract code from get_cluster_offset() into new functions
>             seek_l2_table(), l2_load() and l2_allocate().
>
> [PATCH 2/5] divide get_cluster_offset() into get_cluster_offset() and
>             alloc_cluster_offset().
>
> [PATCH 3/5] divide alloc_cluster_offset() into alloc_cluster_offset()
>             and alloc_compressed_cluster_offset(). Factorize code
>             to free clusters into free_used_clusters().
>
> [PATCH 4/5] modify get_cluster_offset(), alloc_cluster_offset() and 
>             free_used_clusters() to specify how many clusters we
>             want.
>
> [PATCH 5/5] in free_used_clusters(), try to aggregate free clusters
>             and freed clusters.
>
> Laurent
> --
>
>
>