mbox series

[v7,00/21] preallocate filter

Message ID 20201021145859.11201-1-vsementsov@virtuozzo.com
Headers show
Series preallocate filter | expand

Message

Vladimir Sementsov-Ogievskiy Oct. 21, 2020, 2:58 p.m. UTC 
Hi all!

Here is a filter, which does preallocation on write.

v7:
01: add Alberto's r-b
07: don't remove sentence from the comment
08: - drop extra "s->file_end = end;" line
    - improve check/set perm handlers
09: add Max's r-b
10: add Max's r-b
11: new
12: - skip if preallocate unsupported
    - drop auto and quick groups
13: new
14: - improve 'average' field of result spec
    - drop extra 'dim = ...' line
15-18: new
19: a lot of changes
20: new
21: add results dump to json

In Virtuozzo we have to deal with some custom distributed storage
solution, where allocation is very-very expensive operation. We have to
workaround it in Qemu, so here is a new filter.

Still, the filter shows good results for me even for xfs and ext4.

Here are results, produced by new benchmark (last several patches):

All results are in iops (larger means better)

----------------------------------  -------------  ------------
                                    no-prealloc    prealloc
ssd-ext4, aligned sequential 16k    2e+04          2.7e+04
                                                     +36%
hdd-ext4, aligned sequential 16k    4.8e+03 ± 5%   3.1e+03 ± 4%
                                                     -36%
ssd-xfs, aligned sequential 16k     1.6e+04 ± 5%   2.5e+04 ± 3%
                                                     +59%
hdd-xfs, aligned sequential 16k     4.7e+03 ± 22%  3.2e+03 ± 3%
                                                     -31%
ssd-ext4, unaligned sequential 64k  2e+04          2.7e+04
                                                     +36%
hdd-ext4, unaligned sequential 64k  4.9e+03 ± 2%   3.2e+03 ± 8%
                                                     -34%
ssd-xfs, unaligned sequential 64k   1.5e+04        2.6e+04 ± 2%
                                                     +69%
hdd-xfs, unaligned sequential 64k   5.2e+03 ± 2%   3.3e+03 ± 5%
                                                     -35%
----------------------------------  -------------  ------------

Note: it's on Fedora 30, kernel 5.6.13-100.fc30.x86_64

The tests are actually qemu-img bench, run like:

  ./qemu-img create -f qcow2 $img 16G

aligned:
  ./qemu-img bench -c 10000 -d 64 -f qcow2  -s 16k -t none -n -w $img

unaligned
  ./qemu-img bench -c 10000 -d 64 -f qcow2 -o 1k -s 64k -t none -n -w $img

and for preallocation, you'll drop -f qcow2, add --image-opts, and
instead of just $img use
  driver=qcow2,file.driver=preallocate,file.file.driver=file,file.file.filename=$img 

Or, use new benchmark like this:

  ./bench_prealloc.py ../../build/qemu-img \
      ssd-ext4:/path/to/mount/point \
      ssd-xfs:/path2 hdd-ext4:/path3 hdd-xfs:/path4

Vladimir Sementsov-Ogievskiy (21):
  block: simplify comment to BDRV_REQ_SERIALISING
  block/io.c: drop assertion on double waiting for request serialisation
  block/io: split out bdrv_find_conflicting_request
  block/io: bdrv_wait_serialising_requests_locked: drop extra bs arg
  block: bdrv_mark_request_serialising: split non-waiting function
  block: introduce BDRV_REQ_NO_WAIT flag
  block: bdrv_check_perm(): process children anyway
  block: introduce preallocate filter
  qemu-io: add preallocate mode parameter for truncate command
  iotests: qemu_io_silent: support --image-opts
  iotests.py: execute_setup_common(): add required_fmts argument
  iotests: add 298 to test new preallocate filter driver
  scripts/simplebench: fix grammar: s/successed/succeeded/
  scripts/simplebench: support iops
  scripts/simplebench: use standard deviation for +- error
  simplebench: rename ascii() to results_to_text()
  simplebench: move results_to_text() into separate file
  simplebench/results_to_text: improve view of the table
  simplebench/results_to_text: add difference line to the table
  simplebench/results_to_text: make executable
  scripts/simplebench: add bench_prealloc.py

 docs/system/qemu-block-drivers.rst.inc |  26 ++
 qapi/block-core.json                   |  20 +-
 include/block/block.h                  |  20 +-
 include/block/block_int.h              |   3 +-
 block.c                                |   7 +-
 block/file-posix.c                     |   2 +-
 block/io.c                             | 130 +++---
 block/preallocate.c                    | 559 +++++++++++++++++++++++++
 qemu-io-cmds.c                         |  46 +-
 block/meson.build                      |   1 +
 scripts/simplebench/bench-example.py   |   3 +-
 scripts/simplebench/bench_prealloc.py  | 132 ++++++
 scripts/simplebench/bench_write_req.py |   3 +-
 scripts/simplebench/results_to_text.py | 126 ++++++
 scripts/simplebench/simplebench.py     |  66 ++-
 tests/qemu-iotests/298                 | 186 ++++++++
 tests/qemu-iotests/298.out             |   5 +
 tests/qemu-iotests/group               |   1 +
 tests/qemu-iotests/iotests.py          |  16 +-
 19 files changed, 1225 insertions(+), 127 deletions(-)
 create mode 100644 block/preallocate.c
 create mode 100755 scripts/simplebench/bench_prealloc.py
 create mode 100755 scripts/simplebench/results_to_text.py
 create mode 100644 tests/qemu-iotests/298
 create mode 100644 tests/qemu-iotests/298.out

Comments

Max Reitz Nov. 13, 2020, 7:33 p.m. UTC | #1 
On 21.10.20 16:58, Vladimir Sementsov-Ogievskiy wrote:
> Hi all!
> 
> Here is a filter, which does preallocation on write.
> 
> v7:
> 01: add Alberto's r-b
> 07: don't remove sentence from the comment
> 08: - drop extra "s->file_end = end;" line
>      - improve check/set perm handlers
> 09: add Max's r-b
> 10: add Max's r-b
> 11: new
> 12: - skip if preallocate unsupported
>      - drop auto and quick groups
> 13: new
> 14: - improve 'average' field of result spec
>      - drop extra 'dim = ...' line
> 15-18: new
> 19: a lot of changes
> 20: new
> 21: add results dump to json

Thanks, applied to my block-next branch (for 6.0):

https://git.xanclic.moe/XanClic/qemu/commits/branch/block-next

Max
Vladimir Sementsov-Ogievskiy Nov. 16, 2020, 9:32 a.m. UTC | #2 
13.11.2020 22:33, Max Reitz wrote:
> On 21.10.20 16:58, Vladimir Sementsov-Ogievskiy wrote:
>> Hi all!
>>
>> Here is a filter, which does preallocation on write.
>>
>> v7:
>> 01: add Alberto's r-b
>> 07: don't remove sentence from the comment
>> 08: - drop extra "s->file_end = end;" line
>>      - improve check/set perm handlers
>> 09: add Max's r-b
>> 10: add Max's r-b
>> 11: new
>> 12: - skip if preallocate unsupported
>>      - drop auto and quick groups
>> 13: new
>> 14: - improve 'average' field of result spec
>>      - drop extra 'dim = ...' line
>> 15-18: new
>> 19: a lot of changes
>> 20: new
>> 21: add results dump to json
> 
> Thanks, applied to my block-next branch (for 6.0):
> 
> https://git.xanclic.moe/XanClic/qemu/commits/branch/block-next
> 

Thank you!