12.07.2020 20:49, Andrey Shinkevich wrote:
The script 'bench_write_req.py' allows comparing performances of write
request for two qemu-img binary files.
An example with (qemu-img binary 1) and without (qemu-img binary 2) the
applied patch "qcow2: skip writing zero buffers to empty COW areas"
(git commit ID: c8bb23cbdbe32f5) has the following results:
SSD:
----------------- ------------------- -------------------
<qemu-img binary 1> <qemu-img binary 2>
<simple case> 0.34 +- 0.01 10.57 +- 0.96
<general case> 0.33 +- 0.01 9.15 +- 0.85
<cluster middle> 0.33 +- 0.00 8.72 +- 0.05
<cluster overlap> 7.43 +- 1.19 14.35 +- 1.00
----------------- ------------------- -------------------
HDD:
----------------- ------------------- -------------------
<qemu-img binary 1> <qemu-img binary 2>
<simple case> 32.61 +- 1.17 55.11 +- 1.15
<general case> 54.28 +- 8.82 60.11 +- 2.76
<cluster middle> 57.93 +- 0.47 58.53 +- 0.51
<cluster overlap> 11.47 +- 0.94 17.29 +- 4.40
----------------- ------------------- -------------------
Suggested-by: Denis V. Lunev <d...@openvz.org>
Suggested-by: Vladimir Sementsov-Ogievskiy <vsement...@virtuozzo.com>
Signed-off-by: Andrey Shinkevich <andrey.shinkev...@virtuozzo.com>
Andrey wants to drop 02,03 in v5, so this patch is a candidate for v5. Below my
notes.
---
scripts/simplebench/bench_write_req.py | 173 +++++++++++++++++++++++++++++++++
1 file changed, 173 insertions(+)
create mode 100755 scripts/simplebench/bench_write_req.py
diff --git a/scripts/simplebench/bench_write_req.py
b/scripts/simplebench/bench_write_req.py
new file mode 100755
index 0000000..a285ef1
--- /dev/null
+++ b/scripts/simplebench/bench_write_req.py
@@ -0,0 +1,173 @@
+#!/usr/bin/env python3
+#
+# Test to compare performance of write requests for two qemu-img binary files.
Let's note that patch is intended to check benefit of c8bb23cbdbe
"qcow2: skip writing zero buffers to empty COW areas"
+#
+# Copyright (c) 2020 Virtuozzo International GmbH.
+#
+# This program is free software; you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation; either version 2 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program. If not, see <http://www.gnu.org/licenses/>.
+#
+
+
+import sys
+import os
+import subprocess
+import simplebench
+
+
+def bench_func(env, case):
+ """ Handle one "cell" of benchmarking table. """
+ return bench_write_req(env['qemu_img'], env['image_name'],
+ case['block_size'], case['block_offset'],
+ case['requests'])
+
+
+def qemu_img_pipe(*args):
+ '''Run qemu-img and return its output'''
+ subp = subprocess.Popen(list(args),
+ stdout=subprocess.PIPE,
+ stderr=subprocess.STDOUT,
+ universal_newlines=True)
+ exitcode = subp.wait()
+ if exitcode < 0:
+ sys.stderr.write('qemu-img received signal %i: %s\n'
+ % (-exitcode, ' '.join(list(args))))
+ return subp.communicate()[0]
+
+
+def bench_write_req(qemu_img, image_name, block_size, block_offset, requests):
+ """Benchmark write requests
+
+ The function creates a QCOW2 image with the given path/name and fills it
+ with random data optionally.
No, it doesn't fill..
+ Then it runs the 'qemu-img bench' command and
+ makes series of write requests on the image clusters. Finally, it returns
+ the total time of the write operations on the disk.
+
+ qemu_img -- path to qemu_img executable file
+ image_name -- QCOW2 image name to create
+ block_size -- size of a block to write to clusters
+ block_offset -- offset of the block in clusters
+ requests -- number of write requests per cluster
+
+ Returns {'seconds': int} on success and {'error': str} on failure.
+ Return value is compatible with simplebench lib.
+ """
+
+ if not os.path.isfile(qemu_img):
+ print(f'File not found: {qemu_img}')
+ sys.exit(1)
+
+ image_dir = os.path.dirname(os.path.abspath(image_name))
+ if not os.path.isdir(image_dir):
+ print(f'Path not found: {image_name}')
+ sys.exit(1)
+
+ cluster_size = 1024 * 1024
+ image_size = 1024 * cluster_size
+ seek = 4
+ dd_count = int(image_size / cluster_size) - seek
seek and dd_count are unused
+
+ args_create = [qemu_img, 'create', '-f', 'qcow2', '-o',
+ f'cluster_size={cluster_size}',
+ image_name, str(image_size)]
+
+ count = requests * int(image_size / cluster_size)
requests is number of requests per cluster..
+ step = str(cluster_size)
but step is one cluster. So, we have several requests per cluster, but still,
step is cluster?
It sounds strange to me. Assume requests = 5 and we have image with 5 clusters.
count would be 5 * 5 = 25. Trying to make 25 iterations with step=cluster will
go far beyond the image size..
+ offset = str(block_offset)
+ cnt = str(count)
extra variables. Let's just use str() in args below.
+ size = []
+ if block_size:
+ size = ['-s', f'{block_size}']
+
+ args_bench = [qemu_img, 'bench', '-w', '-n', '-t', 'none', '-c', cnt,
+ '-S', step, '-o', offset, '-f', 'qcow2', image_name]
+ if block_size:
+ args_bench.extend(size)
1. You may just write here
if block_size:
args_bench += ['-s', f'{block_size}']
No reason to create extra "size" variable to be used only here.
2. Why do you need this logic? If user pass block_size = 0, we instead rely on
default bufsize of img_bench() which is 4096. And in two test-cases you use
explicit 4096 constant, and in one you use 0 to be implicitly changed to same
4096. Let's instead specify block_size explicitly.
+
+ try:
+ qemu_img_pipe(*args_create)
+ except OSError as e:
+ os.remove(image_name)
+ return {'error': 'qemu_img create failed: ' + str(e)}
+
+ try:
+ ret = qemu_img_pipe(*args_bench)
+ finally:
+ os.remove(image_name)
+ if not ret:
ret may be not defined, if exception shot prior to assignment of ret, isn't it?
I suggest to not bother with "finally", and just make similar "except" like for
previous case, and then just parse ret at function end.
+ return {'error': 'qemu_img bench failed'}
+ if 'seconds' in ret:
+ ret_list = ret.split()
+ index = ret_list.index('seconds.')
+ return {'seconds': float(ret_list[index-1])}
+ else:
+ return {'error': 'qemu_img bench failed: ' + ret}
+
+
+if __name__ == '__main__':
+
+ if len(sys.argv) < 4:
+ program = os.path.basename(sys.argv[0])
+ print(f'USAGE: {program} <path to qemu-img binary file> '
+ '<path to another qemu-img to compare performance with> '
+ '<full or relative name for QCOW2 image to create>')
+ exit(1)
+
+ # Test-cases are "rows" in benchmark resulting table, 'id' is a caption
+ # for the row, other fields are handled by bench_func.
+ test_cases = [
+ {
+ 'id': '<simple case>',
+ 'block_size': 0,
+ 'block_offset': 0,
+ 'requests': 10
+ },
+ {
+ 'id': '<general case>',
+ 'block_size': 4096,
+ 'block_offset': 0,
+ 'requests': 10
+ },
Hmm I don't understand, why simple case and general case are different? As I
already noted, if you don't specify -s for bench, as you do if block_size is 0,
the default value is 4096 anyway in img_bench(). So, how there can be so
different results in commit message? Or what am I missing?
+ {
+ 'id': '<cluster middle>',
+ 'block_size': 4096,
+ 'block_offset': 524288,
+ 'requests': 10
+ },
+ {
+ 'id': '<cluster overlap>',
What is overlapping here? you just write half of cluster with a small offset
from start of cluster. I'm OK with the case itself, I just don't understand the
naming.
+ 'block_size': 524288,
+ 'block_offset': 4096,
+ 'requests': 2
+ },
+ ]
+
+ # Test-envs are "columns" in benchmark resulting table, 'id is a caption
+ # for the column, other fields are handled by bench_func.
+ # Set the paths below to desired values
+ test_envs = [
+ {
+ 'id': '<qemu-img binary 1>',
+ 'qemu_img': f'{sys.argv[1]}',
+ 'image_name': f'{sys.argv[3]}'
+ },
+ {
+ 'id': '<qemu-img binary 2>',
+ 'qemu_img': f'{sys.argv[2]}',
+ 'image_name': f'{sys.argv[3]}'
+ },
+ ]
+
+ result = simplebench.bench(bench_func, test_envs, test_cases, count=3,
+ initial_run=False)
+ print(simplebench.ascii(result))
--
Best regards,
Vladimir