[PATCH v2 0/3] qcow2: relax subclusters allocation dependencies

[PATCH v2 0/3] qcow2: relax subclusters allocation dependencies

[Vladimir Sementsov-Ogievskiy]

Hi all!

v2:
01: improve documentation
02: add Hanna's and Eric's r-bs, add tiny grammar fix
03: fix test by filtering instead of reducing number of writes

Parallel small writes to unallocated cluster works bad when subclusters
enabled.

Look, without subclusters, one of write requests will allocate the whole
cluster, and all other writes to this cluster will be independent of
each other, they depend only on the first one that does allocation.

With subclusters, each write to unallocated subcluster will block the
whole cluster for parallel writing.

So, assume we write 8 consecutive 4k chunks in parallel:

Without subclusters, one of the chunks will block all the cluster and
write L2 entry. The remaining 7 chunks are written in parallel.

With subclusters, each of the chunks will allocate new subcluster and
block the whole cluster. All the chunks are dependent on each other and
queue depth becomes 1. That's not good.

Let's improve the situation.

Vladimir Sementsov-Ogievskiy (3):
  simplebench: add img_bench_templater.py
  qcow2: refactor handle_dependencies() loop body
  qcow2: handle_dependencies(): relax conflict detection

 block/qcow2-cluster.c                      | 60 +++++++++-----
 scripts/simplebench/img_bench_templater.py | 95 ++++++++++++++++++++++
 scripts/simplebench/table_templater.py     | 62 ++++++++++++++
 tests/qemu-iotests/271                     |  5 +-
 tests/qemu-iotests/271.out                 |  4 +-
 5 files changed, 202 insertions(+), 24 deletions(-)
 create mode 100755 scripts/simplebench/img_bench_templater.py
 create mode 100644 scripts/simplebench/table_templater.py

-- 
2.29.2

[PATCH v2 1/3] simplebench: add img_bench_templater.py

[Vladimir Sementsov-Ogievskiy]

Add simple grammar-parsing template benchmark. New tool consume test
template written in bash with some special grammar injections and
produces multiple tests, run them and finally print a performance
comparison table of different tests produced from one template.

Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
---
 scripts/simplebench/img_bench_templater.py | 95 ++++++++++++++++++++++
 scripts/simplebench/table_templater.py     | 62 ++++++++++++++
 2 files changed, 157 insertions(+)
 create mode 100755 scripts/simplebench/img_bench_templater.py
 create mode 100644 scripts/simplebench/table_templater.py

diff --git a/scripts/simplebench/img_bench_templater.py b/scripts/simplebench/img_bench_templater.py
new file mode 100755
index 0000000000..f8e1540ada
--- /dev/null
+++ b/scripts/simplebench/img_bench_templater.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+#
+# Process img-bench test templates
+#
+# Copyright (c) 2021 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 subprocess
+import re
+import json
+
+import simplebench
+from results_to_text import results_to_text
+from table_templater import Templater
+
+
+def bench_func(env, case):
+    test = templater.gen(env['data'], case['data'])
+
+    p = subprocess.run(test, shell=True, stdout=subprocess.PIPE,
+                       stderr=subprocess.STDOUT, universal_newlines=True)
+
+    if p.returncode == 0:
+        try:
+            m = re.search(r'Run completed in (\d+.\d+) seconds.', p.stdout)
+            return {'seconds': float(m.group(1))}
+        except Exception:
+            return {'error': f'failed to parse qemu-img output: {p.stdout}'}
+    else:
+        return {'error': f'qemu-img failed: {p.returncode}: {p.stdout}'}
+
+
+if __name__ == '__main__':
+    if len(sys.argv) > 1:
+        print("""
+Usage: img_bench_templater.py < path/to/test-template.sh
+
+This script generates performance tests from a test template (example below),
+runs them, and displays the results in a table. The template is read from
+stdin.  It must be written in bash and end with a `qemu-img bench` invocation
+(whose result is parsed to get the test instance’s result).
+
+Use the following syntax in the template to create the various different test
+instances:
+
+  column templating: {var1|var2|...} - test will use different values in
+  different columns. You may use several {} constructions in the test, in this
+  case product of all choice-sets will be used.
+
+  row templating: [var1|var2|...] - similar thing to define rows (test-cases)
+
+Test template example:
+
+Assume you want to compare two qemu-img binaries, called qemu-img-old and
+qemu-img-new in your build directory in two test-cases with 4K writes and 64K
+writes. The template may look like this:
+
+qemu_img=/path/to/qemu/build/qemu-img-{old|new}
+$qemu_img create -f qcow2 /ssd/x.qcow2 1G
+$qemu_img bench -c 100 -d 8 [-s 4K|-s 64K] -w -t none -n /ssd/x.qcow2
+
+When passing this to stdin of img_bench_templater.py, the resulting comparison
+table will contain two columns (for two binaries) and two rows (for two
+test-cases).
+
+In addition to displaying the results, script also stores results in JSON
+format into results.json file in current directory.
+""")
+        sys.exit()
+
+    templater = Templater(sys.stdin.read())
+
+    envs = [{'id': ' / '.join(x), 'data': x} for x in templater.columns]
+    cases = [{'id': ' / '.join(x), 'data': x} for x in templater.rows]
+
+    result = simplebench.bench(bench_func, envs, cases, count=5,
+                               initial_run=False)
+    print(results_to_text(result))
+    with open('results.json', 'w') as f:
+        json.dump(result, f, indent=4)
diff --git a/scripts/simplebench/table_templater.py b/scripts/simplebench/table_templater.py
new file mode 100644
index 0000000000..950f3b3024
--- /dev/null
+++ b/scripts/simplebench/table_templater.py
@@ -0,0 +1,62 @@
+# Parser for test templates
+#
+# Copyright (c) 2021 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 itertools
+from lark import Lark
+
+grammar = """
+start: ( text | column_switch | row_switch )+
+
+column_switch: "{" text ["|" text]+ "}"
+row_switch: "[" text ["|" text]+ "]"
+text: /[^|{}\[\]]+/
+"""
+
+parser = Lark(grammar)
+
+class Templater:
+    def __init__(self, template):
+        self.tree = parser.parse(template)
+
+        c_switches = []
+        r_switches = []
+        for x in self.tree.children:
+            if x.data == 'column_switch':
+                c_switches.append([el.children[0].value for el in x.children])
+            elif x.data == 'row_switch':
+                r_switches.append([el.children[0].value for el in x.children])
+
+        self.columns = list(itertools.product(*c_switches))
+        self.rows = list(itertools.product(*r_switches))
+
+    def gen(self, column, row):
+        i = 0
+        j = 0
+        result = []
+
+        for x in self.tree.children:
+            if x.data == 'text':
+                result.append(x.children[0].value)
+            elif x.data == 'column_switch':
+                result.append(column[i])
+                i += 1
+            elif x.data == 'row_switch':
+                result.append(row[j])
+                j += 1
+
+        return ''.join(result)
-- 
2.29.2

[PATCH v2 2/3] qcow2: refactor handle_dependencies() loop body

[Vladimir Sementsov-Ogievskiy]

No logic change, just prepare for the following commit. While being
here do also small grammar fix in a comment.

Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
---
 block/qcow2-cluster.c | 49 ++++++++++++++++++++++++-------------------
 1 file changed, 28 insertions(+), 21 deletions(-)

diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
index bd0597842f..9917e5c28c 100644
--- a/block/qcow2-cluster.c
+++ b/block/qcow2-cluster.c
@@ -1400,29 +1400,36 @@ static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
 
         if (end <= old_start || start >= old_end) {
             /* No intersection */
-        } else {
-            if (start < old_start) {
-                /* Stop at the start of a running allocation */
-                bytes = old_start - start;
-            } else {
-                bytes = 0;
-            }
+            continue;
+        }
 
-            /* Stop if already an l2meta exists. After yielding, it wouldn't
-             * be valid any more, so we'd have to clean up the old L2Metas
-             * and deal with requests depending on them before starting to
-             * gather new ones. Not worth the trouble. */
-            if (bytes == 0 && *m) {
-                *cur_bytes = 0;
-                return 0;
-            }
+        /* Conflict */
 
-            if (bytes == 0) {
-                /* Wait for the dependency to complete. We need to recheck
-                 * the free/allocated clusters when we continue. */
-                qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
-                return -EAGAIN;
-            }
+        if (start < old_start) {
+            /* Stop at the start of a running allocation */
+            bytes = old_start - start;
+        } else {
+            bytes = 0;
+        }
+
+        /*
+         * Stop if an l2meta already exists. After yielding, it wouldn't
+         * be valid any more, so we'd have to clean up the old L2Metas
+         * and deal with requests depending on them before starting to
+         * gather new ones. Not worth the trouble.
+         */
+        if (bytes == 0 && *m) {
+            *cur_bytes = 0;
+            return 0;
+        }
+
+        if (bytes == 0) {
+            /*
+             * Wait for the dependency to complete. We need to recheck
+             * the free/allocated clusters when we continue.
+             */
+            qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
+            return -EAGAIN;
         }
     }
 
-- 
2.29.2

[PATCH v2 3/3] qcow2: handle_dependencies(): relax conflict detection

[Vladimir Sementsov-Ogievskiy]

There is no conflict and no dependency if we have parallel writes to
different subclusters of one cluster when the cluster itself is already
allocated. So, relax extra dependency.

Measure performance:
First, prepare build/qemu-img-old and build/qemu-img-new images.

cd scripts/simplebench
./img_bench_templater.py

Paste the following to stdin of running script:

qemu_img=../../build/qemu-img-{old|new}
$qemu_img create -f qcow2 -o extended_l2=on /ssd/x.qcow2 1G
$qemu_img bench -c 100000 -d 8 [-s 2K|-s 2K -o 512|-s $((1024*2+512))] \
        -w -t none -n /ssd/x.qcow2

The result:

All results are in seconds

------------------  ---------  ---------
                    old        new
-s 2K               6.7 ± 15%  6.2 ± 12%
                                 -7%
-s 2K -o 512        13 ± 3%    11 ± 5%
                                 -16%
-s $((1024*2+512))  9.5 ± 4%   8.4
                                 -12%
------------------  ---------  ---------

So small writes are more independent now and that helps to keep deeper
io queue which improves performance.

271 iotest output becomes racy for three allocation in one cluster.
Second and third writes may finish in different order. Second and
third requests don't depend on each other any more. Still they both
depend on first request anyway. Filter out second and third write
offsets to cover both possible outputs.

Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
---
 block/qcow2-cluster.c      | 11 +++++++++++
 tests/qemu-iotests/271     |  5 ++++-
 tests/qemu-iotests/271.out |  4 ++--
 3 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
index 9917e5c28c..c1c43a891b 100644
--- a/block/qcow2-cluster.c
+++ b/block/qcow2-cluster.c
@@ -1403,6 +1403,17 @@ static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
             continue;
         }
 
+        if (old_alloc->keep_old_clusters &&
+            (end <= l2meta_cow_start(old_alloc) ||
+             start >= l2meta_cow_end(old_alloc)))
+        {
+            /*
+             * Clusters intersect but COW areas don't. And cluster itself is
+             * already allocated. So, there is no actual conflict.
+             */
+            continue;
+        }
+
         /* Conflict */
 
         if (start < old_start) {
diff --git a/tests/qemu-iotests/271 b/tests/qemu-iotests/271
index 599b849cc6..d9d391955e 100755
--- a/tests/qemu-iotests/271
+++ b/tests/qemu-iotests/271
@@ -893,7 +893,10 @@ EOF
 }
 
 _make_test_img -o extended_l2=on 1M
-_concurrent_io     | $QEMU_IO | _filter_qemu_io
+# Second an third writes in _concurrent_io() are independent and may finish in
+# different order. So, filter offset out to match both possible variants.
+_concurrent_io     | $QEMU_IO | _filter_qemu_io | \
+    $SED -e 's/\(20480\|40960\)/OFFSET/'
 _concurrent_verify | $QEMU_IO | _filter_qemu_io
 
 # success, all done
diff --git a/tests/qemu-iotests/271.out b/tests/qemu-iotests/271.out
index 81043ba4d7..5be780de76 100644
--- a/tests/qemu-iotests/271.out
+++ b/tests/qemu-iotests/271.out
@@ -719,8 +719,8 @@ blkdebug: Suspended request 'A'
 blkdebug: Resuming request 'A'
 wrote 2048/2048 bytes at offset 30720
 2 KiB, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
-wrote 2048/2048 bytes at offset 20480
+wrote 2048/2048 bytes at offset OFFSET
 2 KiB, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
-wrote 2048/2048 bytes at offset 40960
+wrote 2048/2048 bytes at offset OFFSET
 2 KiB, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
 *** done
-- 
2.29.2

Re: [PATCH v2 3/3] qcow2: handle_dependencies(): relax conflict detection

[Eric Blake]

On Tue, Aug 24, 2021 at 01:15:17PM +0300, Vladimir Sementsov-Ogievskiy wrote:
> There is no conflict and no dependency if we have parallel writes to
> different subclusters of one cluster when the cluster itself is already
> allocated. So, relax extra dependency.
> 
...
> So small writes are more independent now and that helps to keep deeper
> io queue which improves performance.
> 
> 271 iotest output becomes racy for three allocation in one cluster.
> Second and third writes may finish in different order. Second and
> third requests don't depend on each other any more. Still they both
> depend on first request anyway. Filter out second and third write
> offsets to cover both possible outputs.
> 
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> ---
>  block/qcow2-cluster.c      | 11 +++++++++++
>  tests/qemu-iotests/271     |  5 ++++-
>  tests/qemu-iotests/271.out |  4 ++--
>  3 files changed, 17 insertions(+), 3 deletions(-)
> 

> +++ b/tests/qemu-iotests/271
> @@ -893,7 +893,10 @@ EOF
>  }
>  
>  _make_test_img -o extended_l2=on 1M
> -_concurrent_io     | $QEMU_IO | _filter_qemu_io
> +# Second an third writes in _concurrent_io() are independent and may finish in

and

> +# different order. So, filter offset out to match both possible variants.
> +_concurrent_io     | $QEMU_IO | _filter_qemu_io | \
> +    $SED -e 's/\(20480\|40960\)/OFFSET/'
>  _concurrent_verify | $QEMU_IO | _filter_qemu_io
>

Reviewed-by: Eric Blake <eblake@redhat.com>

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org

Re: [PATCH v2 1/3] simplebench: add img_bench_templater.py

[Hanna Reitz]

On 24.08.21 12:15, Vladimir Sementsov-Ogievskiy wrote:
> Add simple grammar-parsing template benchmark. New tool consume test
> template written in bash with some special grammar injections and
> produces multiple tests, run them and finally print a performance
> comparison table of different tests produced from one template.
>
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> ---
>   scripts/simplebench/img_bench_templater.py | 95 ++++++++++++++++++++++
>   scripts/simplebench/table_templater.py     | 62 ++++++++++++++
>   2 files changed, 157 insertions(+)
>   create mode 100755 scripts/simplebench/img_bench_templater.py
>   create mode 100644 scripts/simplebench/table_templater.py

Reviewed-by: Hanna Reitz <hreitz@redhat.com>

Re: [PATCH v2 3/3] qcow2: handle_dependencies(): relax conflict detection

[Hanna Reitz]

On 24.08.21 12:15, Vladimir Sementsov-Ogievskiy wrote:
> There is no conflict and no dependency if we have parallel writes to
> different subclusters of one cluster when the cluster itself is already
> allocated. So, relax extra dependency.
>
> Measure performance:
> First, prepare build/qemu-img-old and build/qemu-img-new images.
>
> cd scripts/simplebench
> ./img_bench_templater.py
>
> Paste the following to stdin of running script:
>
> qemu_img=../../build/qemu-img-{old|new}
> $qemu_img create -f qcow2 -o extended_l2=on /ssd/x.qcow2 1G
> $qemu_img bench -c 100000 -d 8 [-s 2K|-s 2K -o 512|-s $((1024*2+512))] \
>          -w -t none -n /ssd/x.qcow2
>
> The result:
>
> All results are in seconds
>
> ------------------  ---------  ---------
>                      old        new
> -s 2K               6.7 ± 15%  6.2 ± 12%
>                                   -7%
> -s 2K -o 512        13 ± 3%    11 ± 5%
>                                   -16%
> -s $((1024*2+512))  9.5 ± 4%   8.4
>                                   -12%
> ------------------  ---------  ---------
>
> So small writes are more independent now and that helps to keep deeper
> io queue which improves performance.
>
> 271 iotest output becomes racy for three allocation in one cluster.
> Second and third writes may finish in different order. Second and
> third requests don't depend on each other any more. Still they both
> depend on first request anyway. Filter out second and third write
> offsets to cover both possible outputs.
>
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> ---
>   block/qcow2-cluster.c      | 11 +++++++++++
>   tests/qemu-iotests/271     |  5 ++++-
>   tests/qemu-iotests/271.out |  4 ++--
>   3 files changed, 17 insertions(+), 3 deletions(-)

[...]

> diff --git a/tests/qemu-iotests/271 b/tests/qemu-iotests/271
> index 599b849cc6..d9d391955e 100755
> --- a/tests/qemu-iotests/271
> +++ b/tests/qemu-iotests/271
> @@ -893,7 +893,10 @@ EOF
>   }
>   
>   _make_test_img -o extended_l2=on 1M
> -_concurrent_io     | $QEMU_IO | _filter_qemu_io
> +# Second an third writes in _concurrent_io() are independent and may finish in

s/ an / and /

With that fixed:

Reviewed-by: Hanna Reitz <hreitz@redhat.com>

> +# different order. So, filter offset out to match both possible variants.
> +_concurrent_io     | $QEMU_IO | _filter_qemu_io | \
> +    $SED -e 's/\(20480\|40960\)/OFFSET/'
>   _concurrent_verify | $QEMU_IO | _filter_qemu_io
>   
>   # success, all done

Re: [PATCH v2 0/3] qcow2: relax subclusters allocation dependencies

[Hanna Reitz]

On 24.08.21 12:15, Vladimir Sementsov-Ogievskiy wrote:
> Hi all!
>
> v2:
> 01: improve documentation
> 02: add Hanna's and Eric's r-bs, add tiny grammar fix
> 03: fix test by filtering instead of reducing number of writes

Thanks, I’ve fixed the typo in patch 3 and applied the series to my 
block branch:

https://github.com/XanClic/qemu/commits/block

Hanna