Skip to contentSkip to conversation

ext4: Fix performance regression in writeback of random writes

PATCH Applied Jan Kara 3 participants 5 messages Sep 10, 2013 – Sep 16, 2013

Subsystems (2)
lore
Share & tools
lore

Jan Kara

author reviewer
#0 lore raw

PATCH 1/1 ext4: Fix performance regression in writeback of random writes

Linux Kernel Performance project guys have reported that commit 4e7ea81db5 introduces a performance regression for the following fio workload:
[global]
direct=0
ioengine=mmap
size=1500M
bs=4k
pre_read=1
numjobs=1
overwrite=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/
file_service_type=random:36
file_service_type=random:36

[job0]
startdelay=0
rw=randrw
filename=data0/f1:data0/f2

[job1]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1
...

[job7]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1

The culprit of the problem is that after the commit ext4_writepages() are more aggressive in writing back pages. Thus we have less consecutive dirty pages resulting in more seeking.

This increased aggressivity is caused by a bug in the condition terminating ext4_writepages(). We start writing from the beginning of the file even if we should have terminated ext4_writepages() because wbc->nr_to_write <= 0.

After fixing the condition the throughput of the fio workload is about 20% better than before writeback reorganization.

---
 fs/ext4/inode.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/ext4/inode.c +1 -1
index c79fd7d..7914c05 100644
@@ -2563,7 +2563,7 @@ retry:
 			break;
 	}
 	blk_finish_plug(&plug);
-	if (!ret && !cycled) {
+	if (!ret && !cycled && wbc->nr_to_write > 0) {
 		cycled = 1;
 		mpd.last_page = writeback_index - 1;
 		mpd.first_page = 0;
Signature · 1 line
1.8.1.4
#1 lore raw
Quoted 58 lines · Jan Kara

Linux Kernel Performance project guys have reported that commit 4e7ea81db5 introduces a performance regression for the following fio workload:
[global]
direct=0
ioengine=mmap
size=1500M
bs=4k
pre_read=1
numjobs=1
overwrite=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/
file_service_type=random:36
file_service_type=random:36

[job0]
startdelay=0
rw=randrw
filename=data0/f1:data0/f2

[job1]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1
...

[job7]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1

The culprit of the problem is that after the commit ext4_writepages() are more aggressive in writing back pages. Thus we have less consecutive dirty pages resulting in more seeking.

This increased aggressivity is caused by a bug in the condition terminating ext4_writepages(). We start writing from the beginning of the file even if we should have terminated ext4_writepages() because wbc->nr_to_write <= 0.

After fixing the condition the throughput of the fio workload is about 20% better than before writeback reorganization.

---
  fs/ext4/inode.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
fs/ext4/inode.c +0 -0
index c79fd7d..7914c05 100644

fs/ext4/inode.c retry: · line 2569

Quoted 8 lines
fs/ext4/inode.c +1 -1
@@ -2563,7 +2563,7 @@ retry:
  			break;
  	}
  	blk_finish_plug(&plug);
-	if (!ret && !cycled) {
+	if (!ret && !cycled && wbc->nr_to_write > 0) {
  		cycled = 1;
  		mpd.last_page = writeback_index - 1;
  		mpd.first_page = 0;

Interesting, doesn't that mean generic_writepages (sub-sequent write_cache_pages() ) and all other file systems implementing their own ->writepages() should be updated?

Thanks,
Bernd

author reviewer
#2 lore raw
Quoted 63 lines · Bernd Schubert
Quoted 62 lines · Jan Kara

Linux Kernel Performance project guys have reported that commit 4e7ea81db5 introduces a performance regression for the following fio workload:
[global]
direct=0
ioengine=mmap
size=1500M
bs=4k
pre_read=1
numjobs=1
overwrite=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/
file_service_type=random:36
file_service_type=random:36

[job0]
startdelay=0
rw=randrw
filename=data0/f1:data0/f2

[job1]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1
...

[job7]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1

The culprit of the problem is that after the commit ext4_writepages() are more aggressive in writing back pages. Thus we have less consecutive dirty pages resulting in more seeking.

This increased aggressivity is caused by a bug in the condition terminating ext4_writepages(). We start writing from the beginning of the file even if we should have terminated ext4_writepages() because wbc->nr_to_write <= 0.

After fixing the condition the throughput of the fio workload is about 20% better than before writeback reorganization.

---
 fs/ext4/inode.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/ext4/inode.c +1 -1
index c79fd7d..7914c05 100644
@@ -2563,7 +2563,7 @@ retry:
 			break;
 	}
 	blk_finish_plug(&plug);
-	if (!ret && !cycled) {
+	if (!ret && !cycled && wbc->nr_to_write > 0) {
Quoted 8 lines

fs/ext4/inode.c retry: · line 2569

Quoted 4 lines
fs/ext4/inode.c +0 -0
 		cycled = 1;
 		mpd.last_page = writeback_index - 1;
 		mpd.first_page = 0;

Interesting, doesn't that mean generic_writepages (sub-sequent write_cache_pages() ) and all other file systems implementing their own ->writepages() should be updated?

No. write_cache_pages() has the condition like:
if (!cycled && !done) {

and 'done' is set when wbc->nr_to_write drops to zero. So that function
is OK. We cannot use 'done' in ext4_writepages() because the functions are structured a bit differently and 'done' gets set also when reach end of file.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#3 lore raw
Quoted 72 lines · Jan Kara
Quoted 63 lines · Bernd Schubert
Quoted 62 lines · Jan Kara

Linux Kernel Performance project guys have reported that commit 4e7ea81db5 introduces a performance regression for the following fio workload:
[global]
direct=0
ioengine=mmap
size=1500M
bs=4k
pre_read=1
numjobs=1
overwrite=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/
file_service_type=random:36
file_service_type=random:36

[job0]
startdelay=0
rw=randrw
filename=data0/f1:data0/f2

[job1]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1
...

[job7]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1

The culprit of the problem is that after the commit ext4_writepages() are more aggressive in writing back pages. Thus we have less consecutive dirty pages resulting in more seeking.

This increased aggressivity is caused by a bug in the condition terminating ext4_writepages(). We start writing from the beginning of the file even if we should have terminated ext4_writepages() because wbc->nr_to_write <= 0.

After fixing the condition the throughput of the fio workload is about 20% better than before writeback reorganization.

---
  fs/ext4/inode.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
fs/ext4/inode.c +1 -1
index c79fd7d..7914c05 100644
@@ -2563,7 +2563,7 @@ retry:
  			break;
  	}
  	blk_finish_plug(&plug);
-	if (!ret && !cycled) {
+	if (!ret && !cycled && wbc->nr_to_write > 0) {
Quoted 8 lines

fs/ext4/inode.c retry: · line 2569

Quoted 4 lines
fs/ext4/inode.c +0 -0
  		cycled = 1;
  		mpd.last_page = writeback_index - 1;
  		mpd.first_page = 0;

Interesting, doesn't that mean generic_writepages (sub-sequent write_cache_pages() ) and all other file systems implementing their own ->writepages() should be updated?

Quoted 7 lines

No. write_cache_pages() has the condition like:
if (!cycled && !done) {

and 'done' is set when wbc->nr_to_write drops to zero. So that function
is OK. We cannot use 'done' in ext4_writepages() because the functions are structured a bit differently and 'done' gets set also when reach end of file.

Ah right, I missed that. If pagevec_lookup_tag() returns 0 there is still a way to avoid setting done=1, but I guess wbc->nr_to_write also wouldn't be zero then.
Btrfs' extent_write_cache_pages is another candidate and in combination with the additional blk plug ext4 and generic_writepages are doing, it might explain why I noticed extensive btrfs-raid6-rmw writes some time ago. I'm going to check that and further discuss on that list.

Thanks,
Bernd

maintainer Applied
#4 lore raw
In reply to Jan Kara
Quoted 2 lines · Jan Kara

Linux Kernel Performance project guys have reported that commit 4e7ea81db5 introduces a performance regression for the following fio workload:...

Applied, many thanks to Yan Zheng for finding this performance regression and to Jan for fixing it!

- Ted