| Message ID | cover.1602189832.git.josef@toxicpanda.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Improve preemptive ENOSPC flushing | expand |
On 8.10.20 г. 23:48 ч., Josef Bacik wrote: > There's a lot of individual changes, but most of it revolves around fixing the > O_DIRECT regression that Nikolay noted. With this set of patches we get > slightly better performance in the buffered case than before, and the O_DIRECT > case is slightly improved from baseline as well. > > v1->v2: > - Added a FORCE_COMMIT_TRANS flush operation so we can keep the flush_space > stuff consistent and get all the normal tracepoints. > - Renamed fs_info->dio_bytes to ->ordered_bytes and changed it to count all > ordered extents that were pending, not just DIO ordered extents that were > pending. > - Reworked the clamping to not apply if we're not doing a lot of delalloc > reservations. > - Reworked the preempt flushing loop to be more straightforward. > - Fixed the need_preemptive_flushing() helper to take into account DIO heavy > workloads. > <snip> So indeed it seems to result in slightly better results: dio-josef-v2: WRITE: bw=48.6MiB/s (50.0MB/s), 48.6MiB/s-48.6MiB/s (50.0MB/s-50.0MB/s), io=8192MiB (8590MB), run=168534-168534msec WRITE: bw=50.4MiB/s (52.8MB/s), 50.4MiB/s-50.4MiB/s (52.8MB/s-52.8MB/s), io=8192MiB (8590MB), run=162601-162601msec WRITE: bw=50.9MiB/s (53.4MB/s), 50.9MiB/s-50.9MiB/s (53.4MB/s-53.4MB/s), io=8192MiB (8590MB), run=160964-160964msec WRITE: bw=50.6MiB/s (53.0MB/s), 50.6MiB/s-50.6MiB/s (53.0MB/s-53.0MB/s), io=8192MiB (8590MB), run=161938-161938msec WRITE: bw=49.8MiB/s (52.2MB/s), 49.8MiB/s-49.8MiB/s (52.2MB/s-52.2MB/s), io=8192MiB (8590MB), run=164577-164577msec buffered-josef-v2: WRITE: bw=32.0MiB/s (33.6MB/s), 32.0MiB/s-32.0MiB/s (33.6MB/s-33.6MB/s), io=8192MiB (8590MB), run=255670-255670msec WRITE: bw=29.5MiB/s (30.9MB/s), 29.5MiB/s-29.5MiB/s (30.9MB/s-30.9MB/s), io=8192MiB (8590MB), run=277829-277829msec WRITE: bw=31.8MiB/s (33.4MB/s), 31.8MiB/s-31.8MiB/s (33.4MB/s-33.4MB/s), io=8192MiB (8590MB), run=257554-257554msec WRITE: bw=29.8MiB/s (31.3MB/s), 29.8MiB/s-29.8MiB/s (31.3MB/s-31.3MB/s), io=8192MiB (8590MB), run=274516-274516msec WRITE: bw=29.8MiB/s (31.2MB/s), 29.8MiB/s-29.8MiB/s (31.2MB/s-31.2MB/s), io=8192MiB (8590MB), run=274975-274975msec In comparison with V1 posting: buffered-josef-v1: WRITE: bw=29.1MiB/s (30.5MB/s), 29.1MiB/s-29.1MiB/s (30.5MB/s-30.5MB/s), io=8192MiB (8590MB), run=281678-281678msec WRITE: bw=30.0MiB/s (32.5MB/s), 30.0MiB/s-30.0MiB/s (32.5MB/s-32.5MB/s), io=8192MiB (8590MB), run=264337-264337msec WRITE: bw=29.6MiB/s (31.1MB/s), 29.6MiB/s-29.6MiB/s (31.1MB/s-31.1MB/s), io=8192MiB (8590MB), run=276312-276312msec WRITE: bw=29.8MiB/s (31.2MB/s), 29.8MiB/s-29.8MiB/s (31.2MB/s-31.2MB/s), io=8192MiB (8590MB), run=274916-274916msec WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269030-269030msec buffered-misc-next-no-josef: WRITE: bw=20.2MiB/s (21.2MB/s), 20.2MiB/s-20.2MiB/s (21.2MB/s-21.2MB/s), io=8192MiB (8590MB), run=404831-404831msec WRITE: bw=20.8MiB/s (21.8MB/s), 20.8MiB/s-20.8MiB/s (21.8MB/s-21.8MB/s), io=8192MiB (8590MB), run=394749-394749msec WRITE: bw=20.8MiB/s (21.8MB/s), 20.8MiB/s-20.8MiB/s (21.8MB/s-21.8MB/s), io=8192MiB (8590MB), run=393291-393291msec WRITE: bw=20.7MiB/s (21.8MB/s), 20.7MiB/s-20.7MiB/s (21.8MB/s-21.8MB/s), io=8192MiB (8590MB), run=394918-394918msec WRITE: bw=21.1MiB/s (22.1MB/s), 21.1MiB/s-21.1MiB/s (22.1MB/s-22.1MB/s), io=8192MiB (8590MB), run=388499-388499msec buffered-4.19.x: WRITE: bw=23.3MiB/s (24.4MB/s), 23.3MiB/s-23.3MiB/s (24.4MB/s-24.4MB/s), io=6387MiB (6697MB), run=274460-274460msec WRITE: bw=23.3MiB/s (24.5MB/s), 23.3MiB/s-23.3MiB/s (24.5MB/s-24.5MB/s), io=6643MiB (6966MB), run=284518-284518msec WRITE: bw=23.4MiB/s (24.5MB/s), 23.4MiB/s-23.4MiB/s (24.5MB/s-24.5MB/s), io=6643MiB (6966MB), run=284372-284372msec WRITE: bw=23.6MiB/s (24.7MB/s), 23.6MiB/s-23.6MiB/s (24.7MB/s-24.7MB/s), io=6387MiB (6697MB), run=271200-271200msec WRITE: bw=23.4MiB/s (24.6MB/s), 23.4MiB/s-23.4MiB/s (24.6MB/s-24.6MB/s), io=6387MiB (6697MB), run=272670-272670msec And dio: dio-josef-v1: WRITE: bw=47.1MiB/s (49.4MB/s), 47.1MiB/s-47.1MiB/s (49.4MB/s-49.4MB/s), io=8192MiB (8590MB), run=174049-174049msec WRITE: bw=48.5MiB/s (50.8MB/s), 48.5MiB/s-48.5MiB/s (50.8MB/s-50.8MB/s), io=8192MiB (8590MB), run=169045-169045msec WRITE: bw=45.0MiB/s (48.2MB/s), 45.0MiB/s-45.0MiB/s (48.2MB/s-48.2MB/s), io=8192MiB (8590MB), run=178196-178196msec WRITE: bw=46.1MiB/s (48.3MB/s), 46.1MiB/s-46.1MiB/s (48.3MB/s-48.3MB/s), io=8192MiB (8590MB), run=177861-177861msec WRITE: bw=46.4MiB/s (48.7MB/s), 46.4MiB/s-46.4MiB/s (48.7MB/s-48.7MB/s), io=8192MiB (8590MB), run=176376-176376msec dio-misc-next-no-josef: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163365-163365msec WRITE: bw=50.3MiB/s (52.8MB/s), 50.3MiB/s-50.3MiB/s (52.8MB/s-52.8MB/s), io=8192MiB (8590MB), run=162753-162753msec WRITE: bw=50.6MiB/s (53.1MB/s), 50.6MiB/s-50.6MiB/s (53.1MB/s-53.1MB/s), io=8192MiB (8590MB), run=161766-161766msec WRITE: bw=50.2MiB/s (52.7MB/s), 50.2MiB/s-50.2MiB/s (52.7MB/s-52.7MB/s), io=8192MiB (8590MB), run=163074-163074msec WRITE: bw=50.5MiB/s (52.9MB/s), 50.5MiB/s-50.5MiB/s (52.9MB/s-52.9MB/s), io=8192MiB (8590MB), run=162252-162252msec With this: Tested-by: Nikolay Borisov <nborisov@suse.com>
There's a lot of individual changes, but most of it revolves around fixing the O_DIRECT regression that Nikolay noted. With this set of patches we get slightly better performance in the buffered case than before, and the O_DIRECT case is slightly improved from baseline as well. v1->v2: - Added a FORCE_COMMIT_TRANS flush operation so we can keep the flush_space stuff consistent and get all the normal tracepoints. - Renamed fs_info->dio_bytes to ->ordered_bytes and changed it to count all ordered extents that were pending, not just DIO ordered extents that were pending. - Reworked the clamping to not apply if we're not doing a lot of delalloc reservations. - Reworked the preempt flushing loop to be more straightforward. - Fixed the need_preemptive_flushing() helper to take into account DIO heavy workloads. --- Original email --- A while ago Nikolay started digging into a problem where they were seeing an around 20% regression on random writes, and he bisected it down to btrfs: don't end the transaction for delayed refs in throttle However this wasn't actually the cause of the problem. This patch removed the code that would preemptively end the transactions if we were low on space. Because we had just introduced the ticketing code, this was no longer necessary and was causing a lot of transaction commits. And in Nikolay's testing he validated this, we would see like 100x more transaction commits without that patch than with it, but the write regression clearly appeared when this patch was applied. The root cause of this is that the transaction commits were essentially happening so quickly that we didn't end up needing to wait on space in the ENOSPC ticketing code as much, and thus were able to write pretty quickly. With this gone, we now were getting a sawtoothy sort of behavior where we'd run up, stop while we flushed metadata space, run some more, stop again etc. When I implemented the ticketing infrastructure, I was trying to get us out of excessively flushing space because we would sometimes over create block groups, and thus short circuited flushing if we no longer had tickets. This had the side effect of breaking the preemptive flushing code, where we attempted to flush space in the background before we were forced to wait for space. Enter this patchset. We still have some of this preemption logic sprinkled everywhere, so I've separated it out of the normal ticketed flushing code, and made preemptive flushing it's own thing. The preemptive flushing logic is more specialized than the standard flushing logic. It attempts to flush in whichever pool has the highest usage. This means that if most of our space is tied up in pinned extents, we'll commit the transaction. If most of the space is tied up in delalloc, we'll flush delalloc, etc. To test this out I used the fio job that Nikolay used, this needs to be adjusted so the overall IO size is at least 2x the RAM size for the box you are testing fio --direct=0 --ioengine=sync --thread --directory=/mnt/test --invalidate=1 \ --group_reporting=1 --runtime=300 --fallocate=none --ramp_time=10 \ --name=RandomWrites-async-64512-4k-4 --new_group --rw=randwrite \ --size=2g --numjobs=4 --bs=4k --fsync_on_close=0 --end_fsync=0 \ --filename_format=FioWorkloads.\$jobnum I got the following results misc-next:Josefbw=13.4MiB/s (14.0MB/s), 13.4MiB/s-13.4MiB/s (14.0MB/s-14.0MB/s), io=4015MiB (4210MB), run=300323-300323msec pre-throttling:Josefbw=16.9MiB/s (17.7MB/s), 16.9MiB/s-16.9MiB/s (17.7MB/s-17.7MB/s), io=5068MiB (5314MB), run=300069-300069msec my patches:Josefbw=18.0MiB/s (18.9MB/s), 18.0MiB/s-18.0MiB/s (18.9MB/s-18.9MB/s), io=5403MiB (5666MB), run=300001-300001msec Thanks, Josef Josef Bacik (11): btrfs: add a trace point for reserve tickets btrfs: track ordered bytes instead of just dio ordered bytes btrfs: introduce a FORCE_COMMIT_TRANS flush operation btrfs: improve preemptive background space flushing btrfs: rename need_do_async_reclaim btrfs: check reclaim_size in need_preemptive_reclaim btrfs: rework btrfs_calc_reclaim_metadata_size btrfs: simplify the logic in need_preemptive_flushing btrfs: implement space clamping for preemptive flushing btrfs: adjust the flush trace point to include the source btrfs: add a trace class for dumping the current ENOSPC state fs/btrfs/ctree.h | 4 +- fs/btrfs/disk-io.c | 9 +- fs/btrfs/ordered-data.c | 13 +- fs/btrfs/space-info.c | 269 +++++++++++++++++++++++++++++------ fs/btrfs/space-info.h | 3 + include/trace/events/btrfs.h | 104 +++++++++++++- 6 files changed, 341 insertions(+), 61 deletions(-)