| Message ID | cover.1744822090.git.josef@toxicpanda.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | btrfs: simplify extent buffer writeback | expand |
On Wed, Apr 16, 2025 at 12:51:05PM -0400, Josef Bacik wrote: > Hello, > > We currently have two different paths for writing out extent buffers, a subpage > path and a normal path. This has resulted in subtle bugs with subpage code that > took us a while to figure out. Additionally we have this complex interaction of > get folio, find eb, see if we already started writing that eb out, write out the > eb. > > We already have a radix tree for our extent buffers, so we can use that > similarly to how pagecache uses the radix tree. Tag the buffers with DIRTY when > they're dirty, and WRITEBACK when we start writing them out. > > The unfortunate part is we have to re-implement folio_batch for extent buffers, > so that's where most of the new code comes from. The good part is we are now > down to a single path for writing out extent buffers, it's way simpler, and in > fact quite a bit faster now that we don't have all of these folio->eb > transitions to deal with. > > I ran this through fsperf on a VM with 8 CPUs and 16gib of ram. I used > smallfiles100k, but reduced the files to 1k to make it run faster, the > results are as follows, with the statistically significant improvements > marked with *, there were no regressions. fsperf was run with -n 10 for > both runs, so the baseline is the average 10 runs and the test is the > average of 10 runs. > > smallfiles100k results > metric baseline current stdev diff > ================================================================================ > avg_commit_ms 68.58 58.44 3.35 -14.79% * > commits 270.60 254.70 16.24 -5.88% > dev_read_iops 48 48 0 0.00% > dev_read_kbytes 1044 1044 0 0.00% > dev_write_iops 866117.90 850028.10 14292.20 -1.86% > dev_write_kbytes 10939976.40 10605701.20 351330.32 -3.06% > elapsed 49.30 33 1.64 -33.06% * > end_state_mount_ns 41251498.80 35773220.70 2531205.32 -13.28% * > end_state_umount_ns 1.90e+09 1.50e+09 14186226.85 -21.38% * > max_commit_ms 139 111.60 9.72 -19.71% * > sys_cpu 4.90 3.86 0.88 -21.29% > write_bw_bytes 42935768.20 64318451.10 1609415.05 49.80% * > write_clat_ns_mean 366431.69 243202.60 14161.98 -33.63% * > write_clat_ns_p50 49203.20 20992 264.40 -57.34% * > write_clat_ns_p99 827392 653721.60 65904.74 -20.99% * > write_io_kbytes 2035940 2035940 0 0.00% > write_iops 10482.37 15702.75 392.92 49.80% * > write_lat_ns_max 1.01e+08 90516129 3910102.06 -10.29% * > write_lat_ns_mean 366556.19 243308.48 14154.51 -33.62% * > > As you can see we get about a 33% decrease runtime, with a 50% > throughput increase, which is pretty significant. Thanks, Ignore this for now, the xarray<->radix thing isn't quite one to one, so I've got to convert the buffer radix to a proper xarray first. Thanks, Josef
Hello, We currently have two different paths for writing out extent buffers, a subpage path and a normal path. This has resulted in subtle bugs with subpage code that took us a while to figure out. Additionally we have this complex interaction of get folio, find eb, see if we already started writing that eb out, write out the eb. We already have a radix tree for our extent buffers, so we can use that similarly to how pagecache uses the radix tree. Tag the buffers with DIRTY when they're dirty, and WRITEBACK when we start writing them out. The unfortunate part is we have to re-implement folio_batch for extent buffers, so that's where most of the new code comes from. The good part is we are now down to a single path for writing out extent buffers, it's way simpler, and in fact quite a bit faster now that we don't have all of these folio->eb transitions to deal with. I ran this through fsperf on a VM with 8 CPUs and 16gib of ram. I used smallfiles100k, but reduced the files to 1k to make it run faster, the results are as follows, with the statistically significant improvements marked with *, there were no regressions. fsperf was run with -n 10 for both runs, so the baseline is the average 10 runs and the test is the average of 10 runs. smallfiles100k results metric baseline current stdev diff ================================================================================ avg_commit_ms 68.58 58.44 3.35 -14.79% * commits 270.60 254.70 16.24 -5.88% dev_read_iops 48 48 0 0.00% dev_read_kbytes 1044 1044 0 0.00% dev_write_iops 866117.90 850028.10 14292.20 -1.86% dev_write_kbytes 10939976.40 10605701.20 351330.32 -3.06% elapsed 49.30 33 1.64 -33.06% * end_state_mount_ns 41251498.80 35773220.70 2531205.32 -13.28% * end_state_umount_ns 1.90e+09 1.50e+09 14186226.85 -21.38% * max_commit_ms 139 111.60 9.72 -19.71% * sys_cpu 4.90 3.86 0.88 -21.29% write_bw_bytes 42935768.20 64318451.10 1609415.05 49.80% * write_clat_ns_mean 366431.69 243202.60 14161.98 -33.63% * write_clat_ns_p50 49203.20 20992 264.40 -57.34% * write_clat_ns_p99 827392 653721.60 65904.74 -20.99% * write_io_kbytes 2035940 2035940 0 0.00% write_iops 10482.37 15702.75 392.92 49.80% * write_lat_ns_max 1.01e+08 90516129 3910102.06 -10.29% * write_lat_ns_mean 366556.19 243308.48 14154.51 -33.62% * As you can see we get about a 33% decrease runtime, with a 50% throughput increase, which is pretty significant. Thanks, Josef Josef Bacik (2): btrfs: set DIRTY and WRITEBACK tags on the buffer_radix btrfs: use buffer radix for extent buffer writeback operations fs/btrfs/disk-io.c | 10 ++ fs/btrfs/extent_io.c | 385 ++++++++++++++++++++++------------------- fs/btrfs/extent_io.h | 1 + fs/btrfs/transaction.c | 5 +- 4 files changed, 224 insertions(+), 177 deletions(-)