Excerpts from Liu Bo's message of 2011-06-21 04:49:41 -0400: > I've been working to try to improve the write-ahead log's performance, > and I found that the bottleneck addresses in the checksum items, > especially when we want to make a random write on a large file, e.g a 4G file.
I spent some time last week on this code, because I really wanted to be able to include it. But I hit two problems. Recording the transid of the log tree root doesn't completely solve problems with later mounts expecting generation + 1. If an older kernel were to try and mount a log created by our new code, it wouldn't understand the transid and the mount would fail. I think we just need to force the transid of the root block to generation + 1. It is slightly less optimal but still much better than what we have. The second problem was that I consistently hit crashes during log replay after a crash. The test was just to use synctest: http://oss.oracle.com/~mason/synctest/ synctest -t 32 -f -F -u -n 100 /mnt I waited about 45 seconds and reset the machine. Later mounts would crash during log replay. -chris > > Then a idea for this suggested by Chris is to use sub transaction ids and just > to log the part of inode that had changed since either the last log commit or > the last transaction commit. And as we also push the sub transid into the > btree > blocks, we'll get much faster tree walks. As a result, we abandon the > original > brute force approach, which is "to delete all items of the inode in log", > to making sure we get the most uptodate copies of everything, and instead > we manage to "find and merge", i.e. finding extents in the log tree and > merging > in the new extents from the file. > > This patchset puts the above idea into code, and although the code is now more > complex, it brings us a great deal of performance improvement: > > in my sysbench "write + fsync" test: > > 451.01Kb/sec -> 4.3621Mb/sec > > In v2, thanks to Chris, we worked together to solve 2 bugs, and after that it > works as expected. > > Since there are some vital changes in recent rc, like "kill trans_mutex" and > "use cur_trans", as David asked, I rebase the patchset to the latest for-linus > branch. > > More tests are welcome! > > You can also get this patchset from: > > git://repo.or.cz/linux-btrfs-devel.git sub-trans > > Liu Bo (12): > Btrfs: introduce sub transaction stuff > Btrfs: update block generation if should_cow_block fails > Btrfs: modify btrfs_drop_extents API > Btrfs: introduce first sub trans > Btrfs: still update inode trans stuff when size remains unchanged > Btrfs: improve log with sub transaction > Btrfs: add checksum check for log > Btrfs: fix a bug of log check > Btrfs: kick off useless code > Btrfs: deal with EEXIST after iput > Btrfs: use the right generation number to read log_root_tree > Revert "Btrfs: do not flush csum items of unchanged file data during > treelog" > > fs/btrfs/btrfs_inode.h | 12 ++- > fs/btrfs/ctree.c | 69 +++++++++--- > fs/btrfs/ctree.h | 5 +- > fs/btrfs/disk-io.c | 12 +- > fs/btrfs/extent-tree.c | 10 +- > fs/btrfs/file.c | 22 ++--- > fs/btrfs/inode.c | 33 ++++--- > fs/btrfs/ioctl.c | 6 +- > fs/btrfs/relocation.c | 6 +- > fs/btrfs/transaction.c | 14 ++- > fs/btrfs/transaction.h | 19 +++- > fs/btrfs/tree-defrag.c | 2 +- > fs/btrfs/tree-log.c | 272 ++++++++++++++++++++++++++++++++++------------- > 13 files changed, 331 insertions(+), 151 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
