Ben points out that: When writing files out using O_SYNC, jbd's 1 jiffy delay results in a significant drop in throughput as the disk sits idle. The patch below results in a 4-5x performance improvement (from 6.5MB/s to ~24-30MB/s on my IDE test box) when writing out files using O_SYNC. So optimise the batching code by omitting it entirely if the process which is doing a sync write is the same as the one which did the most recent sync write. If that's true, we're unlikely to get any other processes joining the transaction. (Has been in -mm for ages - it took me a long time to get on to performance testing it) Numbers, on write-cache-disabled IDE: /usr/bin/time -p synctest -n 10 -uf -t 1 -p 1 dir-name Unpatched: 40 seconds Patched: 35 seconds Batching disabled: 35 seconds This is the problematic single-process-doing-fsync case. With multiple fsyncing processes the numbers are AFACIT unaltered by the patch. Aside: performance testing and instrumentation shows that the transaction batching almost doesn't help (testing with synctest -n 1 -uf -t 100 -p 10 dir-name on non-writeback-caching IDE). This is because by the time one process is running a synchronous commit, a bunch of other processes already have a transaction handle open, so they're all going to batch into the same transaction anyway. The batching seems to offer maybe 5-10% speedup with this workload, but I'm pretty sure it was more important than that when it was first developed 4-odd years ago... Cc: "Stephen C. Tweedie" <sct@redhat.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org> |
||
---|---|---|
.. | ||
Makefile | ||
checkpoint.c | ||
commit.c | ||
journal.c | ||
recovery.c | ||
revoke.c | ||
transaction.c |