[PATCH] jbd: fix transaction batching
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>
diff --git a/fs/jbd/transaction.c b/fs/jbd/transaction.c
index 429f4b2..ca91797 100644
--- a/fs/jbd/transaction.c
+++ b/fs/jbd/transaction.c
@@ -1308,6 +1308,7 @@
transaction_t *transaction = handle->h_transaction;
journal_t *journal = transaction->t_journal;
int old_handle_count, err;
+ pid_t pid;
J_ASSERT(transaction->t_updates > 0);
J_ASSERT(journal_current_handle() == handle);
@@ -1333,8 +1334,15 @@
* It doesn't cost much - we're about to run a commit and sleep
* on IO anyway. Speeds up many-threaded, many-dir operations
* by 30x or more...
+ *
+ * But don't do this if this process was the most recent one to
+ * perform a synchronous write. We do this to detect the case where a
+ * single process is doing a stream of sync writes. No point in waiting
+ * for joiners in that case.
*/
- if (handle->h_sync) {
+ pid = current->pid;
+ if (handle->h_sync && journal->j_last_sync_writer != pid) {
+ journal->j_last_sync_writer = pid;
do {
old_handle_count = transaction->t_handle_count;
schedule_timeout_uninterruptible(1);