jbd: improve fsync batching

There is a flaw with the way jbd handles fsync batching.  If we fsync() a
file and we were not the last person to run fsync() on this fs then we
automatically sleep for 1 jiffie in order to wait for new writers to join
into the transaction before forcing the commit.  The problem with this is
that with really fast storage (ie a Clariion) the time it takes to commit
a transaction to disk is way faster than 1 jiffie in most cases, so
sleeping means waiting longer with nothing to do than if we just committed
the transaction and kept going.  Ric Wheeler noticed this when using
fs_mark with more than 1 thread, the throughput would plummet as he added
more threads.

This patch attempts to fix this problem by recording the average time in
nanoseconds that it takes to commit a transaction to disk, and what time
we started the transaction.  If we run an fsync() and we have been running
for less time than it takes to commit the transaction to disk, we sleep
for the delta amount of time and then commit to disk.  We acheive
sub-jiffie sleeping using schedule_hrtimeout.  This means that the wait
time is auto-tuned to the speed of the underlying disk, instead of having
this static timeout.  I weighted the average according to somebody's
comments (Andreas Dilger I think) in order to help normalize random
outliers where we take way longer or way less time to commit than the
average.  I also have a min() check in there to make sure we don't sleep
longer than a jiffie in case our storage is super slow, this was requested
by Andrew.

I unfortunately do not have access to a Clariion, so I had to use a
ramdisk to represent a super fast array.  I tested with a SATA drive with
barrier=1 to make sure there was no regression with local disks, I tested
with a 4 way multipathed Apple Xserve RAID array and of course the
ramdisk.  I ran the following command

fs_mark -d /mnt/ext3-test -s 4096 -n 2000 -D 64 -t $i

where $i was 2, 4, 8, 16 and 32.  I mkfs'ed the fs each time.  Here are my
results

type	threads		with patch	without patch
sata	2		24.6		26.3
sata	4		49.2		48.1
sata	8		70.1		67.0
sata	16		104.0		94.1
sata	32		153.6		142.7

xserve	2		246.4		222.0
xserve	4		480.0		440.8
xserve	8		829.5		730.8
xserve	16		1172.7		1026.9
xserve	32		1816.3		1650.5

ramdisk	2		2538.3		1745.6
ramdisk	4		2942.3		661.9
ramdisk	8		2882.5		999.8
ramdisk	16		2738.7		1801.9
ramdisk	32		2541.9		2394.0

Signed-off-by: Josef Bacik <jbacik@redhat.com>
Cc: Andreas Dilger <adilger@sun.com>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Ric Wheeler <rwheeler@redhat.com>
Cc: <linux-ext4@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 33 insertions, 5 deletions

diff --git a/fs/jbd/transaction.c b/fs/jbd/transaction.c
index 60d4c32c880..b51fbd4b291 100644
--- a/fs/jbd/transaction.c
+++ b/fs/jbd/transaction.c
@@ -25,6 +25,7 @@
 #include <linux/timer.h>
 #include <linux/mm.h>
 #include <linux/highmem.h>
+#include <linux/hrtimer.h>
 
 static void __journal_temp_unlink_buffer(struct journal_head *jh);
 
@@ -49,6 +50,7 @@ get_transaction(journal_t *journal, transaction_t *transaction)
 {
 	transaction->t_journal = journal;
 	transaction->t_state = T_RUNNING;
+	transaction->t_start_time = ktime_get();
 	transaction->t_tid = journal->j_transaction_sequence++;
 	transaction->t_expires = jiffies + journal->j_commit_interval;
 	spin_lock_init(&transaction->t_handle_lock);
@@ -1370,7 +1372,7 @@ int journal_stop(handle_t *handle)
 {
 	transaction_t *transaction = handle->h_transaction;
 	journal_t *journal = transaction->t_journal;
-	int old_handle_count, err;
+	int err;
 	pid_t pid;
 
 	J_ASSERT(journal_current_handle() == handle);
@@ -1399,6 +1401,17 @@ int journal_stop(handle_t *handle)
 	 * on IO anyway.  Speeds up many-threaded, many-dir operations
 	 * by 30x or more...
 	 *
+	 * We try and optimize the sleep time against what the underlying disk
+	 * can do, instead of having a static sleep time.  This is usefull for
+	 * the case where our storage is so fast that it is more optimal to go
+	 * ahead and force a flush and wait for the transaction to be committed
+	 * than it is to wait for an arbitrary amount of time for new writers to
+	 * join the transaction.  We acheive this by measuring how long it takes
+	 * to commit a transaction, and compare it with how long this
+	 * transaction has been running, and if run time < commit time then we
+	 * sleep for the delta and commit.  This greatly helps super fast disks
+	 * that would see slowdowns as more threads started doing fsyncs.
+	 *
 	 * 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
@@ -1406,11 +1419,26 @@ int journal_stop(handle_t *handle)
 	 */
 	pid = current->pid;
 	if (handle->h_sync && journal->j_last_sync_writer != pid) {
+		u64 commit_time, trans_time;
+
 		journal->j_last_sync_writer = pid;
-		do {
-			old_handle_count = transaction->t_handle_count;
-			schedule_timeout_uninterruptible(1);
-		} while (old_handle_count != transaction->t_handle_count);
+
+		spin_lock(&journal->j_state_lock);
+		commit_time = journal->j_average_commit_time;
+		spin_unlock(&journal->j_state_lock);
+
+		trans_time = ktime_to_ns(ktime_sub(ktime_get(),
+						   transaction->t_start_time));
+
+		commit_time = min_t(u64, commit_time,
+				    1000*jiffies_to_usecs(1));
+
+		if (trans_time < commit_time) {
+			ktime_t expires = ktime_add_ns(ktime_get(),
+						       commit_time);
+			set_current_state(TASK_UNINTERRUPTIBLE);
+			schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
+		}
 	}
 
 	current->journal_info = NULL;