jbd: improve fsync batching
Josef Bacik [Thu, 8 Jan 2009 02:07:24 +0000 (18:07 -0800)]
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>

fs/jbd/commit.c
fs/jbd/transaction.c
include/linux/jbd.h

index 25719d9..3fbffb1 100644 (file)
@@ -306,6 +306,8 @@ void journal_commit_transaction(journal_t *journal)
        int flags;
        int err;
        unsigned long blocknr;
+       ktime_t start_time;
+       u64 commit_time;
        char *tagp = NULL;
        journal_header_t *header;
        journal_block_tag_t *tag = NULL;
@@ -418,6 +420,7 @@ void journal_commit_transaction(journal_t *journal)
        commit_transaction->t_state = T_FLUSH;
        journal->j_committing_transaction = commit_transaction;
        journal->j_running_transaction = NULL;
+       start_time = ktime_get();
        commit_transaction->t_log_start = journal->j_head;
        wake_up(&journal->j_wait_transaction_locked);
        spin_unlock(&journal->j_state_lock);
@@ -913,6 +916,18 @@ restart_loop:
        J_ASSERT(commit_transaction == journal->j_committing_transaction);
        journal->j_commit_sequence = commit_transaction->t_tid;
        journal->j_committing_transaction = NULL;
+       commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
+
+       /*
+        * weight the commit time higher than the average time so we don't
+        * react too strongly to vast changes in commit time
+        */
+       if (likely(journal->j_average_commit_time))
+               journal->j_average_commit_time = (commit_time*3 +
+                               journal->j_average_commit_time) / 4;
+       else
+               journal->j_average_commit_time = commit_time;
+
        spin_unlock(&journal->j_state_lock);
 
        if (commit_transaction->t_checkpoint_list == NULL &&
index 60d4c32..b51fbd4 100644 (file)
@@ -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;
index 346e2b8..6384b19 100644 (file)
@@ -543,6 +543,11 @@ struct transaction_s
        unsigned long           t_expires;
 
        /*
+        * When this transaction started, in nanoseconds [no locking]
+        */
+       ktime_t                 t_start_time;
+
+       /*
         * How many handles used this transaction? [t_handle_lock]
         */
        int t_handle_count;
@@ -798,9 +803,19 @@ struct journal_s
        struct buffer_head      **j_wbuf;
        int                     j_wbufsize;
 
+       /*
+        * this is the pid of the last person to run a synchronous operation
+        * through the journal.
+        */
        pid_t                   j_last_sync_writer;
 
        /*
+        * the average amount of time in nanoseconds it takes to commit a
+        * transaction to the disk.  [j_state_lock]
+        */
+       u64                     j_average_commit_time;
+
+       /*
         * An opaque pointer to fs-private information.  ext3 puts its
         * superblock pointer here
         */