Illumos #4045 write throttle & i/o scheduler performance work

4045 zfs write throttle & i/o scheduler performance work

1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync
read, sync write, async read, async write, and scrub/resilver.  The scheduler
issues a number of concurrent i/os from each class to the device.  Once a class
has been selected, an i/o is selected from this class using either an elevator
algorithem (async, scrub classes) or FIFO (sync classes).  The number of
concurrent async write i/os is tuned dynamically based on i/o load, to achieve
good sync i/o latency when there is not a high load of writes, and good write
throughput when there is.  See the block comment in vdev_queue.c (reproduced
below) for more details.

2. The write throttle (dsl_pool_tempreserve_space() and
txg_constrain_throughput()) is rewritten to produce much more consistent delays
when under constant load.  The new write throttle is based on the amount of
dirty data, rather than guesses about future performance of the system.  When
there is a lot of dirty data, each transaction (e.g. write() syscall) will be
delayed by the same small amount.  This eliminates the "brick wall of wait"
that the old write throttle could hit, causing all transactions to wait several
seconds until the next txg opens.  One of the keys to the new write throttle is
decrementing the amount of dirty data as i/o completes, rather than at the end
of spa_sync().  Note that the write throttle is only applied once the i/o
scheduler is issuing the maximum number of outstanding async writes.  See the
block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for
more details.

This diff has several other effects, including:

 * the commonly-tuned global variable zfs_vdev_max_pending has been removed;
use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead.

 * the size of each txg (meaning the amount of dirty data written, and thus the
time it takes to write out) is now controlled differently.  There is no longer
an explicit time goal; the primary determinant is amount of dirty data.
Systems that are under light or medium load will now often see that a txg is
always syncing, but the impact to performance (e.g. read latency) is minimal.
Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this.

 * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression,
checksum, etc.  This improves latency by not allowing these CPU-intensive tasks
to consume all CPU (on machines with at least 4 CPU's; the percentage is
rounded up).

--matt

APPENDIX: problems with the current i/o scheduler

The current ZFS i/o scheduler (vdev_queue.c) is deadline based.  The problem
with this is that if there are always i/os pending, then certain classes of
i/os can see very long delays.

For example, if there are always synchronous reads outstanding, then no async
writes will be serviced until they become "past due".  One symptom of this
situation is that each pass of the txg sync takes at least several seconds
(typically 3 seconds).

If many i/os become "past due" (their deadline is in the past), then we must
service all of these overdue i/os before any new i/os.  This happens when we
enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in
the future.  If we can't complete all the i/os in 2.5 seconds (e.g. because
there were always reads pending), then these i/os will become past due.  Now we
must service all the "async" writes (which could be hundreds of megabytes)
before we service any reads, introducing considerable latency to synchronous
i/os (reads or ZIL writes).

Notes on porting to ZFS on Linux:

- zio_t gained new members io_physdone and io_phys_children.  Because
  object caches in the Linux port call the constructor only once at
  allocation time, objects may contain residual data when retrieved
  from the cache. Therefore zio_create() was updated to zero out the two
  new fields.

- vdev_mirror_pending() relied on the depth of the per-vdev pending queue
  (vq->vq_pending_tree) to select the least-busy leaf vdev to read from.
  This tree has been replaced by vq->vq_active_tree which is now used
  for the same purpose.

- vdev_queue_init() used the value of zfs_vdev_max_pending to determine
  the number of vdev I/O buffers to pre-allocate.  That global no longer
  exists, so we instead use the sum of the *_max_active values for each of
  the five I/O classes described above.

- The Illumos implementation of dmu_tx_delay() delays a transaction by
  sleeping in condition variable embedded in the thread
  (curthread->t_delay_cv).  We do not have an equivalent CV to use in
  Linux, so this change replaced the delay logic with a wrapper called
  zfs_sleep_until(). This wrapper could be adopted upstream and in other
  downstream ports to abstract away operating system-specific delay logic.

- These tunables are added as module parameters, and descriptions added
  to the zfs-module-parameters.5 man page.

  spa_asize_inflation
  zfs_deadman_synctime_ms
  zfs_vdev_max_active
  zfs_vdev_async_write_active_min_dirty_percent
  zfs_vdev_async_write_active_max_dirty_percent
  zfs_vdev_async_read_max_active
  zfs_vdev_async_read_min_active
  zfs_vdev_async_write_max_active
  zfs_vdev_async_write_min_active
  zfs_vdev_scrub_max_active
  zfs_vdev_scrub_min_active
  zfs_vdev_sync_read_max_active
  zfs_vdev_sync_read_min_active
  zfs_vdev_sync_write_max_active
  zfs_vdev_sync_write_min_active
  zfs_dirty_data_max_percent
  zfs_delay_min_dirty_percent
  zfs_dirty_data_max_max_percent
  zfs_dirty_data_max
  zfs_dirty_data_max_max
  zfs_dirty_data_sync
  zfs_delay_scale

  The latter four have type unsigned long, whereas they are uint64_t in
  Illumos.  This accommodates Linux's module_param() supported types, but
  means they may overflow on 32-bit architectures.

  The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most
  likely to overflow on 32-bit systems, since they express physical RAM
  sizes in bytes.  In fact, Illumos initializes zfs_dirty_data_max_max to
  2^32 which does overflow. To resolve that, this port instead initializes
  it in arc_init() to 25% of physical RAM, and adds the tunable
  zfs_dirty_data_max_max_percent to override that percentage.  While this
  solution doesn't completely avoid the overflow issue, it should be a
  reasonable default for most systems, and the minority of affected
  systems can work around the issue by overriding the defaults.

- Fixed reversed logic in comment above zfs_delay_scale declaration.

- Clarified comments in vdev_queue.c regarding when per-queue minimums take
  effect.

- Replaced dmu_tx_write_limit in the dmu_tx kstat file
  with dmu_tx_dirty_delay and dmu_tx_dirty_over_max.  The first counts
  how many times a transaction has been delayed because the pool dirty
  data has exceeded zfs_delay_min_dirty_percent.  The latter counts how
  many times the pool dirty data has exceeded zfs_dirty_data_max (which
  we expect to never happen).

- The original patch would have regressed the bug fixed in
  zfsonlinux/zfs@c418410, which prevented users from setting the
  zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE.
  A similar fix is added to vdev_queue_aggregate().

- In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the
  heap instead of the stack.  In Linux we can't afford such large
  structures on the stack.

Reviewed by: George Wilson <[email protected]>
Reviewed by: Adam Leventhal <[email protected]>
Reviewed by: Christopher Siden <[email protected]>
Reviewed by: Ned Bass <[email protected]>
Reviewed by: Brendan Gregg <[email protected]>
Approved by: Robert Mustacchi <[email protected]>

References:
  http://www.illumos.org/issues/4045
  illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e

Ported-by: Ned Bass <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #1913

1 files changed, 197 insertions, 10 deletions

diff --git a/module/zfs/dmu_tx.c b/module/zfs/dmu_tx.c
index ece6b14b3..47cb86b08 100644
--- a/module/zfs/dmu_tx.c
+++ b/module/zfs/dmu_tx.c
@@ -53,7 +53,8 @@ dmu_tx_stats_t dmu_tx_stats = {
 	{ "dmu_tx_memory_reclaim",	KSTAT_DATA_UINT64 },
 	{ "dmu_tx_memory_inflight",	KSTAT_DATA_UINT64 },
 	{ "dmu_tx_dirty_throttle",	KSTAT_DATA_UINT64 },
-	{ "dmu_tx_write_limit",		KSTAT_DATA_UINT64 },
+	{ "dmu_tx_dirty_delay",		KSTAT_DATA_UINT64 },
+	{ "dmu_tx_dirty_over_max",	KSTAT_DATA_UINT64 },
 	{ "dmu_tx_quota",		KSTAT_DATA_UINT64 },
 };
 
@@ -70,6 +71,7 @@ dmu_tx_create_dd(dsl_dir_t *dd)
 	    offsetof(dmu_tx_hold_t, txh_node));
 	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
 	    offsetof(dmu_tx_callback_t, dcb_node));
+	tx->tx_start = gethrtime();
 #ifdef DEBUG_DMU_TX
 	refcount_create(&tx->tx_space_written);
 	refcount_create(&tx->tx_space_freed);
@@ -614,6 +616,7 @@ dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
 	if (txh == NULL)
 		return;
 	dn = txh->txh_dnode;
+	dmu_tx_count_dnode(txh);
 
 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
 		return;
@@ -931,6 +934,142 @@ dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
 }
 #endif
 
+/*
+ * If we can't do 10 iops, something is wrong.  Let us go ahead
+ * and hit zfs_dirty_data_max.
+ */
+hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
+int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
+
+/*
+ * We delay transactions when we've determined that the backend storage
+ * isn't able to accommodate the rate of incoming writes.
+ *
+ * If there is already a transaction waiting, we delay relative to when
+ * that transaction finishes waiting.  This way the calculated min_time
+ * is independent of the number of threads concurrently executing
+ * transactions.
+ *
+ * If we are the only waiter, wait relative to when the transaction
+ * started, rather than the current time.  This credits the transaction for
+ * "time already served", e.g. reading indirect blocks.
+ *
+ * The minimum time for a transaction to take is calculated as:
+ *     min_time = scale * (dirty - min) / (max - dirty)
+ *     min_time is then capped at zfs_delay_max_ns.
+ *
+ * The delay has two degrees of freedom that can be adjusted via tunables.
+ * The percentage of dirty data at which we start to delay is defined by
+ * zfs_delay_min_dirty_percent. This should typically be at or above
+ * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
+ * delay after writing at full speed has failed to keep up with the incoming
+ * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
+ * speaking, this variable determines the amount of delay at the midpoint of
+ * the curve.
+ *
+ * delay
+ *  10ms +-------------------------------------------------------------*+
+ *       |                                                             *|
+ *   9ms +                                                             *+
+ *       |                                                             *|
+ *   8ms +                                                             *+
+ *       |                                                            * |
+ *   7ms +                                                            * +
+ *       |                                                            * |
+ *   6ms +                                                            * +
+ *       |                                                            * |
+ *   5ms +                                                           *  +
+ *       |                                                           *  |
+ *   4ms +                                                           *  +
+ *       |                                                           *  |
+ *   3ms +                                                          *   +
+ *       |                                                          *   |
+ *   2ms +                                              (midpoint) *    +
+ *       |                                                  |    **     |
+ *   1ms +                                                  v ***       +
+ *       |             zfs_delay_scale ---------->     ********         |
+ *     0 +-------------------------------------*********----------------+
+ *       0%                    <- zfs_dirty_data_max ->               100%
+ *
+ * Note that since the delay is added to the outstanding time remaining on the
+ * most recent transaction, the delay is effectively the inverse of IOPS.
+ * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
+ * was chosen such that small changes in the amount of accumulated dirty data
+ * in the first 3/4 of the curve yield relatively small differences in the
+ * amount of delay.
+ *
+ * The effects can be easier to understand when the amount of delay is
+ * represented on a log scale:
+ *
+ * delay
+ * 100ms +-------------------------------------------------------------++
+ *       +                                                              +
+ *       |                                                              |
+ *       +                                                             *+
+ *  10ms +                                                             *+
+ *       +                                                           ** +
+ *       |                                              (midpoint)  **  |
+ *       +                                                  |     **    +
+ *   1ms +                                                  v ****      +
+ *       +             zfs_delay_scale ---------->        *****         +
+ *       |                                             ****             |
+ *       +                                          ****                +
+ * 100us +                                        **                    +
+ *       +                                       *                      +
+ *       |                                      *                       |
+ *       +                                     *                        +
+ *  10us +                                     *                        +
+ *       +                                                              +
+ *       |                                                              |
+ *       +                                                              +
+ *       +--------------------------------------------------------------+
+ *       0%                    <- zfs_dirty_data_max ->               100%
+ *
+ * Note here that only as the amount of dirty data approaches its limit does
+ * the delay start to increase rapidly. The goal of a properly tuned system
+ * should be to keep the amount of dirty data out of that range by first
+ * ensuring that the appropriate limits are set for the I/O scheduler to reach
+ * optimal throughput on the backend storage, and then by changing the value
+ * of zfs_delay_scale to increase the steepness of the curve.
+ */
+static void
+dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
+{
+	dsl_pool_t *dp = tx->tx_pool;
+	uint64_t delay_min_bytes =
+	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
+	hrtime_t wakeup, min_tx_time, now;
+
+	if (dirty <= delay_min_bytes)
+		return;
+
+	/*
+	 * The caller has already waited until we are under the max.
+	 * We make them pass us the amount of dirty data so we don't
+	 * have to handle the case of it being >= the max, which could
+	 * cause a divide-by-zero if it's == the max.
+	 */
+	ASSERT3U(dirty, <, zfs_dirty_data_max);
+
+	now = gethrtime();
+	min_tx_time = zfs_delay_scale *
+	    (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
+	min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
+	if (now > tx->tx_start + min_tx_time)
+		return;
+
+	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
+	    uint64_t, min_tx_time);
+
+	mutex_enter(&dp->dp_lock);
+	wakeup = MAX(tx->tx_start + min_tx_time,
+	    dp->dp_last_wakeup + min_tx_time);
+	dp->dp_last_wakeup = wakeup;
+	mutex_exit(&dp->dp_lock);
+
+	zfs_sleep_until(wakeup);
+}
+
 static int
 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
 {
@@ -965,6 +1104,13 @@ dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
 		return (SET_ERROR(ERESTART));
 	}
 
+	if (!tx->tx_waited &&
+	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
+		tx->tx_wait_dirty = B_TRUE;
+		DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
+		return (ERESTART);
+	}
+
 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
 	tx->tx_needassign_txh = NULL;
 
@@ -1092,6 +1238,10 @@ dmu_tx_unassign(dmu_tx_t *tx)
  *	blocking, returns immediately with ERESTART.  This should be used
  *	whenever you're holding locks.  On an ERESTART error, the caller
  *	should drop locks, do a dmu_tx_wait(tx), and try again.
+ *
+ * (3)	TXG_WAITED.  Like TXG_NOWAIT, but indicates that dmu_tx_wait()
+ *	has already been called on behalf of this operation (though
+ *	most likely on a different tx).
  */
 int
 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
@@ -1100,11 +1250,15 @@ dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
 	int err;
 
 	ASSERT(tx->tx_txg == 0);
-	ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT);
+	ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
+	    txg_how == TXG_WAITED);
 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
 
 	before = gethrtime();
 
+	if (txg_how == TXG_WAITED)
+		tx->tx_waited = B_TRUE;
+
 	/* If we might wait, we must not hold the config lock. */
 	ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
 
@@ -1128,17 +1282,47 @@ void
 dmu_tx_wait(dmu_tx_t *tx)
 {
 	spa_t *spa = tx->tx_pool->dp_spa;
+	dsl_pool_t *dp = tx->tx_pool;
 
 	ASSERT(tx->tx_txg == 0);
 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
 
-	/*
-	 * It's possible that the pool has become active after this thread
-	 * has tried to obtain a tx. If that's the case then his
-	 * tx_lasttried_txg would not have been assigned.
-	 */
-	if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
-		txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
+	if (tx->tx_wait_dirty) {
+		uint64_t dirty;
+
+		/*
+		 * dmu_tx_try_assign() has determined that we need to wait
+		 * because we've consumed much or all of the dirty buffer
+		 * space.
+		 */
+		mutex_enter(&dp->dp_lock);
+		if (dp->dp_dirty_total >= zfs_dirty_data_max)
+			DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
+		while (dp->dp_dirty_total >= zfs_dirty_data_max)
+			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
+		dirty = dp->dp_dirty_total;
+		mutex_exit(&dp->dp_lock);
+
+		dmu_tx_delay(tx, dirty);
+
+		tx->tx_wait_dirty = B_FALSE;
+
+		/*
+		 * Note: setting tx_waited only has effect if the caller
+		 * used TX_WAIT.  Otherwise they are going to destroy
+		 * this tx and try again.  The common case, zfs_write(),
+		 * uses TX_WAIT.
+		 */
+		tx->tx_waited = B_TRUE;
+	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
+		/*
+		 * If the pool is suspended we need to wait until it
+		 * is resumed.  Note that it's possible that the pool
+		 * has become active after this thread has tried to
+		 * obtain a tx.  If that's the case then tx_lasttried_txg
+		 * would not have been set.
+		 */
+		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
 	} else if (tx->tx_needassign_txh) {
 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
 
@@ -1148,6 +1332,10 @@ dmu_tx_wait(dmu_tx_t *tx)
 		mutex_exit(&dn->dn_mtx);
 		tx->tx_needassign_txh = NULL;
 	} else {
+		/*
+		 * A dnode is assigned to the quiescing txg.  Wait for its
+		 * transaction to complete.
+		 */
 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
 	}
 }
@@ -1268,7 +1456,6 @@ dmu_tx_pool(dmu_tx_t *tx)
 	return (tx->tx_pool);
 }
 
-
 void
 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
 {