Print this page
4045 zfs write throttle & i/o scheduler performance work
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
@@ -52,10 +52,11 @@
tx->tx_pool = dd->dd_pool;
list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
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 ZFS_DEBUG
refcount_create(&tx->tx_space_written);
refcount_create(&tx->tx_space_freed);
#endif
return (tx);
@@ -595,17 +596,17 @@
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
object, THT_FREE, off, len);
if (txh == NULL)
return;
dn = txh->txh_dnode;
+ dmu_tx_count_dnode(txh);
if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
return;
if (len == DMU_OBJECT_END)
len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
- dmu_tx_count_dnode(txh);
/*
* For i/o error checking, we read the first and last level-0
* blocks if they are not aligned, and all the level-1 blocks.
*
@@ -909,10 +910,160 @@
(u_longlong_t)db->db.db_object, db->db_level,
(u_longlong_t)db->db_blkid);
}
#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 = MSEC2NSEC(100);
+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);
+ if (now > tx->tx_start + min_tx_time)
+ return;
+
+ min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
+
+ 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);
+
+#ifdef _KERNEL
+ mutex_enter(&curthread->t_delay_lock);
+ while (cv_timedwait_hires(&curthread->t_delay_cv,
+ &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
+ CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
+ continue;
+ mutex_exit(&curthread->t_delay_lock);
+#else
+ hrtime_t delta = wakeup - gethrtime();
+ struct timespec ts;
+ ts.tv_sec = delta / NANOSEC;
+ ts.tv_nsec = delta % NANOSEC;
+ (void) nanosleep(&ts, NULL);
+#endif
+}
+
static int
dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
{
dmu_tx_hold_t *txh;
spa_t *spa = tx->tx_pool->dp_spa;
@@ -939,10 +1090,16 @@
return (SET_ERROR(EIO));
return (SET_ERROR(ERESTART));
}
+ if (!tx->tx_waited &&
+ dsl_pool_need_dirty_delay(tx->tx_pool)) {
+ tx->tx_wait_dirty = B_TRUE;
+ return (SET_ERROR(ERESTART));
+ }
+
tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
tx->tx_needassign_txh = NULL;
/*
* NB: No error returns are allowed after txg_hold_open, but
@@ -1063,23 +1220,31 @@
*
* (2) TXG_NOWAIT. If we can't assign into the current open txg without
* 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)
{
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));
/* If we might wait, we must not hold the config lock. */
ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
+ if (txg_how == TXG_WAITED)
+ tx->tx_waited = B_TRUE;
+
while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
dmu_tx_unassign(tx);
if (err != ERESTART || txg_how != TXG_WAIT)
return (err);
@@ -1094,22 +1259,52 @@
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));
+ if (tx->tx_wait_dirty) {
/*
- * 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.
+ * dmu_tx_try_assign() has determined that we need to wait
+ * because we've consumed much or all of the dirty buffer
+ * space.
*/
- if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
- txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
+ mutex_enter(&dp->dp_lock);
+ while (dp->dp_dirty_total >= zfs_dirty_data_max)
+ cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
+ uint64_t 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) {
+ /*
+ * A dnode is assigned to the quiescing txg. Wait for its
+ * transaction to complete.
+ */
dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
mutex_enter(&dn->dn_mtx);
while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
cv_wait(&dn->dn_notxholds, &dn->dn_mtx);