| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
txg_kick() fails to see that we are quiescing, forcing transactions to
their next stages without leaving them accumulate changes
Creating a fragmented pool in a DCenter VM and continuously writing to it with
multiple instances of randwritecomp, we get the following output from txg.d:
0ms 311MB in 4114ms (95% p1) 75MB/s 544MB (76%) 336us 153ms 0ms
0ms 8MB in 51ms ( 0% p1) 163MB/s 474MB (66%) 129us 34ms 0ms
0ms 366MB in 4454ms (93% p1) 82MB/s 572MB (79%) 498us 20ms 0ms
0ms 406MB in 5212ms (95% p1) 77MB/s 591MB (82%) 661us 37ms 0ms
0ms 340MB in 5110ms (94% p1) 66MB/s 622MB (86%) 1048us 41ms 1ms
0ms 3MB in 61ms ( 0% p1) 51MB/s 419MB (58%) 33us 0ms 0ms
0ms 361MB in 3555ms (88% p1) 101MB/s 542MB (75%) 335us 40ms 0ms
0ms 356MB in 4592ms (92% p1) 77MB/s 561MB (78%) 430us 89ms 1ms
0ms 11MB in 129ms (13% p1) 90MB/s 507MB (70%) 222us 15ms 0ms
0ms 281MB in 2520ms (89% p1) 111MB/s 542MB (75%) 334us 42ms 0ms
0ms 383MB in 3666ms (91% p1) 104MB/s 557MB (77%) 411us 133ms 0ms
0ms 404MB in 5757ms (94% p1) 70MB/s 635MB (88%) 1274us 123ms 2ms
4ms 367MB in 4172ms (89% p1) 88MB/s 556MB (77%) 401us 51ms 0ms
0ms 42MB in 470ms (44% p1) 90MB/s 557MB (77%) 412us 43ms 0ms
0ms 261MB in 2273ms (88% p1) 114MB/s 556MB (77%) 407us 27ms 0ms
0ms 394MB in 3646ms (85% p1) 108MB/s 552MB (77%) 393us 304ms 0ms
0ms 275MB in 2416ms (89% p1) 113MB/s 510MB (71%) 200us 53ms 0ms
0ms 9MB in 53ms ( 0% p1) 169MB/s 483MB (67%) 140us 100ms 1ms
The TXGs that are getting synced and don't have lots of changes are pushed by
txg_kick() which basically forces the current open txg to get to the quiesced
state:
if (tx->tx_syncing_txg == 0 &&
tx->tx_quiesce_txg_waiting <= tx->tx_open_txg &&
tx->tx_sync_txg_waiting <= tx->tx_synced_txg &&
tx->tx_quiesced_txg <= tx->tx_synced_txg) {
tx->tx_quiesce_txg_waiting = tx->tx_open_txg + 1;
cv_broadcast(&tx->tx_quiesce_more_cv);
}
The problem is that the above code doesn't check if we are currently quiescing
anything (only if a quiesce or a sync has been requested, ..etc) so the
following scenario can happen:
1] We have an open txg A that had enough dirty data (more than
zfs_dirty_data_sync) and it was pushed to the quiesced state, and opened
a new txg B. No txg is currently being synced.
2] Immediately after the opening of B, txg_kick() was run by some other write
(and because of A's dirty data) and saw that we are not currently syncing
any txg and no one has requested quiescing so it requests one by bumping
tx_quiesce_txg_waiting and broadcasts the quiesce thread.
3] The quiesce thread just passed txg A to be synced and sees that a quiescing
request has been sent to it so it immediately grabs B without letting it
gather enough data, putting it in a quiesced state and opening a new txg C.
In this scenario txg B, is an example of how the entries of interest show up in
the txg.d output.
Ideally we would like txg_kick() to get triggered only when we are sure that
we are not syncing AND not quiescing any txg. This way we can kick an open TXG
to the quiescing state when we are sure that there is nothing going on and we
would benefit from the different states running concurrently.
Authored by: Serapheim Dimitropoulos <[email protected]>
Reviewed by: Matt Ahrens <[email protected]>
Reviewed by: Brad Lewis <[email protected]>
Reviewed by: Andriy Gapon <[email protected]>
Approved by: Dan McDonald <[email protected]>
Ported-by: Brian Behlendorf <[email protected]>
OpenZFS-issue: https://illumos.org/issues/9464
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/1cd7635b
Closes #7587
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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
|
