To maintain a strict limit on the metadata contained in the arc, while
preventing the arc buffer headers from completely consuming the
"arc_meta_used" space, we need to evict metadata buffers from the arc's
ghost lists along with the regular lists.
This change modifies arc_adjust_meta such that it more closely models
the adjustments made in arc_adjust. "arc_meta_used" is used similarly to
"arc_size", and "arc_meta_limit" is used similarly to "arc_c".
Testing metadata intensive workloads (e.g. creating, copying, and
removing millions of small files and/or directories) has shown this
change to make a dramatic improvement to the hit rate maintained in the
arc. While I think there is still room for improvement, this is a big
step in the right direction.
In addition, zpl_free_cached_objects was made into a no-op as I'm not
yet sure how to properly implement that function.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #2110
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 <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Ned Bass <bass6@llnl.gov>
Reviewed by: Brendan Gregg <brendan.gregg@joyent.com>
Approved by: Robert Mustacchi <rm@joyent.com>
References:
http://www.illumos.org/issues/4045illumos/illumos-gate@69962b5647
Ported-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#1913
3112 ztest does not honor ZFS_DEBUG
3113 ztest should use watchpoints to protect frozen arc bufs
3114 some leaked nvlists in zfsdev_ioctl
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Matt Amdur <Matt.Amdur@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Christopher Siden <chris.siden@delphix.com>
Approved by: Eric Schrock <eric.schrock@delphix.com>
References:
https://www.illumos.org/issues/3112https://www.illumos.org/issues/3113https://www.illumos.org/issues/3114illumos/illumos-gate@cd1c8b85eb
The /proc/self/cmd watchpoint interface is specific to Solaris.
Therefore, the #3113 implementation was reworked to use the more
portable mprotect(2) system call. When the pages are watched they
are marked read-only for protection. Any write to the protected
address range immediately trigger a SIGSEGV. The pages are marked
writable again when they are unwatched.
Ported-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1489
Currently there is no mechanism to inspect which dbufs are being
cached by the system. There are some coarse counters in arcstats
by they only give a rough idea of what's being cached. This patch
aims to improve the current situation by adding a new dbufs kstat.
When read this new kstat will walk all cached dbufs linked in to
the dbuf_hash. For each dbuf it will dump detailed information
about the buffer. It will also dump additional information about
the referenced arc buffer and its related dnode. This provides a
more complete view in to exactly what is being cached.
With this generic infrastructure in place utilities can be written
to post-process the data to understand exactly how the caching is
working. For example, the data could be processed to show a list
of all cached dnodes and how much space they're consuming. Or a
similar list could be generated based on dnode type. Many other
ways to interpret the data exist based on what kinds of questions
you're trying to answer.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Prakash Surya <surya1@llnl.gov>
3137 L2ARC compression
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Approved by: Dan McDonald <danmcd@nexenta.com>
References:
illumos/illumos-gate@aad02571bchttps://www.illumos.org/issues/3137http://wiki.illumos.org/display/illumos/L2ARC+Compression
Notes for Linux port:
A l2arc_nocompress module option was added to prevent the
compression of l2arc buffers regardless of how a dataset's
compression property is set. This allows the legacy behavior
to be preserved.
Ported by: James H <james@kagisoft.co.uk>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#1379
These functions are used in neither Illumos nor ZFSOnLinux. They appear
to have been replaced by arc_buf_alloc()/arc_buf_free(), so lets remove
them.
Signed-off-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1614
3805 arc shouldn't cache freed blocks
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Richard Elling <richard.elling@dey-sys.com>
Reviewed by: Will Andrews <will@firepipe.net>
Approved by: Dan McDonald <danmcd@nexenta.com>
References:
illumos/illumos-gate@6e6d5868f5https://www.illumos.org/issues/3805
ZFS should proactively evict freed blocks from the cache.
On dcenter, we saw that we were caching ~256GB of metadata, while the
pool only had <4GB of metadata on disk. We were wasting about half the
system's RAM (252GB) on blocks that have been freed.
Even though these freed blocks will never be used again, and thus will
eventually be evicted, this causes us to use memory inefficiently for 2
reasons:
1. A block that is freed has no chance of being accessed again, but will
be kept in memory preferentially to a block that was accessed before it
(and is thus older) but has not been freed and thus has at least some
chance of being accessed again.
2. We partition the ARC into several buckets:
user data that has been accessed only once (MRU)
metadata that has been accessed only once (MRU)
user data that has been accessed more than once (MFU)
metadata that has been accessed more than once (MFU)
The user data vs metadata split is somewhat arbitrary, and the primary
control on how much memory is used to cache data vs metadata is to
simply try to keep the proportion the same as it has been in the past
(each bucket "evicts against" itself). The secondary control is to
evict data before evicting metadata.
Because of this bucketing, we may end up with one bucket mostly
containing freed blocks that are very old, while another bucket has more
recently accessed, still-allocated blocks. Data in the useful bucket
(with still-allocated blocks) may be evicted in preference to data in
the useless bucket (with old, freed blocks).
On dcenter, we saw that the MFU metadata bucket was 230MB, while the MFU
data bucket was 27GB and the MRU metadata bucket was 256GB. However,
the vast majority of data in the MRU metadata bucket (256GB) was freed
blocks, and thus useless. Meanwhile, the MFU metadata bucket (230MB)
was constantly evicting useful blocks that will be soon needed.
The problem of cache segmentation is a larger problem that needs more
investigation. However, if we stop caching freed blocks, it should
reduce the impact of this more fundamental issue.
Ported-by: Richard Yao <ryao@cs.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#1503
Commit c409e4647f introduced a
number of module parameters. This required several types to be
changed to accomidate the required module parameters Linux macros.
Unfortunately, arc.c contained its own extern definition of the
zfs_write_limit_max variable and its type was not updated to be
consistent with its dsl_pool.c counterpart. If the variable had
been properly marked extern in a common header, then gcc would
have generated a warning and this would not have slipped through.
The result of this was that the ARC unconditionally expected
zfs_write_limit_max to be 64-bit. Unfortunately, the largest size
integer module parameter that Linux supports is unsigned long, which
varies in size depending on the host system's native word size. The
effect was that on 32-bit systems, ARC incorrectly performed 64-bit
operations on a 32-bit value by reading the neighboring 32 bits as
the upper 32 bits of the 64-bit value.
We correct that by changing the extern declaration to use the unsigned
long type and move these extern definitions in to the common arc.h
header. This should make ARC correctly treat zfs_write_limit_max as a
32-bit value on 32-bit systems.
Reported-by: Jorgen Lundman <lundman@lundman.net>
Signed-off-by: Richard Yao <ryao@cs.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#749
The Linux 3.1 kernel has introduced the concept of per-filesystem
shrinkers which are directly assoicated with a super block. Prior
to this change there was one shared global shrinker.
The zfs code relied on being able to call the global shrinker when
the arc_meta_limit was exceeded. This would cause the VFS to drop
references on a fraction of the dentries in the dcache. The ARC
could then safely reclaim the memory used by these entries and
honor the arc_meta_limit. Unfortunately, when per-filesystem
shrinkers were added the old interfaces were made unavailable.
This change adds support to use the new per-filesystem shrinker
interface so we can continue to honor the arc_meta_limit. The
major benefit of the new interface is that we can now target
only the zfs filesystem for dentry and inode pruning. Thus we
can minimize any impact on the caching of other filesystems.
In the context of making this change several other important
issues related to managing the ARC were addressed, they include:
* The dnlc_reduce_cache() function which was called by the ARC
to drop dentries for the Posix layer was replaced with a generic
zfs_prune_t callback. The ZPL layer now registers a callback to
drop these dentries removing a layering violation which dates
back to the Solaris code. This callback can also be used by
other ARC consumers such as Lustre.
arc_add_prune_callback()
arc_remove_prune_callback()
* The arc_reduce_dnlc_percent module option has been changed to
arc_meta_prune for clarity. The dnlc functions are specific to
Solaris's VFS and have already been largely eliminated already.
The replacement tunable now represents the number of bytes the
prune callback will request when invoked.
* Less aggressively invoke the prune callback. We used to call
this whenever we exceeded the arc_meta_limit however that's not
strictly correct since it results in over zeleous reclaim of
dentries and inodes. It is now only called once the arc_meta_limit
is exceeded and every effort has been made to evict other data from
the ARC cache.
* More promptly manage exceeding the arc_meta_limit. When reading
meta data in to the cache if a buffer was unable to be recycled
notify the arc_reclaim thread to invoke the required prune.
* Added arcstat_prune kstat which is incremented when the ARC
is forced to request that a consumer prune its cache. Remember
this will only occur when the ARC has no other choice. If it
can evict buffers safely without invoking the prune callback
it will.
* This change is also expected to resolve the unexpect collapses
of the ARC cache. This would occur because when exceeded just the
arc_meta_limit reclaim presure would be excerted on the arc_c
value via arc_shrink(). This effectively shrunk the entire cache
when really we just needed to reclaim meta data.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#466Closes#292
One of the neat tricks an autoconf style project is capable of
is allow configurion/building in a directory other than the
source directory. The major advantage to this is that you can
build the project various different ways while making changes
in a single source tree.
For example, this project is designed to work on various different
Linux distributions each of which work slightly differently. This
means that changes need to verified on each of those supported
distributions perferably before the change is committed to the
public git repo.
Using nfs and custom build directories makes this much easier.
I now have a single source tree in nfs mounted on several different
systems each running a supported distribution. When I make a
change to the source base I suspect may break things I can
concurrently build from the same source on all the systems each
in their own subdirectory.
wget -c http://github.com/downloads/behlendorf/zfs/zfs-x.y.z.tar.gz
tar -xzf zfs-x.y.z.tar.gz
cd zfs-x-y-z
------------------------- run concurrently ----------------------
<ubuntu system> <fedora system> <debian system> <rhel6 system>
mkdir ubuntu mkdir fedora mkdir debian mkdir rhel6
cd ubuntu cd fedora cd debian cd rhel6
../configure ../configure ../configure ../configure
make make make make
make check make check make check make check
This change also moves many of the include headers from individual
incude/sys directories under the modules directory in to a single
top level include directory. This has the advantage of making
the build rules cleaner and logically it makes a bit more sense.
Prakash Surya
Matthew Ahrens
Brian Behlendorf
Saso Kiselkov
Richard Yao
George Wilson
Richard Yao