Below are tips for various workloads.
Descriptions of ZFS internals that have an effect on application performance follow.
For decades, operating systems have used RAM as a cache to avoid the necessity of waiting on disk IO, which is extremely slow. This concept is called page replacement. Until ZFS, virtually all filesystems used the Least Recently Used (LRU) page replacement algorithm in which the least recently used pages are the first to be replaced. Unfortunately, the LRU algorithm is vulnerable to cache flushes, where a brief change in workload that occurs occasionally removes all frequently used data from cache. The Adaptive Replacement Cache (ARC) algorithm was implemented in ZFS to replace LRU. It solves this problem by maintaining four lists:
A list for recently cached entries.
A list for recently cached entries that have been accessed more than once.
A list for entries evicted from #1.
A list of entries evicited from #2.
Data is evicted from the first list while an effort is made to keep data in the second list. In this way, ARC is able to outperform LRU by providing a superior hit rate.
In addition, a dedicated cache device (typically a SSD) can be added to
the pool, with
zpool add POOLNAME cache DEVICENAME. The cache
device is managed by the L2ARC, which scans entries that are next to be
evicted and writes them to the cache device. The data stored in ARC and
L2ARC can be controlled via the
zfs properties respectively, which can be set on both zvols and
datasets. Possible settings are
is possible to improve performance when a zvol or dataset hosts an
application that does its own caching by caching only metadata. One
example would be a virtual machine using ZFS. Another would be a
database system which manages its own cache (Oracle for instance).
PostgreSQL, by contrast, depends on the OS-level file cache for the
majority of cache.
Top-level vdevs contain an internal property called ashift, which stands
for alignment shift. It is set at vdev creation and it is immutable. It
can be read using the
zdb command. It is calculated as the maximum
base 2 logarithm of the physical sector size of any child vdev and it
alters the disk format such that writes are always done according to it.
This makes 2^ashift the smallest possible IO on a vdev. Configuring
ashift correctly is important because partial sector writes incur a
penalty where the sector must be read into a buffer before it can be
written. ZFS makes the implicit assumption that the sector size reported
by drives is correct and calculates ashift based on that.
In an ideal world, physical sector size is always reported correctly and therefore, this requires no attention. Unfortunately, this is not the case. The sector size on all storage devices was 512-bytes prior to the creation of flash-based solid state drives. Some operating systems, such as Windows XP, were written under this assumption and will not function when drives report a different sector size.
Flash-based solid state drives came to market around 2007. These devices report 512-byte sectors, but the actual flash pages, which roughly correspond to sectors, are never 512-bytes. The early models used 4096-byte pages while the newer models have moved to an 8192-byte page. In addition, “Advanced Format” hard drives have been created which also use a 4096-byte sector size. Partial page writes suffer from similar performance degradation as partial sector writes. In some cases, the design of NAND-flash makes the performance degradation even worse, but that is beyond the scope of this description.
Reporting the correct sector sizes is the responsibility the block device layer. This unfortunately has made proper handling of devices that misreport drives different across different platforms. The respective methods are as follows:
sd.conf on illumos
ashift= on ZFS on Linux
-o ashift= also works with both MacZFS (pool version 8) and ZFS-OSX (pool version 5000).
-o ashift= is convenient, but it is flawed in that the creation of pools containing top level vdevs that have multiple optimal sector sizes require the use of multiple commands. A newer syntax that will rely on the actual sector sizes has been discussed as a cross platform replacement and will likely be implemented in the future.
In addition, there is a database of drives known to misreport sector sizes to the ZFS on Linux project. It is used to automatically adjust ashift without the assistance of the system administrator. This approach is unable to fully compensate for misreported sector sizes whenever drive identifiers are used ambiguously (e.g. virtual machines, iSCSI LUNs, some rare SSDs), but it does a great amount of good. The format is roughly compatible with illumos’ sd.conf and it is expected that other implementations will integrate the database in future releases. Strictly speaking, this database does not belong in ZFS, but the difficulty of patching the Linux kernel (especially older ones) necessitated that this be implemented in ZFS itself for Linux. The same is true for MacZFS. However, FreeBSD and illumos are both able to implement this in the correct layer.
Internally, ZFS allocates data using multiples of the device’s sector
size, typically either 512 bytes or 4KB (see above). When compression is
enabled, a smaller number of sectors can be allocated for each block.
The uncompressed block size is set by the
recordsize (defaults to
volblocksize (defaults to 16KB since v2.2) property (for filesystems
The following compression algorithms are available:
New algorithm added after feature flags were created. It is significantly superior to LZJB in all metrics tested. It is new default compression algorithm (compression=on) in OpenZFS. It is available on all platforms as of 2020.
Original default compression algorithm (compression=on) for ZFS. It was created to satisfy the desire for a compression algorithm suitable for use in filesystems. Specifically, that it provides fair compression, has a high compression speed, has a high decompression speed and detects incompressible data quickly.
GZIP (1 through 9)
Classic Lempel-Ziv implementation. It provides high compression, but it often makes IO CPU-bound.
ZLE (Zero Length Encoding)
A very simple algorithm that only compresses zeroes.
Zstandard is a modern, high performance, general compression algorithm which provides similar or better compression levels to GZIP, but with much better performance. Zstandard offers a very wide range of performance/compression trade-off, and is backed by an extremely fast decoder. It is available from OpenZFS 2.0 version.
If you want to use compression and are uncertain which to use, use LZ4. It averages a 2.1:1 compression ratio while gzip-1 averages 2.7:1, but gzip is much slower. Both figures are obtained from testing by the LZ4 project on the Silesia corpus. The greater compression ratio of gzip is usually only worthwhile for rarely accessed data.
Choose a RAID-Z stripe width based on your IOPS needs and the amount of space you are willing to devote to parity information. If you need more IOPS, use fewer disks per stripe. If you need more usable space, use more disks per stripe. Trying to optimize your RAID-Z stripe width based on exact numbers is irrelevant in nearly all cases. See this blog post for more details.
ZFS datasets use an internal recordsize of 128KB by default. The dataset recordsize is the basic unit of data used for internal copy-on-write on files. Partial record writes require that data be read from either ARC (cheap) or disk (expensive). recordsize can be set to any power of 2 from 512 bytes to 1 megabyte. Software that writes in fixed record sizes (e.g. databases) will benefit from the use of a matching recordsize.
Changing the recordsize on a dataset will only take effect for new files. If you change the recordsize because your application should perform better with a different one, you will need to recreate its files. A cp followed by a mv on each file is sufficient. Alternatively, send/recv should recreate the files with the correct recordsize when a full receive is done.
Record sizes of up to 16M are supported with the large_blocks pool feature, which is enabled by default on new pools on systems that support it.
Record sizes larger than 1M were disabled by default before openZFS v2.2, unless the zfs_max_recordsize kernel module parameter was set to allow sizes higher than 1M.
`zfs send` operations must specify -L to ensure that larger than 128KB blocks are sent and the receiving pools must support the large_blocks feature.
Zvols have a
volblocksize property that is analogous to
Current default (16KB since v2.2) balances the metadata overhead, compression
opportunities and decent space efficiency on majority of pool configurations
due to 4KB disk physical block rounding (especially on RAIDZ and DRAID),
while incurring some write amplification on guest FSes that run with smaller
block sizes .
Users are advised to test their scenarios and see whether the
needs to be changed to favor one or the other:
sector alignment of guest FS is crucial
most of guest FSes use default block size of 4-8KB, so:
volblocksizecan help with mostly sequential workloads and will gain a compression efficiency
volblocksizecan help with random workloads and minimize IO amplification, but will use more metadata (e.g. more small IOs will be generated by ZFS) and may have worse space efficiency (especially on RAIDZ and DRAID)
It’s meaningless to set
volblocksizeless than guest FS’s block size or ashift
See Dataset recordsize for additional information
Deduplication uses an on-disk hash table, using extensible
implemented in the ZAP (ZFS Attribute Processor). Each cached entry uses
slightly more than 320 bytes of memory. The DDT code relies on ARC for
caching the DDT entries, such that there is no double caching or
internal fragmentation from the kernel memory allocator. Each pool has a
global deduplication table shared across all datasets and zvols on which
deduplication is enabled. Each entry in the hash table is a record of a
unique block in the pool. (Where the block size is set by the
The hash table (also known as the DDT or DeDup Table) must be accessed for every dedup-able block that is written or freed (regardless of whether it has multiple references). If there is insufficient memory for the DDT to be cached in memory, each cache miss will require reading a random block from disk, resulting in poor performance. For example, if operating on a single 7200RPM drive that can do 100 io/s, uncached DDT reads would limit overall write throughput to 100 blocks per second, or 400KB/s with 4KB blocks.
The consequence is that sufficient memory to store deduplication data is
required for good performance. The deduplication data is considered
metadata and therefore can be cached if the
secondarycache properties are set to
metadata. In addition, the
deduplication table will compete with other metadata for metadata
storage, which can have a negative effect on performance. Simulation of
the number of deduplication table entries needed for a given pool can be
done using the -D option to zdb. Then a simple multiplication by
320-bytes can be done to get the approximate memory requirements.
Alternatively, you can estimate an upper bound on the number of unique
blocks by dividing the amount of storage you plan to use on each dataset
(taking into account that partial records each count as a full
recordsize for the purposes of deduplication) by the recordsize and each
zvol by the volblocksize, summing and then multiplying by 320-bytes.
ZFS top level vdevs are divided into metaslabs from which blocks can be independently allocated so allow for concurrent IOs to perform allocations without blocking one another. At present, there is a regression on the Linux and Mac OS X ports that causes serialization to occur.
By default, the selection of a metaslab is biased toward lower LBAs to improve performance of spinning disks, but this does not make sense on solid state media. This behavior can be adjusted globally by setting the ZFS module’s global metaslab_lba_weighting_enabled tuanble to 0. This tunable is only advisable on systems that only use solid state media for pools.
The metaslab allocator will allocate blocks on a first-fit basis when a
metaslab has more than or equal to 4 percent free space and a best-fit
basis when a metaslab has less than 4 percent free space. The former is
much faster than the latter, but it is not possible to tell when this
behavior occurs from the pool’s free space. However, the command
-mmm $POOLNAME will provide this information.
If small random IOPS are of primary importance, mirrored vdevs will outperform raidz vdevs. Read IOPS on mirrors will scale with the number of drives in each mirror while raidz vdevs will each be limited to the IOPS of the slowest drive.
If sequential writes are of primary importance, raidz will outperform mirrored vdevs. Sequential write throughput increases linearly with the number of data disks in raidz while writes are limited to the slowest drive in mirrored vdevs. Sequential read performance should be roughly the same on each.
Both IOPS and throughput will increase by the respective sums of the IOPS and throughput of each top level vdev, regardless of whether they are raidz or mirrors.
ZFS will behave differently on different platforms when given a whole disk.
On illumos, ZFS attempts to enable the write cache on a whole disk. The illumos UFS driver cannot ensure integrity with the write cache enabled, so by default Sun/Solaris systems using UFS file system for boot were shipped with drive write cache disabled (long ago, when Sun was still an independent company). For safety on illumos, if ZFS is not given the whole disk, it could be shared with UFS and thus it is not appropriate for ZFS to enable write cache. In this case, the write cache setting is not changed and will remain as-is. Today, most vendors ship drives with write cache enabled by default.
On Linux, the Linux IO elevator is largely redundant given that ZFS has its own IO elevator.
ZFS will also create a GPT partition table own partitions when given a whole disk under illumos on x86/amd64 and on Linux. This is mainly to make booting through UEFI possible because UEFI requires a small FAT partition to be able to boot the system. The ZFS driver will be able to tell the difference between whether the pool had been given the entire disk or not via the whole_disk field in the label.
This is not done on FreeBSD. Pools created by FreeBSD will always have the whole_disk field set to true, such that a pool imported on another platform that was created on FreeBSD will always be treated as the whole disks were given to ZFS.
Some Linux distributions (at least Debian, Ubuntu) enable
init_on_alloc option as security precaution by default.
This option can help to :
prevent possible information leaks and make control-flow bugs that depend on uninitialized values more deterministic.
Unfortunately, it can lower ARC throughput considerably (see bug).
If you’re ready to cope with these security risks ,
you may disable it
init_on_alloc=0 in the GRUB kernel boot parameters.
Make sure that you create your pools such that the vdevs have the correct alignment shift for your storage device’s size. if dealing with flash media, this is going to be either 12 (4K sectors) or 13 (8K sectors). For SSD ephemeral storage on Amazon EC2, the proper setting is 12.
Set either relatime=on or atime=off to minimize IOs used to update access time stamps. For backward compatibility with a small percentage of software that supports it, relatime is preferred when available and should be set on your entire pool. atime=off should be used more selectively.
Keep pool free space above 10% to avoid many metaslabs from reaching the 4% free space threshold to switch from first-fit to best-fit allocation strategies. When the threshold is hit, the Metaslab Allocator becomes very CPU intensive in an attempt to protect itself from fragmentation. This reduces IOPS, especially as more metaslabs reach the 4% threshold.
The recommendation is 10% rather than 5% because metaslabs selection considers both location and free space unless the global metaslab_lba_weighting_enabled tunable is set to 0. When that tunable is 0, ZFS will consider only free space, so the the expense of the best-fit allocator can be avoided by keeping free space above 5%. That setting should only be used on systems with pools that consist of solid state drives because it will reduce sequential IO performance on mechanical disks.
Set compression=lz4 on your pools’ root datasets so that all datasets inherit it unless you have a reason not to enable it. Userland tests of LZ4 compression of incompressible data in a single thread has shown that it can process 10GB/sec, so it is unlikely to be a bottleneck even on incompressible data. Furthermore, incompressible data will be stored without compression such that reads of incompressible data with compression enabled will not be subject to decompression. Writes are so fast that in-compressible data is unlikely to see a performance penalty from the use of LZ4 compression. The reduction in IO from LZ4 will typically be a performance win.
Note that larger record sizes will increase compression ratios on compressible data by allowing compression algorithms to process more data at a time.
Do not put more than ~16 disks in raidz. The rebuild times on mechanical disks will be excessive when the pool is full.
If your workload involves fsync or O_SYNC and your pool is backed by mechanical storage, consider adding one or more SLOG devices. Pools that have multiple SLOG devices will distribute ZIL operations across them. The best choice for SLOG device(s) are likely Optane / 3D XPoint SSDs. See Optane / 3D XPoint SSDs for a description of them. If an Optane / 3D XPoint SSD is an option, the rest of this section on synchronous I/O need not be read. If Optane / 3D XPoint SSDs is not an option, see NAND Flash SSDs for suggestions for NAND flash SSDs and also read the information below.
To ensure maximum ZIL performance on NAND flash SSD-based SLOG devices, you should also overprovison spare area to increase IOPS . Only about 4GB is needed, so the rest can be left as overprovisioned storage. The choice of 4GB is somewhat arbitrary. Most systems do not write anything close to 4GB to ZIL between transaction group commits, so overprovisioning all storage beyond the 4GB partition should be alright. If a workload needs more, then make it no more than the maximum ARC size. Even under extreme workloads, ZFS will not benefit from more SLOG storage than the maximum ARC size. That is half of system memory on Linux and 3/4 of system memory on illumos.
You can do this with a mix of a secure erase and a partition table trick, such as the following:
Run a secure erase on the NAND-flash SSD.
Create a partition table on the NAND-flash SSD.
Create a 4GB partition.
Give the partition to ZFS to use as a log device.
If using the secure erase and partition table trick, do not use the unpartitioned space for other things, even temporarily. That will reduce or eliminate the overprovisioning by marking pages as dirty.
Alternatively, some devices allow you to change the sizes that they report.This would also work, although a secure erase should be done prior to changing the reported size to ensure that the SSD recognizes the additional spare area. Changing the reported size can be done on drives that support it with `hdparm -N ` on systems that have laptop-mode-tools.
On NVMe, you can use namespaces to achieve overprovisioning:
Do a sanitize command as a precaution to ensure the device is completely clean.
Delete the default namespace.
Create a new namespace of size 4GB.
Give the namespace to ZFS to use as a log device. e.g. zfs add tank log /dev/nvme1n1
Whole disks should be given to ZFS rather than partitions. If you must use a partition, make certain that the partition is properly aligned to avoid read-modify-write overhead. See the section on Alignment Shift (ashift) for a description of proper alignment. Also, see the section on Whole Disks versus Partitions for a description of changes in ZFS behavior when operating on a partition.
Single disk RAID 0 arrays from RAID controllers are not equivalent to whole disks. The Hardware RAID controllers page explains in detail.
Bit torrent performs 16KB random reads/writes. The 16KB writes cause read-modify-write overhead. The read-modify-write overhead can reduce performance by a factor of 16 with 128KB record sizes when the amount of data written exceeds system memory. This can be avoided by using a dedicated dataset for bit torrent downloads with recordsize=16KB.
When the files are read sequentially through a HTTP server, the random nature in which the files were generated creates fragmentation that has been observed to reduce sequential read performance by a factor of two on 7200RPM hard disks. If performance is a problem, fragmentation can be eliminated by rewriting the files sequentially in either of two ways:
The first method is to configure your client to download the files to a temporary directory and then copy them into their final location when the downloads are finished, provided that your client supports this.
The second method is to use send/recv to recreate a dataset sequentially.
In practice, defragmenting files obtained through bit torrent should only improve performance when the files are stored on magnetic storage and are subject to significant sequential read workloads after creation.
redundant_metadata=most can increase IOPS by at least a few
percentage points by eliminating redundant metadata at the lowest level
of the indirect block tree. This comes with the caveat that data loss
will occur if a metadata block pointing to data blocks is corrupted and
there are no duplicate copies, but this is generally not a problem in
production on mirrored or raidz vdevs.
Make separate datasets for InnoDB’s data files and log files. Set
recordsize=16K on InnoDB’s data files to avoid expensive partial record
writes and leave recordsize=128K on the log files. Set
primarycache=metadata on both to prefer InnoDB’s
logbias=throughput on the data to stop ZIL from writing twice.
skip-innodb_doublewrite in my.cnf to prevent innodb from writing
twice. The double writes are a data integrity feature meant to protect
against corruption from partially-written records, but those are not
possible on ZFS. It should be noted that Percona’s
blog had advocated
using an ext4 configuration where double writes were
turned off for a performance gain, but later recanted it because it
caused data corruption. Following a well timed power failure, an in
place filesystem such as ext4 can have half of a 8KB record be old while
the other half would be new. This would be the corruption that caused
Percona to recant its advice. However, ZFS’ copy on write design would
cause it to return the old correct data following a power failure (no
matter what the timing is). That prevents the corruption that the double
write feature is intended to prevent from ever happening. The double
write feature is therefore unnecessary on ZFS and can be safely turned
off for better performance.
On Linux, the driver’s AIO implementation is a compatibility shim that
just barely passes the POSIX standard. InnoDB performance suffers when
using its default AIO codepath. Set
innodb_use_atomic_writes=0 in my.cnf to disable AIO. Both of these
settings must be disabled to disable AIO.
Make separate datasets for PostgreSQL’s data and WAL. Set
recordsize=32K (64K also work well, as
does the 128K default) on both. Configure
full_page_writes = off
for PostgreSQL, as ZFS will never commit a partial write. For a database
with large updates, experiment with
PostgreSQL’s data to avoid writing twice, but be aware that with this
setting smaller updates can cause severe fragmentation.
Make a separate dataset for the database. Set the recordsize to 64K. Set the SQLite page size to 65536 bytes .
Note that SQLite databases typically are not exercised enough to merit special tuning, but this will provide it. Note the side effect on cache size mentioned at SQLite.org .
Create a dedicated dataset for files being served.
See Sequential workloads for configuration recommendations.
Windows/DOS clients doesn’t support case sensitive file names.
If your main workload won’t need case sensitivity for other supported clients,
create dataset with
zfs create -o casesensitivity=insensitive
so Samba may search filenames faster in future .
case sensitive option in
recordsize=1M on datasets that are subject to sequential workloads.
Larger record sizes
for documentation on things that should be known before setting 1M
compression=lz4 as per the general recommendation for LZ4
Create a dedicated dataset, use chown to make it user accessible (or create a directory under it and use chown on that) and then configure the game download application to place games there. Specific information on how to configure various ones is below.
See Sequential workloads for configuration recommendations before installing games.
Note that the performance gains from this tuning are likely to be small and limited to load times. However, the combination of 1M records and LZ4 will allow more games to be stored, which is why this tuning is documented despite the performance gains being limited. A steam library of 300 games (mostly from humble bundle) that had these tweaks applied to it saw 20% space savings. Both faster load times and significant space savings are possible on compressible games when this tuning has been done. Games whose assets are already compressed will see little to no benefit.
Open the context menu by left clicking on the triple bar icon in the upper right. Go to “Preferences” and then the “System options” tab. Change the default installation directory and click save.
Go to “Settings” -> “Downloads” -> “Steam Library Folders” and use “Add Library Folder” to set the directory for steam to use to store games. Make sure to set it to the default by right clicking on it and clicking “Make Default Folder” before closing the dialogue.
If you’ll use Proton to run non-native games,
create dataset with
zfs create -o casesensitivity=insensitive
so Wine may search filenames faster in future .
Windows file systems’ standard behavior is to be case-insensitive.
Create dataset with
zfs create -o casesensitivity=insensitive
so Wine may search filenames faster in future .
Virtual machine images on ZFS should be stored using either zvols or raw files to avoid unnecessary overhead. The recordsize/volblocksize and guest filesystem may be configured to match to avoid overhead from partial record modification, see zvol volblocksize. If raw files are used, a separate dataset should be used to make it easy to configure recordsize independently of other things stored on ZFS.
AIO should be used to maximize IOPS when using files for guest storage.