Table of Contents
Storage before ZFS involved rather expensive hardware that was unable to protect against silent corruption and did not scale very well. The introduction of ZFS has enabled people to use far less expensive hardware than previously used in the industry with superior scaling. This page attempts to provide some basic guidance to people buying hardware for use in ZFS-based servers and workstations.
Hardware that adheres to this guidance will enable ZFS to reach its full potential for performance and reliability. Hardware that does not adhere to it will serve as a handicap. Unless otherwise stated, such handicaps apply to all storage stacks and are by no means specific to ZFS. Systems built using competing storage stacks will also benefit from these suggestions.
Running the latest BIOS and CPU microcode is highly recommended.
Computer microprocessors are very complex designs that often have bugs, which are called errata. Modern microprocessors are designed to utilize microcode. This puts part of the hardware design into quasi-software that can be patched without replacing the entire chip. Errata are often resolved through CPU microcode updates. These are often bundled in BIOS updates. In some cases, the BIOS interactions with the CPU through machine registers can be modified to fix things with the same microcode. If a newer microcode is not bundled as part of a BIOS update, it can often be loaded by the operating system bootloader or the operating system itself.
Bit flips can have fairly dramatic consequences for all computer filesystems and ZFS is no exception. No technique used in ZFS (or any other filesystem) is capable of protecting against bit flips. Consequently, ECC Memory is highly recommended.
Ordinary background radiation will randomly flip bits in computer memory, which causes undefined behavior. These are known as “bit flips”. Each bit flip can have any of four possible consequences depending on which bit is flipped:
Bit flips can have no effect.
Bit flips that have no effect occur in unused memory.
Bit flips can cause runtime failures.
This is the case when a bit flip occurs in something read from disk.
Failures are typically observed when program code is altered.
If the bit flip is in a routine within the system’s kernel or /sbin/init, the system will likely crash. Otherwise, reloading the affected data can clear it. This is typically achieved by a reboot.
It can cause data corruption.
This is the case when the bit is in use by data being written to disk.
If the bit flip occurs before ZFS’ checksum calculation, ZFS will not realize that the data is corrupt.
If the bit flip occurs after ZFS’ checksum calculation, but before write-out, ZFS will detect it, but it might not be able to correct it.
It can cause metadata corruption.
This is the case when a bit flips in an on-disk structure being written to disk.
If the bit flip occurs before ZFS’ checksum calculation, ZFS will not realize that the metadata is corrupt.
If the bit flip occurs after ZFS’ checksum calculation, but before write-out, ZFS will detect it, but it might not be able to correct it.
Recovery from such an event will depend on what was corrupted. In the worst, case, a pool could be rendered unimportable.
All filesystems have poor reliability in their absolute worst case bit-flip failure scenarios. Such scenarios should be considered extraordinarily rare.
ZFS depends on the block device layer for storage. Consequently, ZFS is affected by the same things that affect other filesystems, such as driver support and non-working hardware. Consequently, there are a few things to note:
Never place SATA disks into a SAS expander without a SAS interposer.
If you do this and it does work, it is the exception, rather than the rule.
Do not expect SAS controllers to be compatible with SATA port multipliers.
This configuration is typically not tested.
The disks could be unrecognized.
Support for SATA port multipliers is inconsistent across OpenZFS platforms
Linux drivers generally support them.
Illumos drivers generally do not support them.
FreeBSD drivers are somewhere between Linux and Illumos in terms of support.
These have problems involving sector size reporting, SMART passthrough, the ability to set ERC and other areas. ZFS will perform as well on such devices as they are capable of allowing, but try to avoid them. They should not be expected to have the same up-time as SAS and SATA drives and should be considered unreliable.
The ideal storage controller for ZFS has the following attributes:
Driver support on major OpenZFS platforms
Stability is important.
High per-port bandwidth
PCI Express interface bandwidth divided by the number of ports
Support for RAID, Battery Backup Units and hardware write caches is unnecessary.
Marc Bevand’s blog post From 32 to 2 ports: Ideal SATA/SAS Controllers for ZFS & Linux MD RAID contains an excellent list of storage controllers that meet these criteria. He regularly updates it as newer controllers become available.
Hardware RAID controllers should not be used with ZFS. While ZFS will likely be more reliable than other filesystems on Hardware RAID, it will not be as reliable as it would be on its own.
Hardware RAID will limit opportunities for ZFS to perform self healing on checksum failures. When ZFS does RAID-Z or mirroring, a checksum failure on one disk can be corrected by treating the disk containing the sector as bad for the purpose of reconstructing the original information. This cannot be done when a RAID controller handles the redundancy unless a duplicate copy is stored by ZFS in the case that the corruption involving as metadata, the copies flag is set or the RAID array is part of a mirror/raid-z vdev within ZFS.
Sector size information is not necessarily passed correctly by hardware RAID on RAID 1 and cannot be passed correctly on RAID 5/6. Hardware RAID 1 is more likely to experience read-modify-write overhead from partial sector writes and Hardware RAID 5/6 will almost certainty suffer from partial stripe writes (i.e. the RAID write hole). Using ZFS with the disks directly will allow it to obtain the sector size information reported by the disks to avoid read-modify-write on sectors while ZFS avoids partial stripe writes on RAID-Z by desing from using copy-on-write.
There can be sector alignment problems on ZFS when a drive misreports its sector size. Such drives are typically NAND-flash based solid state drives and older SATA drives from the advanced format (4K sector size) transition before Windows XP EoL occurred. This can be manually corrected at vdev creation.
It is possible for the RAID header to cause misalignment of sector writes on RAID 1 by starting the array within a sector on an actual drive, such that manual correction of sector alignment at vdev creation does not solve the problem.
Controller failures can require that the controller be replaced with the same model, or in less extreme cases, a model from the same manufacturer. Using ZFS by itself allows any controller to be used.
If a hardware RAID controller’s write cache is used, an additional failure point is introduced that can only be partially mitigated by additional complexity from adding flash to save data in power loss events. The data can still be lost if the battery fails when it is required to survive a power loss event or there is no flash and power is not restored in a timely manner. The loss of the data in the write cache can severely damage anything stored on a RAID array when many outstanding writes are cached. In addition, all writes are stored in the cache rather than just synchronous writes that require a write cache, which is inefficient, and the write cache is relatively small. ZFS allows synchronous writes to be written directly to flash, which should provide similar acceleration to hardware RAID and the ability to accelerate many more in-flight operations.
Behavior during RAID reconstruction when silent corruption damages data is undefined. There are reports of RAID 5 and 6 arrays being lost during reconstruction when the controller encounters silent corruption. ZFS’ checksums allow it to avoid this situation by determining if not enough information exists to reconstruct data. In which case, the file is listed as damaged in zpool status and the system administrator has the opportunity to restore it from a backup.
IO response times will be reduced whenever the OS blocks on IO operations because the system CPU blocks on a much weaker embedded CPU used in the RAID controller. This lowers IOPS relative to what ZFS could have achieved.
The controller’s firmware is an additional layer of complexity that cannot be inspected by arbitrary third parties. The ZFS source code is open source and can be inspected by anyone.
If multiple RAID arrays are formed by the same controller and one fails, the identifiers provided by the arrays exposed to the OS might become inconsistent. Giving the drives directly to the OS allows this to be avoided via naming that maps to a unique port or unique drive identifier.
e.g. If you have arrays A, B, C and D; array B dies, the interaction between the hardware RAID controller and the OS might rename arrays C and D to look like arrays B and C respectively. This can fault pools verbatim imported from the cachefile.
Not all RAID controllers behave this way. However, this issue has been observed on both Linux and FreeBSD when system administrators used single drive RAID 0 arrays. It has also been observed with controllers from different vendors.
One might be inclined to try using single-drive RAID 0 arrays to try to use a RAID controller like a HBA, but this is not recommended for many of the reasons listed for other hardware RAID types. It is best to use a HBA instead of a RAID controller, for both performance and reliability.
Historically, all hard drives had 512-byte sectors, with the exception of some SCSI drives that could be modified to support slightly larger sectors. In 2009, the industry migrated from 512-byte sectors to 4096-byte “Advanced Format” sectors. Since Windows XP is not compatible with 4096-byte sectors or drives larger than 2TB, some of the first advanced format drives implemented hacks to maintain Windows XP compatibility.
The first advanced format drives on the market misreported their sector size as 512-bytes for Windows XP compatibility. As of 2013, it is believed that such hard drives are no longer in production. Advanced format hard drives made during or after this time should report their true physical sector size.
Drives storing 2TB and smaller might have a jumper that can be set to map all sectors off by 1. This to provide proper alignment for Windows XP, which started its first partition at sector 63. This jumper setting should be off when using such drives with ZFS.
As of 2014, there are still 512-byte and 4096-byte drives on the market, but they are known to properly identify themselves unless behind a USB to SATA controller. Replacing a 512-byte sector drive with a 4096-byte sector drives in a vdev created with 512-byte sector drives will adversely affect performance. Replacing a 4096-byte sector drive with a 512-byte sector drive will have no negative effect on performance.
ZFS is said to be able to use cheap drives. This was true when it was introduced and hard drives supported Error recovery control. Since ZFS’ introduction, error recovery control has been removed from low-end drives from certain manufacturers, most notably Western Digital. Consistent performance requires hard drives that support error recovery control.
Hard drives store data using small polarized regions a magnetic surface. Reading from and/or writing to this surface poses a few reliability problems. One is that imperfections in the surface can corrupt bits. Another is that vibrations can cause drive heads to miss their targets. Consequently, hard drive sectors are composed of three regions:
A sector number
The actual data
The sector number and ECC enables hard drives to detect and respond to such events. When either event occurs during a read, hard drives will retry the read many times until they either succeed or conclude that the data cannot be read. The latter case can take a substantial amount of time and consequently, IO to the drive will stall.
Enterprise hard drives and some consumer hard drives implement a feature called Time-Limited Error Recovery (TLER) by Western Digital, Error Recovery Control (ERC) by Seagate and Command Completion Time Limit by Hitachi and Samsung, which permits the time drives are willing to spend on such events to be limited by the system administrator.
Drives that lack such functionality can be expected to have arbitrarily high limits. Several minutes is not impossible. Drives with this functionality typically default to 7 seconds. ZFS does not currently adjust this setting on drives. However, it is advisable to write a script to set the error recovery time to a low value, such as 0.1 seconds until ZFS is modified to control it. This must be done on every boot.
High RPM drives have lower seek times, which is historically regarded as being desirable. They increase cost and sacrifice storage density in order to achieve what is typically no more than a factor of 6 improvement over their lower RPM counterparts.
To provide some numbers, a 15k RPM drive from a major manufacturer is rated for 3.4 millisecond average read and 3.9 millisecond average write. Presumably, this number assumes that the target sector is at most half the number of drive tracks away from the head and half the disk away. Being even further away is worst-case 2 times slower. Manufacturer numbers for 7200 RPM drives are not available, but they average 13 to 16 milliseconds in empirical measurements. 5400 RPM drives can be expected to be slower.
ARC and ZIL are able to mitigate much of the benefit of lower seek times. Far larger increases in IOPS performance can be obtained by adding additional RAM for ARC, L2ARC devices and SLOG devices. Even higher increases in performance can be obtained by replacing hard drives with solid state storage entirely. Such things are typically more cost effective than high RPM drives when considering IOPS.
Drives with command queues are able to reorder IO operations to increase IOPS. This is called Native Command Queuing on SATA and Tagged Command Queuing on PATA/SCSI/SAS. ZFS stores objects in metaslabs and it can use several metastabs at any given time. Consequently, ZFS is not only designed to take advantage of command queuing, but good ZFS performance requires command queuing. Almost all drives manufactured within the past 10 years can be expected to support command queuing. The exceptions are:
Consumer PATA/IDE drives
First generation SATA drives, which used IDE to SATA translation chips, from 2003 to 2004.
SATA drives operating under IDE emulation that was configured in the system BIOS.
Each OpenZFS system has different methods for checking whether command
queuing is supported. On Linux,
hdparm -I /path/to/device \| grep
Queue is used. On FreeBSD,
camcontrol identify $DEVICE is used.
As of 2014, Solid state storage is dominated by NAND-flash and most articles on solid state storage focus on it exclusively. As of 2014, the most popular form of flash storage used with ZFS involve drives with SATA interfaces. Enterprise models with SAS interfaces are beginning to become available.
As of 2017, Solid state storage using NAND-flash with PCI-E interfaces are widely available on the market. They are predominantly enterprise drives that utilize a NVMe interface that has lower overhead than the ATA used in SATA or SCSI used in SAS. There is also an interface known as M.2 that is primarily used by consumer SSDs, although not necessarily limited to them. It can provide electrical connectivity for multiple buses, such as SATA, PCI-E and USB. M.2 SSDs appear to use either SATA or NVME.
Many NVMe SSDs support both 512-byte sectors and 4096-byte sectors. They often ship with 512-byte sectors, which are less performant than 4096-byte sectors. Some also support metadata for T10/DIF CRC to try to improve reliability, although this is unnecessary with ZFS.
NVMe drives should be
to use 4096-byte sectors without metadata prior to being given to ZFS
for best performance unless they indicate that 512-byte sectors are as
performant as 4096-byte sectors, although this is unlikely. Lower
numbers in the Rel_Perf of Supported LBA Sizes from
/dev/$device_namespace (for example
smartctl -a /dev/nvme1n1)
indicate higher performance low level formats, with 0 being the best.
The current formatting will be marked by a plus sign under the format
You may format a drive using
nvme format /dev/nvme1n1 -l $ID. The $ID
corresponds to the Id field value from the Supported LBA Sizes SMART
On-flash data structures are highly complex and traditionally have been highly vulnerable to corruption. In the past, such corruption would result in the loss of *all* drive data and an event such as a PSU failure could result in multiple drives simultaneously failing. Since the drive firmware is not available for review, the traditional conclusion was that all drives that lack hardware features to avoid power failure events cannot be trusted, which was found to be the case multiple times in the past 1 2 3. Discussion of power failures bricking NAND flash SSDs appears to have vanished from literature following the year 2015. SSD manufacturers now claim that firmware power loss protection is robust enough to provide equivalent protection to hardware power loss protection. Kingston is one example. Firmware power loss protection is used to guarantee the protection of flushed data and the drives’ own metadata, which is all that filesystems such as ZFS need.
However, those that either need or want strong guarantees that firmware bugs are unlikely to be able to brick drives following power loss events should continue to use drives that provide hardware power loss protection. The basic concept behind how hardware power failure protection works has been documented by Intel for those who wish to read about the details. As of 2020, use of hardware power loss protection is now a feature solely of enterprise SSDs that attempt to protect unflushed data in addition to drive metadata and flushed data. This additional protection beyond protecting flushed data and the drive metadata provides no additional benefit to ZFS, but it does not hurt it.
It should also be noted that drives in data centers and laptops are unlikely to experience power loss events, reducing the usefulness of hardware power loss protection. This is especially the case in datacenters where redundant power, UPS power and the use of IPMI to do forced reboots should prevent most drives from experiencing power loss events.
Lists of drives that provide hardware power loss protection are maintained below for those who need/want it. Since ZFS, like other filesystems, only requires power failure protection for flushed data and drive metadata, older drives that only protect these things are included on the lists.
A non-exhaustive list of NVMe drives with power failure protection is as follows:
Intel DC P3500/P3600/P3608/P3700
Micron 7300/7400/7450 PRO/MAX
Samsung PM963 (M.2 form factor)
Seagate Nytro 5000 M.2 (XP1920LE30002 tested; read notes below before buying)
Inexpensive 22110 M.2 enterprise drive using consumer MLC that is optimized for read mostly workloads. It is not a good choice for a SLOG device, which is a write mostly workload.
The manual for this drive specifies airflow requirements. If the drive does not receive sufficient airflow from case fans, it will overheat at idle. It’s thermal throttling will severely degrade performance such that write throughput performance will be limited to 1/10 of the specification and read latencies will reach several hundred milliseconds. Under continuous load, the device will continue to become hotter until it suffers a “degraded reliability” event where all data on at least one NVMe namespace is lost. The NVMe namespace is then unusable until a secure erase is done. Even with sufficient airflow under normal circumstances, data loss is possible under load following the failure of fans in an enterprise environment. Anyone deploying this into production in an enterprise environment should be mindful of this failure mode.
Those who wish to use this drive in a low airflow situation can workaround this failure mode by placing a passive heatsink such as this on the NAND flash controller. It is the chip under the sticker closest to the capacitors. This was tested by placing the heatsink over the sticker (as removing it was considered undesirable). The heatsink will prevent the drive from overheating to the point of data loss, but it will not fully alleviate the overheating situation under load without active airflow. A scrub will cause it to overheat after a few hundred gigabytes are read. However, the thermal throttling will quickly cool the drive from 76 degrees Celsius to 74 degrees Celsius, restoring performance.
It might be possible to use the heatsink in an enterprise environment to provide protection against data loss following fan failures. However, this was not evaluated. Furthermore, operating temperatures for consumer NAND flash should be at or above 40 degrees Celsius for long term data integrity. Therefore, the use of a heatsink to provide protection against data loss following fan failures in an enterprise environment should be evaluated before deploying drives into production to ensure that the drive is not overcooled.
A non-exhaustive list of SAS drives with power failure protection is as follows:
A non-exhaustive list of SATA drives with power failure protection is as follows:
Early reports claimed that the 330 and 335 had power failure protection too, but they do not.
Intel DC S3500/S3510/S3610/S3700/S3710
Micron 5210 Ion
First QLC drive on the list. High capacity with a low price per gigabyte.
Samsung SM843T (do not confuse with SM843)
Samsung 845DC Evo
Samsung 845DC Pro
These lists have been compiled on a volunteer basis by OpenZFS contributors (mainly Richard Yao) from trustworthy sources of information. The lists are intended to be vendor neutral and are not intended to benefit any particular manufacturer. Any perceived bias toward any manufacturer is caused by a lack of awareness and a lack of time to research additional options. Confirmation of the presence of adequate power loss protection by a reliable source is the only requirement for inclusion into this list. Adequate power loss protection means that the drive must protect both its own internal metadata and all flushed data. Protection of unflushed data is irrelevant and therefore not a requirement. ZFS only expects storage to protect flushed data. Consequently, solid state drives whose power loss protection only protects flushed data is sufficient for ZFS to ensure that data remains safe.
Anyone who believes an unlisted drive to provide adequate power failure protection may contact the Mailing Lists with a request for inclusion and substantiation for the claim that power failure protection is provided. Examples of substantiation include pictures of drive internals showing the presence of capacitors, statements by well regarded independent review sites such as Anandtech and manufacturer specification sheets. The latter are accepted on the honor system until a manufacturer is found to misstate reality on the protection of the drives’ own internal metadata structures and/or the protection of flushed data. Thus far, all manufacturers have been honest.
The smallest unit on a NAND chip that can be written is a flash page. The first NAND-flash SSDs on the market had 4096-byte pages. Further complicating matters is that the the page size has been doubled twice since then. NAND flash SSDs should report these pages as being sectors, but so far, all of them incorrectly report 512-byte sectors for Windows XP compatibility. The consequence is that we have a similar situation to what we had with early advanced format hard drives.
As of 2014, most NAND-flash SSDs on the market have 8192-byte page sizes. However, models using 128-Gbit NAND from certain manufacturers have a 16384-byte page size. Maximum performance requires that vdevs be created with correct ashift values (13 for 8192-byte and 14 for 16384-byte). However, not all OpenZFS platforms support this. The Linux port supports ashift=13, while others are limited to ashift=12 (4096-byte).
As of 2017, NAND-flash SSDs are tuned for 4096-byte IOs. Matching the flash page size is unnecessary and ashift=12 is usually the correct choice. Public documentation on flash page size is also nearly non-existent.
It should be noted that this is a separate case from discard on zvols or hole punching on filesystems. Those work regardless of whether ATA TRIM / SCSI UNMAP is sent to the actual block devices.
The ATA TRIM command in SATA 3.0 and earlier is a non-queued command. Issuing a TRIM command on a SATA drive conforming to SATA 3.0 or earlier will cause the drive to drain its IO queue and stop servicing requests until it finishes, which hurts performance. SATA 3.1 removed this limitation, but very few SATA drives on the market are conformant to SATA 3.1 and it is difficult to distinguish them from SATA 3.0 drives. At the same time, SCSI UNMAP has no such problems.
These are SSDs with far better latencies and write endurance than NAND flash SSDs. They are byte addressable, such that ashift=9 is fine for use on them. Unlike NAND flash SSDs, they do not require any special power failure protection circuitry for reliability. There is also no need to run TRIM on them. However, they cost more per GB than NAND flash (as of 2020). The enterprise models make excellent SLOG devices. Here is a list of models that are known to perform well:
Note that SLOG devices rarely have more than 4GB in use at any given time, so the smaller sized devices are generally the best choice in terms of cost, with larger sizes giving no benefit. Larger sizes could be a good choice for other vdev types, depending on performance needs and cost considerations.
Ensuring that computers are properly grounded is highly recommended. There have been cases in user homes where machines experienced random failures when plugged into power receptacles that had open grounds (i.e. no ground wire at all). This can cause random failures on any computer system, whether it uses ZFS or not.
Power should also be relatively stable. Large dips in voltages from brownouts are preferably avoided through the use of UPS units or line conditioners. Systems subject to unstable power that do not outright shutdown can exhibit undefined behavior. PSUs with longer hold-up times should be able to provide partial protection against this, but hold up times are often undocumented and are not a substitute for a UPS or line conditioner.
PSUs are supposed to deassert a PWR_OK signal to indicate that provided voltages are no longer within the rated specification. This should force an immediate shutdown. However, the system clock of a developer workstation was observed to significantly deviate from the expected value following during a series of ~1 second brown outs. This machine did not use a UPS at the time. However, the PWR_OK mechanism should have protected against this. The observation of the PWR_OK signal failing to force a shutdown with adverse consequences (to the system clock in this case) suggests that the PWR_OK mechanism is not a strict guarantee.
A PSU hold-up time is the amount of time that a PSU can continue to output power at maximum output within standard voltage tolerances following the loss of input power. This is important for supporting UPS units because the transfer time taken by a standard UPS to supply power from its battery can leave machines without power for “5-12 ms”. Intel’s ATX Power Supply design guide specifies a hold up time of 17 milliseconds at maximum continuous output. The hold-up time is a inverse function of how much power is being output by the PSU, with lower power output increasing holdup times.
Capacitor aging in PSUs will lower the hold-up time below what it was when new, which could cause reliability issues as equipment ages. Machines using substandard PSUs with hold-up times below the specification therefore require higher end UPS units for protection to ensure that the transfer time does not exceed the hold-up time. A hold-up time below the transfer time during a transfer to battery power can cause undefined behavior should the PWR_OK signal not become deasserted to force the machine to power off.
If in doubt, use a double conversion UPS unit. Double conversion UPS units always run off the battery, such that the transfer time is 0. This is unless they are high efficiency models that are hybrids between standard UPS units and double conversion UPS units, although these are reported to have much lower transfer times than standard PSUs. You could also contact your PSU manufacturer for the hold up time specification, but if reliability for years is a requirement, you should use a higher end UPS with a low transfer time.
Note that double conversion units are at most 94% efficient unless they support a high efficiency mode, which adds latency to the time to transition to battery power.
The lead acid batteries in UPS units generally need to be replaced regularly to ensure that they provide power during power outages. For home systems, this is every 3 to 5 years, although this varies with temperature 4. For enterprise systems, contact your vendor.