Applies to openSUSE Leap 42.3

12 Tuning I/O Performance

I/O scheduling controls how input/output operations will be submitted to storage. openSUSE Leap offers various I/O algorithms—called elevators—suiting different workloads. Elevators can help to reduce seek operations and can prioritize I/O requests.

Choosing the best suited I/O elevator not only depends on the workload, but on the hardware, too. Single ATA disk systems, SSDs, RAID arrays, or network storage systems, for example, each require different tuning strategies.

12.1 Switching I/O Scheduling

openSUSE Leap picks a default I/O scheduler at boot-time, which can be changed on the fly per block device. This makes it possible to set different algorithms, for example, for the device hosting the system partition and the device hosting a database.

The default I/O scheduler is chosen for each device based on whether the device reports to be rotational disk or not. For non-rotational disks DEADLINE I/O scheduler is picked. Other devices default to CFQ (Completely Fair Queuing). To change this default, use the following boot parameter:

elevator=SCHEDULER

Replace SCHEDULER with one of the values cfq, noop, or deadline. See Section 12.2, “Available I/O Elevators” for details.

To change the elevator for a specific device in the running system, run the following command:

echo SCHEDULER > /sys/block/DEVICE/queue/scheduler

Here, SCHEDULER is one of cfq, noop, or deadline. DEVICE is the block device (sda for example). Note that this change will not persist during reboot. For permanent I/O scheduler change for a particular device either place the command switching the I/O scheduler into init scripts or add appropriate udev rule into /lib/udev/rules.d/. See /lib/udev/rules.d/60-ssd-scheduler.rules for an example of such tuning.

12.2 Available I/O Elevators

In the following elevators available on openSUSE Leap are listed. Each elevator has a set of tunable parameters, which can be set with the following command:

echo VALUE > /sys/block/DEVICE/queue/iosched/TUNABLE

where VALUE is the desired value for the TUNABLE and DEVICE the block device.

To find out which elevator is the current default, run the following command. The currently selected scheduler is listed in brackets:

jupiter:~ # cat /sys/block/sda/queue/scheduler
noop deadline [cfq]

This file can also contain the string none meaning that I/O scheduling does not happen for this device. This is usually because the device uses multi-queue queueing mechanism (refer to Section 12.4, “Enable blk-mq I/O Path for SCSI by Default”).

12.2.1 CFQ (Completely Fair Queuing)

CFQ is a fairness-oriented scheduler and is used by default on openSUSE Leap. The algorithm assigns each thread a time slice in which it is allowed to submit I/O to disk. This way each thread gets a fair share of I/O throughput. It also allows assigning tasks I/O priorities which are taken into account during scheduling decisions (see Section 8.3.3, “Prioritizing Disk Access with ionice). The CFQ scheduler has the following tunable parameters:

/sys/block/DEVICE/queue/iosched/slice_idle_us

When a task has no more I/O to submit in its time slice, the I/O scheduler waits for a while before scheduling the next thread. The slice_idle_us is the time in microseconds the I/O scheduler waits. File slice_idle controls the same tunable but in millisecond units. Waiting for more I/O from a thread can improve locality of I/O. Additionally, it avoids starving processes doing dependent I/O. A process does dependent I/O if it needs a result of one I/O in order to submit another I/O. For example, if you first need to read an index block in order to find out a data block to read, these two reads form a dependent I/O.

For media where locality does not play a big role (SSDs, SANs with lots of disks) setting /sys/block/<device>/queue/iosched/slice_idle_us to 0 can improve the throughput considerably.

/sys/block/DEVICE/queue/iosched/quantum

This option limits the maximum number of requests that are being processed at once by the device. The default value is 4. For a storage with several disks, this setting can unnecessarily limit parallel processing of requests. Therefore, increasing the value can improve performance. However, it can also cause latency of certain I/O operations to increase because more requests are buffered inside the storage. When changing this value, you can also consider tuning /sys/block/DEVICE/queue/iosched/slice_async_rq (the default value is 2). This limits the maximum number of asynchronous requests—usually write requests—that are submitted in one time slice.

/sys/block/DEVICE/queue/iosched/low_latency

When enabled (which is the default on openSUSE Leap) the scheduler may dynamically adjust the length of the time slice by aiming to meet a tuning parameter called the target_latency. Time slices are recomputed to meet this target_latency and ensure that processes get fair access within a bounded length of time.

/sys/block/DEVICE/queue/iosched/target_latency

Contains an estimated latency time for the CFQ. CFQ will use it to calculate the time slice used for every task.

/sys/block/DEVICE/queue/iosched/group_idle_us

To avoid starving of blkio cgroups doing dependent I/O, CFQ waits a bit after completion of I/O for one blkio cgroup before scheduling I/O for a different blkio cgroup. When slice_idle_us is set, this parameter does not have a big impact. However, for fast media, the overhead of slice_idle_us is generally undesirable. Disabling slice_idle_us and setting group_idle_us is a method to avoid starvation of blkio cgroups doing dependent I/O with lower overhead. Note that the file group_idle controls the same tunable however with millisecond granularity.

Example 12.1: Increasing individual thread throughput using CFQ

In openSUSE Leap 42.3, the low_latency tuning parameter is enabled by default to ensure that processes get fair access within a bounded length of time. (Note that this parameter was not enabled in versions prior to openSUSE Leap.)

This is usually preferred in a server scenario where processes are executing I/O as part of transactions, as it makes the time needed for each transaction predictable. However, there are scenarios where that is not the desired behavior:

  • If the performance metric of interest is the peak performance of a single process when there is I/O contention.

  • If a workload must complete as quickly as possible and there are multiple sources of I/O. In this case, unfair treatment from the I/O scheduler may allow the transactions to complete faster: Processes take their full slice and exit quickly, resulting in reduced overall contention.

To address this, there are two options—increase target_latency or disable low_latency. As with all tuning parameters it is important to verify your workload behaves as expected before and after the tuning modification. Take careful note of whether your workload depends on individual process peak performance or scales better with fairness. It should also be noted that the performance will depend on the underlying storage and the correct tuning option for one installation may not be universally true.

Find below an example that does not control when I/O starts but is simple enough to demonstrate the point. 32 processes are writing a small amount of data to disk in parallel. Using the openSUSE Leap default (enabling low_latency), the result looks as follows:

root # echo 1 > /sys/block/sda/queue/iosched/low_latency
root # time ./dd-test.sh
10485760 bytes (10 MB) copied, 2.62464 s, 4.0 MB/s
10485760 bytes (10 MB) copied, 3.29624 s, 3.2 MB/s
10485760 bytes (10 MB) copied, 3.56341 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.56908 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.53043 s, 3.0 MB/s
10485760 bytes (10 MB) copied, 3.57511 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.53672 s, 3.0 MB/s
10485760 bytes (10 MB) copied, 3.5433 s, 3.0 MB/s
10485760 bytes (10 MB) copied, 3.65474 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.63694 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.90122 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 3.88507 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 3.86135 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 3.84553 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 3.88871 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 3.94943 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 4.12731 s, 2.5 MB/s
10485760 bytes (10 MB) copied, 4.15106 s, 2.5 MB/s
10485760 bytes (10 MB) copied, 4.21601 s, 2.5 MB/s
10485760 bytes (10 MB) copied, 4.35004 s, 2.4 MB/s
10485760 bytes (10 MB) copied, 4.33387 s, 2.4 MB/s
10485760 bytes (10 MB) copied, 4.55434 s, 2.3 MB/s
10485760 bytes (10 MB) copied, 4.52283 s, 2.3 MB/s
10485760 bytes (10 MB) copied, 4.52682 s, 2.3 MB/s
10485760 bytes (10 MB) copied, 4.56176 s, 2.3 MB/s
10485760 bytes (10 MB) copied, 4.62727 s, 2.3 MB/s
10485760 bytes (10 MB) copied, 4.78958 s, 2.2 MB/s
10485760 bytes (10 MB) copied, 4.79772 s, 2.2 MB/s
10485760 bytes (10 MB) copied, 4.78004 s, 2.2 MB/s
10485760 bytes (10 MB) copied, 4.77994 s, 2.2 MB/s
10485760 bytes (10 MB) copied, 4.86114 s, 2.2 MB/s
10485760 bytes (10 MB) copied, 4.88062 s, 2.1 MB/s

real    0m4.978s
user    0m0.112s
sys     0m1.544s

Note that each process completes in similar times. This is the CFQ scheduler meeting its target_latency: Each process has fair access to storage.

Note that the earlier processes complete somewhat faster. This happens because the start time of the processes is not identical. In a more complicated example, it is possible to control for this.

This is what happens when low_latency is disabled:

root # echo 0 > /sys/block/sda/queue/iosched/low_latency
root # time ./dd-test.sh
10485760 bytes (10 MB) copied, 0.813519 s, 12.9 MB/s
10485760 bytes (10 MB) copied, 0.788106 s, 13.3 MB/s
10485760 bytes (10 MB) copied, 0.800404 s, 13.1 MB/s
10485760 bytes (10 MB) copied, 0.816398 s, 12.8 MB/s
10485760 bytes (10 MB) copied, 0.959087 s, 10.9 MB/s
10485760 bytes (10 MB) copied, 1.09563 s, 9.6 MB/s
10485760 bytes (10 MB) copied, 1.18716 s, 8.8 MB/s
10485760 bytes (10 MB) copied, 1.27661 s, 8.2 MB/s
10485760 bytes (10 MB) copied, 1.46312 s, 7.2 MB/s
10485760 bytes (10 MB) copied, 1.55489 s, 6.7 MB/s
10485760 bytes (10 MB) copied, 1.64277 s, 6.4 MB/s
10485760 bytes (10 MB) copied, 1.78196 s, 5.9 MB/s
10485760 bytes (10 MB) copied, 1.87496 s, 5.6 MB/s
10485760 bytes (10 MB) copied, 1.9461 s, 5.4 MB/s
10485760 bytes (10 MB) copied, 2.08351 s, 5.0 MB/s
10485760 bytes (10 MB) copied, 2.28003 s, 4.6 MB/s
10485760 bytes (10 MB) copied, 2.42979 s, 4.3 MB/s
10485760 bytes (10 MB) copied, 2.54564 s, 4.1 MB/s
10485760 bytes (10 MB) copied, 2.6411 s, 4.0 MB/s
10485760 bytes (10 MB) copied, 2.75171 s, 3.8 MB/s
10485760 bytes (10 MB) copied, 2.86162 s, 3.7 MB/s
10485760 bytes (10 MB) copied, 2.98453 s, 3.5 MB/s
10485760 bytes (10 MB) copied, 3.13723 s, 3.3 MB/s
10485760 bytes (10 MB) copied, 3.36399 s, 3.1 MB/s
10485760 bytes (10 MB) copied, 3.60018 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.58151 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.67385 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.69471 s, 2.8 MB/s
10485760 bytes (10 MB) copied, 3.66658 s, 2.9 MB/s
10485760 bytes (10 MB) copied, 3.81495 s, 2.7 MB/s
10485760 bytes (10 MB) copied, 4.10172 s, 2.6 MB/s
10485760 bytes (10 MB) copied, 4.0966 s, 2.6 MB/s

real    0m3.505s
user    0m0.160s
sys     0m1.516s

Note that the time processes take to complete is spread much wider as processes are not getting fair access. Some processes complete faster and exit, allowing the total workload to complete faster, and some processes measure higher apparent I/O performance. It is also important to note that this example may not behave similarly on all systems as the results depend on the resources of the machine and the underlying storage.

It is important to emphasize that neither tuning option is inherently better than the other. Both are best in different circumstances and it is important to understand the requirements of your workload and tune accordingly.

12.2.2 NOOP

A trivial scheduler that only passes down the I/O that comes to it. Useful for checking whether complex I/O scheduling decisions of other schedulers are causing I/O performance regressions.

This scheduler is recommended for setups with devices that do I/O scheduling themselves, such as intelligent storage or in multipathing environments. If you choose a more complicated scheduler on the host, the scheduler of the host and the scheduler of the storage device compete with each other. This can decrease performance. The storage device can usually determine best how to schedule I/O.

For similar reasons, this scheduler is also recommended for use within virtual machines.

The NOOP scheduler can be useful for devices that do not depend on mechanical movement, like SSDs. Usually, the DEADLINE I/O scheduler is a better choice for these devices. However, NOOP creates less overhead and thus can on certain workloads increase performance.

12.2.3 DEADLINE

DEADLINE is a latency-oriented I/O scheduler. Each I/O request is assigned a deadline. Usually, requests are stored in queues (read and write) sorted by sector numbers. The DEADLINE algorithm maintains two additional queues (read and write) in which requests are sorted by deadline. As long as no request has timed out, the sector queue is used. When timeouts occur, requests from the deadline queue are served until there are no more expired requests. Generally, the algorithm prefers reads over writes.

This scheduler can provide a superior throughput over the CFQ I/O scheduler in cases where several threads read and write and fairness is not an issue. For example, for several parallel readers from a SAN and for databases (especially when using TCQ disks). The DEADLINE scheduler has the following tunable parameters:

/sys/block/<device>/queue/iosched/writes_starved

Controls how many reads can be sent to disk before it is possible to send writes. A value of 3 means, that three read operations are carried out for one write operation.

/sys/block/<device>/queue/iosched/read_expire

Sets the deadline (current time plus the read_expire value) for read operations in milliseconds. The default is 500.

/sys/block/<device>/queue/iosched/write_expire

/sys/block/<device>/queue/iosched/read_expire Sets the deadline (current time plus the read_expire value) for read operations in milliseconds. The default is 500.

12.3 I/O Barrier Tuning

Most file systems (such as XFS, Ext3, Ext4, or reiserfs) send write barriers to disk after fsync or during transaction commits. Write barriers enforce proper ordering of writes, making volatile disk write caches safe to use (at some performance penalty). If your disks are battery-backed in one way or another, disabling barriers can safely improve performance.

Sending write barriers can be disabled using the nobarrier mount option.

Warning
Warning: Disabling Barriers Can Lead to Data Loss

Disabling barriers when disks cannot guarantee caches are properly written in case of power failure can lead to severe file system corruption and data loss.

12.4 Enable blk-mq I/O Path for SCSI by Default

Block multiqueue (blk-mq) is a multi-queue block I/O queueing mechanism. Blk-mq uses per-cpu software queues to queue I/O requests. The software queues are mapped to one or more hardware submission queues. Blk-mq significantly reduces lock contention. In particular blk-mq improves performance for devices that support a high number of input/output operations per second (IOPS). Blk-mq is already the default for some devices, e.g. NVM Express devices.

Currently blk-mq has no I/O scheduling support (no CFQ, no deadline I/O scheduler). This lack of I/O scheduling can cause significant performance degradation when spinning disks are used. Therefore blk-mq is not enabled by default for SCSI devices.

If you have fast SCSI devices (e.g. SSDs) instead of spinning SCSI devices attached to your system you might consider switching to blk-mq for SCSI. This is done using the kernel command line option scsi_mod.use_blk_mq=1.

Print this page