How to Choose and Tune Linux I/O Schedulers for Maximum Performance

Linux I/O scheduler is a component of the Linux kernel that sits between the generic block layer and block device drivers, allowing performance optimization by selecting different scheduling algorithms.

1. Scheduling Algorithms

The main I/O schedulers in Linux are:

noop (No Operation): an elevator algorithm based on FIFO queues, suitable for SSDs because of fast random reads/writes.

cfq (Completely Fair Scheduler): creates a separate queue for each process/thread to distribute I/O evenly; default in CentOS 6.

deadline: assigns a latest execution time to each I/O request; best for mechanical disks in database workloads; also default in CentOS 7.

2. Viewing and Changing the Scheduler

In CentOS 7 the available schedulers are listed by:

[[email protected] www]# dmesg | grep -i scheduler
[    1.307640] io scheduler noop registered
[    1.308944] io scheduler deadline registered (default)
[    1.310445] io scheduler cfq registered
[    1.311733] io scheduler mq-deadline registered
[    1.313133] io scheduler kyber registered
[[email protected] www]#

In CentOS 6 the list is:

[[email protected] www]# dmesg | grep -i scheduler
io scheduler noop registered
io scheduler anticipatory registered
io scheduler deadline registered
io scheduler cfq registered (default)

CentOS 7.x uses deadline as the default scheduler, while CentOS 6.x defaults to cfq. You can also view or change the scheduler for each block device, e.g.:

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

Temporary change to noop:

# echo 'noop' >/sys/block/sda/queue/scheduler
# cat /sys/block/sda/queue/scheduler
[noop] deadline cfq

Temporary change to cfq:

# echo 'cfq' >/sys/block/sda/queue/scheduler
# cat /sys/block/sda/queue/scheduler
noop deadline [cfq]

To make the change permanent:

# For CentOS 7.x
grubby --update-kernel=ALL --args="elevator=deadline"
reboot
cat /sys/block/sda/queue/scheduler   # shows noop [deadline] cfq

# For CentOS 6.x
Edit /boot/grub/menu.lst or /boot/grub/grub.conf and add:
kernel /boot/vmlinuz... elevator=deadline rhgb quiet

3. Performance Benchmark Comparison

In summary, for SSDs the noop or deadline scheduler performs best, while SAS disks benefit from deadline. Since CentOS 7 defaults to deadline, no adjustment is usually needed, but on CentOS 6 you should consider the I/O scheduler.

A sysbench test on MySQL with an SSD shows that the noop scheduler yields the highest IOPS and throughput:

Algorithm   IOPS   Write Speed   Read IOPS   Read Speed   oltp_write_only (TPS)   oltp_read_only (TPS)
deadline    6935.37 118.37MiB/s 7956.88 124.33MiB/s 491.43 TPS (2948.62 queries) 393.13 TPS (6290.13 queries)
noop        7057.60 110.27MiB/s 8399.89 131.25MiB/s 544.38 TPS (3266.28 queries) 379.97 TPS (6079.59 queries)
cfq         6614.37 103.35MiB/s 7481.39 116.90MiB/s 498.54 TPS (2991.25 queries) 344.66 TPS (5514.58 queries)