linuxexpert-practice-labs/labs/lfcs/3-7-storage-performance

LFCS 3.7: Storage performance

Finding a busy disk and the process behind it, measuring a disk with fio, I/O schedulers, and keeping disk history with sysstat.

Before you start

The lab adds three spare disks, which come back at every boot:

• /dev/loop40 (1 GiB), with an ext4 filesystem mounted at /data, for your test files;
• /dev/loop41 (256 MiB), empty, for the scheduler tasks;
• /dev/loop42 (256 MiB), which a process the lab started keeps busy.

If sysstat collection was on, the lab turns it off.

Tasks

1. One disk is busy. Find the busiest device and the process responsible. Write the device name (such as vdb) on the first line of /root/lab-answers/1, and the process's command name (as ps or iotop shows it) on the second line.
2. Measure /data's disk with fio, using a test file in /data: first sequential write throughput, then random 4 KiB read IOPS. Write the first test's BW= figure (such as 188MiB/s) on the first line of /root/lab-answers/2 and the second test's IOPS= figure (such as 48.2k) on the second line. Remove the test file afterwards.
3. Find the I/O scheduler of /dev/loop41 and write its name to /root/lab-answers/3. Then change it to mq-deadline until the next reboot.
4. Make that scheduler choice for /dev/loop41 survive a reboot.
5. Turn on sysstat collection and, once it has collected some data, show today's disk activity.

Hints

• iostat -xz 2 (ignore the first report), sudo iotop -o, pidstat -d 2 (packages sysstat and iotop).
• fio --name=... --filename=/data/fio.test --rw=write --bs=1M --size=... --direct=1, then --rw=randread --bs=4k.
• cat /sys/block/loop41/queue/scheduler; echo ... | sudo tee it.
• A rule in /etc/udev/rules.d/ (KERNEL==, ATTR{queue/scheduler}=), udevadm control --reload, udevadm trigger.
• systemctl enable --now sysstat; data comes every 10 minutes (the sa1 in /usr/lib64/sa/ or /usr/lib/sysstat/ collects a sample by hand: sa1 1 2); sar -d -p.