Performance
See also
Contents
dstat
`dstat` is one of the more valuable tools for monitoring system performance. The output columns can be easily customized.
The default options are -cdngy. The following are options I commonly use. Many other are described in the manpage.
-c --cpu system, user, idle, wait, hardware interrupt, software interrupt -d --disk disk read, write -f --full full listing when using certain options (--cpu, --int, --disk, --net, --swap) -g --page page in, out -i --int interrupts (see also --full option, --I option, and review /proc/interrupts) -l --load load average -m --mem memory used, buffers, cache, free -n --net network receive, send -r --io I/O read, write -s --swap swap used, free -y --sys system interrupts, context switches --vm vm hard pagefaults, soft pagefaults, allocated, free
`dstat` also has many Python plugins stored in /usr/share/dstat/.
Some statistics require the lm-sensors package. Run `sensors-detect` after installing.
tools
apttitude -q -y install iozone3 stress cpuburn sysstat iotop hddtemp fio sysbench lmbench phoronix-test-suite iperf
The fio tool does not have a lot of documentation, but it looks interesting. The homepage is just a git repository: fio. Under Ubuntu install the fio package. For documentation see /usr/share/doc/fio/, especially /usr/share/doc/fio/examples/.
lmbench is ancient (its homepage is nearly 20 years old!), but it still works.
drive IO testing and performance measurement
Basic read and write speed testing. The options used below tests for IOPS, not bytes/sec. These options also favor sequential streaming of large blocks of data.
iozone -a -s 1048576 -g 1G -i 0 -i 1 -O
Bonnie++ tests will work OK with the defaults. You do have to set the user. Note that root is not normally recommended. Note that bonnie is the same as bonnie++ (bonnie is a sym-link to bonnie++). The output of bonnie++ is stupidly difficult to read. There is no way to fix this. It also dumps out CSV data, which is even harder to read without a spreadsheet.
bonnie -u root:root
drive stress testing
The stress command generates pure loads. It does not attempt to measure how the system handles this. You can combine this with other tools to get performance measurements.
This generates stress on /dev/sda. While this is running you may want to run "iostat 1 300 /dev/sda" in a different window.
stress --hdd 10 /dev/sda
CPU stress and burn
Install the Ubuntu package cpuburn. For each CPU core your system has run one instance of `burnP6` (for Intel P6 processors). Monitor the CPU usage and system load using `htop` or the tool of your choice. Monitor the temperature using `sensors` or some ACPI tool.
burnP6 & burnP6 & burnP6 & burnP6 & watch -n1 sensors killall burnP6