IBM TotalStorage DS6000 Series: Performance Monitoring and Tuning

300 IBM TotalStorage DS6000 Series: Performance Monitoring and Tuning

granularity in kilobytes. Note that this must be a power of 2, and that the coarsest granularity

is 128 kilobytes.

8.6 Bonnie

Bonnie is a performance measurement tool written by Tim Bray. For a more complete

description and documentation on Bonnie go to the following Web site:

http://www.textuality.com/bonnie/

Bonnie performs a series of tests on a file of known size. If the size is not specified, Bonnie

uses 100 Mb, but that probably is not enough for a big modern server. Bonnie works with

64-bit pointers if you have them.

For each test, Bonnie reports the bytes processed per elapsed second, per CPU second, and

the % CPU usage (user and system).

8.6.1 Benchmarks

Bonnie does the following benchmarks:

򐂰 char output with put() / putc_unlocked()

The result is the performance a program will see that uses putc() to write single

characters. On most systems, the speed for this is limited by the overhead of the library

calls into the libc, not by the underlying device. The _unlocked version (used if bonnie is

called with -u) may be considerably faster, as it involves less overhead.

򐂰 char input with getc() / getc_unlocked()

The result is the performance a program will see that uses getc() to read single characters.

The same comments apply as to putc().

򐂰 Block output with write()

This is the speed with which your program can output data to the underlying file system

and device writing blocks to a file with write(). As writes are buffered on most systems, you

will see numbers that are much higher than the actual speed of your device, unless you

sync() after the writes (option -y) or use a considerably larger size for your test file than

your OS will buffer. For Linux, this is almost all your main memory.

If called with the -o_direct option, this operation (and the ones described in the following

two paragraphs) is done with the O_DIRECT flag set, which results in direct DMA from

your hardware to userspace, thus avoiding CPU overhead copying buffers around. This

will prevent buffering, and gives a much better estimate of real hardware speed, also for

small test sizes.

򐂰 Block input with read()

This is the speed with which you can read blocks of data from a file with read(). The same

comment as for block output regarding your OS doing buffering for you applies, with the

exception that using -y does not help to get realistic numbers for reading. You would need

to flush the buffers of the underlying block device, but this turns out to not be trivial, as you

first have to find out the block device. It would be a Linux-only feature anyway.

򐂰 Block in/out rewrite

Bonnie does a read(), changes a few bytes, write()s the data back and reread()s it. This is

a pattern that occurs on some database applications. Its result tells you how your

operating root (/) file system can handle such access patterns.