Linux AIO is notoriously bad; but it is hard to understand what the weaknesses of a filesystem implementation in a particular kernel version are. The fsqual tool provides a way to qualify a filesystem's aio implementation.
See this blog post for more information.
The test matrix checks the following variations:
- Whether the write changes the size of the file or not
- Whether the write first touches its offset range, or alternatively overwrites already-written data
- Whether the write uses sector granularity (typically 512 bytes) or block granularity (typically 4096 bytes)
- Whether the writes happen concurrently, or only one at a time
- Whether O_DSYNC is in use (as is typical for commitlogs) or not
Install the following packages:
libaio-devel(libaio-devon Debian and derivatives)xfsprogs-devel.x86_64(xfslibs-devon Debian and derivatives)gcc-c++(g++on Debian and derivatives)make
To build, simply run
make
Change to a directory under the mountpoint to be tested, and run fsqual.
fsqual will report, for each test scenario, whether io_submit was truly asynchronous. This is done by measuring context switches during the io_submit call.