[Pvfs2-developers] I/O test 0.1 beta
Julian Martin Kunkel
Julian.Kunkel at web.de
Sat Aug 19 11:01:22 EDT 2006
Hi,
I uploaded a unfinished I/O testsuite for PVFS2 to
http://www.rzuser.uni-heidelberg.de/~jkunkel2/direct-io-0.1-dist.tgz.
This testsuite tries several I/O strategies.
There is still a lot to do, but I wanted to get some feedback from you guys,
and maybe one of you can run the test on a RAID system ?
The package consists of 2 source files, one uses either AIO or threads with
direct I/O or without and the other one uses a extension contained in a
modified libposix-aio directory.
This extension uses the linux kernels I/O submit and io_getevents function to
submit as many jobs as possible to the kernel. On the other hand io_getevents
tries to wait for the completion of the current pending job number. This is
the difference between my modifications and the aio default, without signals
one must periodically check if one of the jobs is finished or use aio_suspend
which blocks until at least one op of a given list is finished. Then the
whole list has to be iterated to figure out which one finished. With the
simple modifications this is not necessary, io_getevents returns only
finished ops, thus the iteration is unncecessary. This is only experimental
and I want to see the performance difference (if any) by feeding the kernel
scheduler as good as possible. So, far I haven't seen difference to the
threaded version, maybe I have to set the iteration to a higher value or try
on a different system ?
Right now I do not run the glibc / aio test, due to the fact that there
shouldn't be a performance difference for a single file. However, it can be
linked against libposix-aio, however this library is by default not thread
safe...
Also the package contains some (incomplete) results I got with one of our
machines. You can simple build the graphs by running "make graphs".
If you got some cool ideas in which direction the test should go, I would be
happy, right now I hope more data will help to figure out which test might be
appropriate.
Thanks,
julian
More information about the Pvfs2-developers
mailing list