[Beowulf] Incorrect information.

Iozone capps at iozone.org
Tue Feb 21 08:34:19 PST 2006


Beowulf users,

    The information posted to this list concerning the proper mechanism
for running Iozone in parallel across the nodes is incorrect. You do
NOT use dsh or any other parallel tool. It is NOT needed. Iozone
already knows how to go parallel across nodes. See -+m option.
    If you follow this person's suggestion you will get wrong answers,
and prove that you have never read the manual, nor the man page,
nor even the help screen.  
    If you use the -+m option then Iozone will eliminate the 
straggler effect (some finish early, others finish late) and 
be able to obtain accurate throughput results.


Below is the mis-information that was posted. ( It is wrong, please
ignore this mis-information)

Enjoy,
Don Capps
capps at iozone.org

-----------
Egan Ford egan at sense.net 
Thu Jan 5 15:54:00 PST 2006 

  a.. Previous message: [Beowulf] Iozone on a Cluster 
  b.. Next message: [Beowulf] Newbie 
  c.. Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] 

--------------------------------------------------------------------------------

Find a parallel shell (e.g. pdsh), then write a script to run iozone,
something like:

echo ready
while [ ! -r /other/nfs/fs/go ]
do
	sleep .1
done
cd /nas/fs
iozone option option option ... >/other/nfs/fs/iozone.$(hostname -s).out
exit 0

Use pdsh or other to run the script on all nodes, e.g.:

date >foo;pdsh ...;date >>foo

Poor man parallel shell example:

date >foo;for i in node1 node2 node3 ...
do
ssh $i script &
done;wait;date >>foo

then type:

touch /other/nfs/fs/go

I do not recommend summing the output of iozone when running in parallel.
If you have a lot of clients they will not all start at the same time, some
may block on I/O for some period of time before starting inflating the
performance.  It is better to do the math yourself by using the timestamps
in the foo file.

A better solution to all of this is to use iometer.  With iometer you launch
a daemon on each node, then startup a iometer GUI (on Windows).  The GUI can
contact all the daemons and coordinate a benchmark in parallel, isolating
read from write.  With iozone in parallel you may get read and write
overlap.  You will need to patch iometer to use a directory instead of the
root of the FS or all daemons will try to read/write the same name.  I have
a patch somewhere to use a shell variable.

Another solution would be to use an MPIIO benchmark.

Normally I only use iozone for testing individual drive performance on nodes
and then check for a normal distribution.  I use MPIIO benchmarks and
iometer for testing SAN, NAS, and Parallel file systems.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.scyld.com/pipermail/beowulf/attachments/20060221/3ede7080/attachment.html


More information about the Beowulf mailing list