[Bioclusters] mpiblast and its database

[Jeremy Mann]

It has come unto this list, questions about mpiblast and its copying of
database segments to individual nodes. Here, we have a 20 node dual CPU
beowulf. Previously, I was using mpiformatdb -N 20. Then it hit me, even
though I start lamboot with cpus=2 for each node, in reality, blast only
runs one instance on each node. Now I get one segment per node. But what
happens when I start another job? Maybe mpiblast doesn't use the same CPU
for this next job, so it copies another segment, and so forth...

So this morning, I did mpiformatdb -N 44 (42 cpus + 1 dual master as Jason
Gans suggested) for protein nr. Then I ran the tests I used previously,
its a simple protein nr sequence which takes about 3 mins on one node.
Started top on various nodes, and ls in the local storage. The very first
time I ran it, of course, the segments (this time 2 per node) copied to
local storage. Speed wasn't great (nfsd was the problem), but obviously
faster that one node, total time for the first run was 1 min, 23 secs.

With the same top and ls windows open, I ran it again. This time, no
copying and each node still had its segment. Total time for the second run
was 5.9 seconds!

Then I thought, "Ok, maybe it caches the sequence so it didn't need to
copy segments anymore." So I took a different nr protein sequence and ran
this in mpiblast. To my surprise, there was no copying, and each node
still had its segments. Total time for the new sequence was 5.9 seconds.

I can only assume after this little experiment that formatting a db to the
exact number of nodes/cpus is the key.

Now, on to my next problem.... Keeping the segments in local storage so
another user doesn't have to go thru the copy process.

I'll keep you all posted.





-- 
Jeremy Mann
jeremy@biochem.uthscsa.edu

University of Texas Health Science Center
Bioinformatics Core Facility
http://www.bioinformatics.uthscsa.edu
Phone: (210) 567-2672