[BioBrew Users] Unable to get mpiblast running

[Bastian Friedrich]

Hi,

On Tuesday 11 April 2006 06:50, Glen Otero wrote:
> Bastian-
>
> Any progress running mpiBLAST?

yesterday, I compiled mpiblast 1.3, linked against LAM/MPI (lam-gnu 
7.1.1 as included in the Rocks "hpc" roll). This seems to be running 
fine so far.

As I said, your mpiblast 1.4 with MPICH worked for the "mini data set" 
as described by you (p53/p53db). I was not successful getting it 
running with any other data set I tested - the Hs.seq.uniq/IL2RA 
combination as described in your tutorial, as well as genbank. This 
seems to be an MPICH problem I'm encountering (although it is unclear 
whether openmpi would run if this problem did not exist... :)

When compiling mpiblast 1.4 with LAM/MPI, I get an interesting result: 
The Hs.seq.uniq/IL2RA combination _does_ run (I just found out about 
the --debug flag yesterday), but it takes more than 8 hours (!) to 
finish. This is a little too long, I think ;) Here is a little snippet 
of it's output:
=============================
[...]
[9]     26.7763 rank 9 exiting successfully
[9]     26.7763 MPI startup  time is 0
[0]     26.9766 Processing result_id: 0
[0]     26.9767 Query 0 now has 2 / 28 fragments searched
[0]     26.9767 Got result count of 1
[0]     26.9767 receiveResults to read 1 records
[0]     27.4961 Read SeqAnnot
[0]     27.4961 Reading bioseqs for frag/query 4/0
[0]     108.928 Receive was successful -- about to merge
[0]     108.94  Query results have been merged
[0]     108.952 Processing result_id: 0
[0]     108.952 Query 0 now has 3 / 28 fragments searched
[0]     108.952 Got result count of 1
[0]     108.953 receiveResults to read 1 records
[0]     109.473 Read SeqAnnot
[0]     109.473 Reading bioseqs for frag/query 5/0
[0]     245.66  Receive was successful -- about to merge
[...]
=============================

The concrete blast jobs seem to finish in a reasonable time (the 26 
seconds are more or less the same as when running mpiblast 1.3), but 
the receiving and merging of results takes hours. The load on the 
compute nodes drops to zero, but stays high (1) on the head node. I 
have not fully investigated this topic, as even a rather small job 
takes hours to finish :)

This topic - or a similar one - is currently discussed on the mpiblast 
mailinglist, to which I subscribed yesterday.

When compiling mpiblast 1.4 with OpenMPI, load on all nodes (head and 
compute nodes) stays high for hours; I have not fully investigated this 
topic, but there seems to be some incompatibility between mpiblast and 
OpenMPI... mpiblast 1.3 does not work, either; OpenMPI is not actively 
supported by mpiblast, so I'll stick with the alternatives for now.

So, once more: mpiblast 1.3.0 with LAM/MPI works well; all alternatives 
failed so far. We will see if mpiblast 1.3.0 is sufficient for our 
needs later today.

If there is any more information I can provide, feel free to ask.

Thx again, Regards,
   Bastian

-- 
 Bastian Friedrich                  bastian at bastian-friedrich.de
 Adress & Fon available on my HP   http://www.bastian-friedrich.de/
\~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\
\ ... All the world's a stage, and I missed rehearsal.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://bioinformatics.org/pipermail/biobrew-users/attachments/20060411/f7c3e0e9/attachment.bin