[maker-devel] 12core speed check

[Carson Holt]

I can give a similar setup a try as well to see if anything is amiss in the
development version.  The expected behavior is that 1 and 2 cores should
have identical performance (as one process is always fully dedicated to
communication).

--Carson


From:  Ramón Fallon <ramonfallon at gmail.com>
Date:  Thursday, 14 March, 2013 11:19 AM
To:  <maker-devel at yandell-lab.org>
Subject:  [maker-devel] 12core speed check

Hi,

I was trying to tweak some of our machines to maximise Mpich2/Maker (svn rev
997)  throughput and describe one small set of results on this mailing list
to allow sharing of experiences.

I use the example input dataset "dpp_contig.fasta" with the original
sequence repeated 125 times within the same file (under different names of
course) to allow for a decent size run. This file totalled 4.019 megabases.
I use the dpp_proteins.fasta and  The maker_opts.ctl has "cpus=1" set as the
docs recommend for MPI.

Hardware is a standalone HP Proliant SL390 with two Intel X5675 @ 3.07GHz,
totalling 12 cores with 192GB RAM and 1TB disk (local, no NFS) running
Ubuntu 10.04 with 2.6.32-41 linux kernel

commandline was "mpiexec -n <#cores> maker" within a dedicated directory
containing all relevant files.

         #cores      time(mins)    Megabases/hr
              1           27.00            8.93
              2          126.25            1.91
              4           42.57            5.66
              6           25.42            9.49
              8           18.60           12.96
             10           16.67           14.47
             12           13.98           17.24

I attach a png file with graph. The upshot of this particular experiment is
that 2 processes show anomalous behaviour and that 6 processors are needed
to gain an advantage on the 1 processor run, while 12 processors achieves a
speed-up of nearly 2 on the 1 processor version.

I am now going to move on to a three node cluster with 2x 8core processors
each (so I can go up to 48 processors), so will report back with higher core
numbers. Any suggestions on further speed optimizations welcome.

Cheers / Ramón.
_______________________________________________ maker-devel mailing list
maker-devel at box290.bluehost.com
http://box290.bluehost.com/mailman/listinfo/maker-devel_yandell-lab.org

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://yandell-lab.org/pipermail/maker-devel_yandell-lab.org/attachments/20130314/30866d43/attachment-0003.html>