[Bioperl-l] operating systems and bio-computing
Mark Dalphin
mdalphin at amgen.com
Wed May 14 15:06:58 EDT 2003
Dear Thomas,
Here at Amgen, we are running DEC Alphas, SGI Octanes, HP-UX and a few Suns in
the Computational Biology group. We are switching to Linux due to the huge
cost advantage.
I have recently given up my SGI Octane running Irix in exchange for a Linux
box. It has saved me a great deal of time downloading and building
"freeware" which wasn't available (or out of date at the SGI web site). I
find my Linux box to be MUCH faster than the SGI (dual P4 at 1.4 GHz 1 Gbyte
RAM vs R10000 256MByte RAM) and easily as stable. And freeware for
bioinformatics installs quickly and easily.
When I benchmarked Blast (wu-blast) as well as several other programs we used
(FGenes, GenScan, RepeatMasker, various protein modeling and threading
tools), I found that a 1 GHz, dual CPU Linux box ran about 1/2 as fast as an
Alpha DS-10 (the SGI, HP and Suns weren't in the running for creating a
Compute Cluster) but cost about 1/5 as much. Since our problems scale well by
CPU (2x CPUs mean 2x faster), the cost advantage of Linux was obvious. We
built a Linux cluster (not Beowolf, just a cluster of headless work-stations
with a queueing system for distributing jobs) and almost all of our work is
done there instead of on the older Alpha compute cluster.
In this heterogenous environment, people are migrating to Linux (RedHat 7.3)
out of choice for almost all tasks except viewing protein models; there SGI
still prevails. And it sounds from your email as if people will migrate to
Linux for that work as well.
One final caveat: our SysAdmins find that running a Linux Compute Cluster is
somewhat more time-consuming than running the Alpha Cluster. Certain tasks
are well supported, but others are not nearly as clean as for the commercial
OSs. For example, we had a problem getting NFS to work at close to wire-speed
from the Linux boxes. Our SysAdmins finally located some patches to the
network drivers and the throughput increased significantly. A commercial OS
seller would probably have worked on that for us; instead we needed Google,
lots of time to experiment and some knowledgable SysAdmins. A summary of this
might be: for 'enterprise scale' computing the commercial OSs have some
advantages that cost alone ignores. For in individual work station, this is
probable much less of a problem as Red Hat has already tuned their product to
a single user.
Cheers,
Mark
PS: This reflects my personal opinion and personal experience; Amgen did buy
the Linux compute cluster, but my benchmarks were only part of the reason, I
am sure.
--
Mark Dalphin email: mdalphin at amgen.com
Mail Stop: 29-2-A phone: +1-805-447-4951 (work)
One Amgen Center Drive +1-805-375-0680 (home)
Thousand Oaks, CA 91320 fax: +1-805-499-9955 (work)
More information about the Bioperl-l
mailing list