On May 7, 2004, at 11:36 AM, Michael Cariaso wrote: > After seeing this abstract > http://bioinformatics.oupjournals.org/cgi/content/abstract/20/7/1193 > about a newly released MPI version of clustalw, coming out of of > University of Western Australia > I was wondering of anyone had any experience to compare it to this > version > > http://bioinformatics.oupjournals.org/cgi/content/abstract/19/12/1585? > ijkey=12ACOyqyrKpFz&keytype=ref > coming from singapore? So here's my two pence on this. I never really got the whole benefit of mpi/pvm speed up deal for compute farms. There I said it, I'm sorry. :-) IMHO it takes away the flexibility of the embarrassingly parallel nature of clusters. With any of these approaches, you make hardware failure a critical component of your environment. However, and it's a real big however - you can turn around results really, really fast with these setups. They are great for providing rapid sub second turn around, where you need it for single jobs. For example web servers and services where you need rapid response for single, or small numbers of jobs. Other than that, if you are pipelining huge numbers of jobs you are still better off running 1,000 copies of clustal/blast/hmmer than passing messages about the place. See Amdahl for further details, keep q close to 0.95... etc. etc. Even on smaller clusters, you still win, I like 8 cpus to work as 8, not 7.5... Anyway, both these articles are great, and I'm sure the code works just fine. I used an early version of the SGI // clustal back at the EBI, and it was splendid, again it was for a web server, so it was ideal. I'd like to see how this stuff runs on a GB ethernet interconnect, but again it is not the bandwidth, it is always the latency that kills you... Guy/Tim - did you ever deploy that HMMer PVM cluster we talked about for the Pfam web site? j. -- James Cuff, D. Phil. Group Leader, Applied Production Systems The Broad Institute. 320 Charles Street, Cambridge, MA. 02141-2023. Tel: 617-252-1925 Fax: 617-258-0903