Dear All,
I am using LAM/MPI to parallelize computations that are accessible
from web-based applications (e.g., http://signs.bioinfo.cnio.es) that
run from a cluster. When the application starts, it is parallelized
over all the CPUs in the cluster.
After experiencing a few problems with having a single lam universe
for the apache user, I have now started using multiple lamboots (all
issued by the apache user), one for each web service request.
Basically, the CGI sets a different LAM_MPI_SESSION_SUFFIX before
issuing the lamboot [1] and launching the parallelized program. Once
the job is completed (or after it has run for the maximum allowed
time) a lamhalt and lamwipe are issued after setting again the
LAM_MPI_SESSION_SUFFIX environment variable. This seems to work just
fine, and is a lot more tolerant to problems with the application,
communication problems, etc. Problems that affect or result from one
run are contained to that run, and do not affect or perturb other lam
universes.
However, the FAQ and the manual seem to suggest that issuing multiple
lamboots should only be done in very special circumstances. I've been
searching the mailing list and doing several google searches, and
haven't found anything else relevant. What are the problems I am
likely to run into? Would anyone care to comment?
Thanks,
R.
[1] As explained, e.g., in the FAQ
(http://www.lam-mpi.org/faq/category4.php3#question20) or the manual
---p. 115.
--
Ramon Diaz-Uriarte
Bioinformatics Unit
Spanish National Cancer Centre (CNIO)
http://ligarto.org/rdiaz
|
