A Survey
of MPI Implementations
| <-
Prev | Index |
Next -> |
NHSE ReviewTM:
Comments · Archive ·
Search
MPI is well-integrated with the PE/POE infrastructure. This
infrastructure provides, among other things:
- Parallel job startup, including optional automatic space sharing
of parallel applications.
- Signal propagation to all processes in a parallel application.
- Flexible handling of standard I/O: standard output may be
labeled by processor number and/or ordered by processor number;
standard input may be broadcast to all processes or sent to a single
process.
- An integrated batch queuing system called Loadleveler.
Unfortunately there remain a number of usability problems related to
PE/POE. Though not problems with the MPI implementation itself, these
interfere with the usability of MPI. For example:
- Numerous user-settable options for poe do not have reasonable
defaults.
- Signal propagation is not entirely foolproof, so that
orphaned processes are not uncommon
- Parallel jobs are still second-class citizens. For example,
there is no good way to see what parallel jobs are running on the
system. The standard utility, jm_status, produces verbose output
that is not easily parsed by humans.
- Despite improvement since earlier releases, the
batch system Loadleveler requires substantial local customization and
tools to be useful.
Copyright © 1996
| <-
Prev | Index |
Next ->
|
NHSE ReviewTM:
Comments · Archive
· Search