6 – MPI Sample Applications
D000006-000 Rev A
6-5
Q
Pallas has known scalability limitations, especially in its
AllToAll
phase. This phase
can simultaneously perform up to 4MB transfers to-and-from all nodes at once. The
downside is a system must have approx 10*NP MB of memory available per process
for Pallas data to run this benchmark. Therefore, for a small cluster (approximately
16 processors or less), it is modest at 160MB. However, for a larger cluster
(approximately 256 processors or greater), it is rather large at 2.5GB.
As such, it is recommended that Pallas be used for smaller runs (2-32 processes)
or that it be recognized that the benchmark is likely to fail (or swap Linux to death)
at larger process counts. Depending upon the amount of memory in the system and
the numbers of processes to run, the
VIADEV_MEM_REG_MAX
parameter in
/opt/iba/src/mpi_apps/mpi.param.pallas
may need to be edited.
To run pallas:
1.
cd /opt/iba/src/mpi_apps
2.
./run_pmb NP
where
NP
is the number of processes to run, such as:
./run_pmb 4
During this benchmark the
/opt/iba/src/mpi_apps/mpi.param.pallas
config
file is used.
Summary of Contents for Fast Fabric
Page 1: ...D000006 000 Rev A Page i Q S i m p l i f y Fast Fabric Users Guide...
Page 2: ...Fast Fabric Users Guide Q Page ii D000006 000 Rev A...
Page 38: ...3 Getting Started Upgrading IB software 3 24 D000006 000 Rev A Q...
Page 148: ...6 MPI Sample Applications Pallas 6 6 D000006 000 Rev A Q...
Page 166: ...B Fast Fabric Configuration Files Port List Files B 14 D000006 000 Rev A Q...
Page 168: ...C Configuration of IPoIB Name Mapping C 2 D000006 000 Rev A Q...
