User guide

6–SHMEM Description and Configuration
SHMEM Benchmark Programs
6-28 IB0054606-02 A
Additional SHMEM micro-benchmark programs are included to measure get and
put performance with randomized PE selection and randomized target memory
locations, all-to-all communication patterns using put, barrier and reduce:
QLogic SHMEM random access benchmark
shmem-rand: randomized put/get benchmark
This is actually a hybrid SHMEM/MPI code, so a binary is provided per
supported MPI implementation. It has the following command line options:
Usage: shmem-rand [options] [list of message sizes].
Message sizes are specified in bytes (default = 8)
Options: See Table 6-5
-b INT batch size, number of concurrent operations (default = 64)
-f force order for bifurcation of PEs based on rank order
-h displays the help page
-l INT set minimum message size (default = 2)
-m INT sets the maximum message size (default = 4194304)
Table 6-5. QLogic SHMEM random access benchmark options
Option Description
-a use automatic (NULL) handles for NB ops (default explicit han-
dles)
-b use a barrier every window
-c INTEGER specify loop count (see also -t)
-f fixed window size (default is scaled)
-h displays the help page
-l enable communication to local ranks
-m INTEGER[K] memory size in MB (default = 8MB): or in KB with a K suffix
-n use non-pipelined mode for NB ops (default pipelined)
Table 6-4. QLogic SHMEM micro-benchmarks options
Option Description