Shared memory job (OpenMP)
|
|
- In shared-memory parallel computing, multiple processors (or "threads") perform tasks independently but share a common global memory. If a processor modified an array in the memory, all other processors can see this update as well
- A serial code can be parallelized by marking code sections, which should be run in parallel, with openMP directives.
- An openMP code can run on a single compute node and use up to maximum number of processors in that compute node, e.g., 24, 36, or 128 processors.
- To run an openMP code, define number of processors(threads) in $OMP_NUM_THREADS
$ export OMP_NUM_THREADS=8
$ bsub -n 8 ./program
|
Distributed memory job (MPI)
|
|
- In distributed-memory parallel computing, multiple processors (or "cores") perform tasks independently while each processor possesses its own private memory resource. If a processor modify an array in its memory, this processor has to communicate to other processors so that the others see this update.
- Massage Passing Interface (MPI) library implements distributed memory parallel computing which can be programmed in C, C++ and Fortran
- An MPI program can run on a single compute node as well as on multiple compute nodes
- An MPI program must be launched using "mpirun"
$ module load gcc/6.3.0 openmpi/4.0.2
$ bsub -n 240 mpirun ./program
|
Examples
Launch an openMP program
$ export OMP_NUM_THREADS=8
$ bsub -n 8 ./hello_omp
Generic job.
Job <8147290> is submitted to queue <normal.4h>.
Try launching an MPI program
$ unset OMP_NUM_THREADS
$ bsub -n 240 mpirun ./hello_mpi
MPI job.
Your environment is not configured for MPI.
Please load the module(s) needed by your job before executing 'bsub'.
Request aborted by esub. Job not submitted.
Load a compiler and compatible MPI module before launching an MPI code
$ module load intel openmpi
$ bsub -n 240 mpirun ./hello_mpi
MPI job.
Job <8147303> is submitted to queue <normal.4h>.
