PHAML Software: parallel models

chart showing parallel performance
Computation time for phases of PHAML with 1 to 16 processors and 150,000 vertices per processor. At 16 processors, PHAML get about 50% parallel efficiency.

PHAML supports both message passing for distributed memory parallel computers using MPI and shared memory parallelism using OpenMP.

PHAML uses one of six models of parallelism:

  • Master/Slave -- in this message passing model, the user executes one program, the master. The master program spawns the slave processes which perform the actual computations. If graphics are used, addition graphics processes are spawned.
  • SPMD (single program multiple data) -- in this message passing model, the user executes all of the programs, usually through some command like mpirun that starts them all. All of the processes run an identical program, but each program takes on the task of being either the master, one of the slaves, or a graphics engine.
  • OpenMP shared memory -- the user runs a single program. The program creates multiple threads that execute parallel sections of the code in parallel.
  • Hybrid OpenMP/MPI master/slave -- this model would normally be used on a parallel computer in which each node has multiple cores. It combines the master/slave model for message passing between nodes with the OpenMP model for parallelism between the cores on each node.
  • Hybrid OpenMP/MPI SPMD -- this model would normally be used on a parallel computer in which each node has multiple cores. It combines the SPMD model for message passing between nodes with the OpenMP model for parallelism between the cores on each node.
  • Sequential -- this is not parallel at all. This is a mode in which PHAML is run on a regular uniprocessor computer (or on one processor of a multiprocessor or multicore computer).

    The parallelism in PHAML is hidden from the user. One conceptualization is that the computational processes are part of a phaml_solution_type object, and hidden like all the other data in the object. In fact, one of the components of this data structure is a structure that contains information about the parallel processes. The user works only with the master process. In the master/slave model, when the master process calls phaml_create to initialize the data structure, the slave processes are spawned. When the master process calls any of the other public subroutines in module phaml, it sends a message to the slaves with a request to perform the desired operation. When phaml_destroy is called from the master process, the slave processes are terminated.

    PHAML was written to be portable not only across different architectures and compilers, but also across different message passing means. All communication between processes is isolated in one module. This module contains the data structures to maintain the information needed about the parallel configuration, and all operations that PHAML uses for communication, such as comm_init, phaml_send, phaml_recv, phaml_alltoall, phaml_global_max, etc. Thus to introduce a new means of message passing, one need only write a new version of this module that adheres to the defined API. PHAML contains versions for MPI 1 (without spawning), MPI 2 (with spawning), and a dummy version for sequential and OpenMP programs.

    Publications

    Mitchell, W.F., PHAML User's Guide , NISTIR 7374 , 2006. (original, pdf, 3.2M ) (latest revision, pdf)

    Mitchell, W.F., The Design of a Parallel Adaptive Multi-Level Code in Fortran 90, Proceedings of the 2002 International Conference on Computational Science, 2002. ( gzipped postscript, 50k)

    Back to PHAML home page


    Last change to this page: December 14, 2011
    Date this page created: April 2, 2007
    Home Page