Parallelizing the Symbolic Manipulation Program FORM

TL;DR

An analysis of the parallel platforms used and the structure of a parallel prototype of FORM is explained, and the status of the ongoing project of its parallelization is reported on.

Abstract

After an introduction to the sequential version of FORM and the mechanisms behind it we report on the status of our ongoing project of its parallelization. An analysis of the parallel platforms used is given and the structure of a parallel prototype of FORM is explained.

Figures (5)

• Figure 1: MPI(CH) bandwidth and latency results for the PingPong benchmark.
• Figure 2: Runtimes [sec] for parallel sorting on the alpha cluster, shown is the runtime vs. number of processors for different communication soft/hardware.
• Figure 3: Runtimes [sec] for parallel sorting on two other architectures: IBM SP2 (left) and Intel SMP (right) shown are the version with blocking and nonblocking communication.
• Figure 4: Runtimes [sec] vs. number of processors for a "simple" 3-loop Feynman-diagram on a IBM SP2 (left) and a 4 processor SMP machine (right) using its proprietary MPI-library. Shown are runtimes for different granularities of the input-term distribution.
• Figure 5: Runtimes [sec] vs. number of processors for a "simple" 3-loop Feynman-diagram on the DEC alpha cluster using MPI(CH) over the ParaStation II hard/software. Shown are runtimes for different granularities of the input-term distribution.