Basic HTML version of Foils prepared 15 March 1996

Foil 3 General Analysis of Parallel ADI Performance -- the Pipeline start-up

From Further PDE Solvers for the NAS Benchmarks CPSP713 Case studies in Computational Science -- Spring Semester 1996. by Geoffrey C. Fox
Consider N 3NAS grid decomposed onto P processors
We must start each ADI solve at the beginning i,j,k=1 or the end i,j,k= NNAS
For each sweep (where elements not all stored in same processor) start half the solves at the beginning and half at the end
  • in each case work towards the middle
After at most NNAS/2 steps, we will have "made it" to the middle on the initial solve and all processors will be active.
This algorithm is sensitive to latency as in racing to get middle fast(est), we have only performed part of one solve in each processor.
  • We communicate only one (5 component) element
Maximizing message size by partial solution of several solves before communication:
  • Decreases communication cost as fewer messages
  • But worsens load balance as delays start of processors near middle

Northeast Parallel Architectures Center, Syracuse University, npac@npac.syr.edu

If you have any comments about this server, send e-mail to webmaster@npac.syr.edu.

Page produced by wwwfoil on Sun Feb 22 1998