A Performance Comparison Between the Programming Models of MPI and OpenMP

1
A Performance Comparison Between the Programming
Models of MPI and OpenMP

• Jay Hoeflinger

2
Our first attempt at making an OpenMP version of
ROCFLO involved using a traditional loop-parallel
paradigm. It showed poor speedup compared to
the MPI version of ROCFLO. We guessed bad
locality was to blame.
3
(No Transcript)
4
The Goal

• Determine the performance costs of OpenMP
  programs, especially the constructs which do not
  scale well.

5
The Methodology

• Start with the MPI version of ROCFLO, which
  scales very well.
• Do performance experiments.
• Write an OpenMP version of ROCFLO which
  duplicates the structure of the MPI code exactly
  (including message passing calls). This removes
  locality as a problem.
• Measure performance and if any parts do not scale
  well, determine why.

6
Barrier Cost MPI vs OpenMP(Origin 2000)
7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
So Locality was not the Problem!

• The problems turned out to be
• I/O which doesnt scale
• ALLOCATE which doesnt scale
• our non-scaling reduction implementation
• our first-cut messaging infrastructure which,
  could be improved