Parallel benchmark on multi-core CPUs

From Gerris

(Difference between revisions)
Jump to: navigation, search
Revision as of 23:09, 4 February 2010
Popinet (Talk | contribs)
(→MPICH1)
← Previous diff		 Revision as of 00:00, 5 February 2010
Popinet (Talk | contribs)

Next diff →
Line 10: Line 10:
; MPICH1 : 1.2.7-9.1ubuntu1 (packages mpich-shmem-bin, libmpich-shmem1.0-dev), ; MPICH1 : 1.2.7-9.1ubuntu1 (packages mpich-shmem-bin, libmpich-shmem1.0-dev),
; MPICH2 : 1.2-1ubuntu1.1 (packages mpich2, libmpich2-dev, libmpich2-1.2), ; MPICH2 : 1.2-1ubuntu1.1 (packages mpich2, libmpich2-dev, libmpich2-1.2),
-; OpenMPI : +; Open MPI : 1.3.2-3ubuntu1.1 (packages libopenmpi1.3, openmpi-common, libopenmpi-dev, openmpi-bin)
== MPICH1 == == MPICH1 ==
Line 34: Line 34:
[[Image:balance-mpich2.png]] [[Image:balance-mpich2.png]]
 +
 +{| border="1"
 +|-
 +! #CPUs
 +! Relative speedup
 +|-
 +| 1
 +| 1
 +|-
 +| 2 (not load-balanced)
 +| 1.46
 +|-
 +| 4 (not load-balanced)
 +| 2.44
 +|-
 +| 2 (load-balanced)
 +| 2.1
 +|-
 +| 4 (load-balanced)
 +| 3.5
 +|}
 +
 +== Open MPI ==
 +
 +[[Image:balance-openmpi.png]]
 +
 +{| border="1"
 +|-
 +! #CPUs
 +! Relative speedup
 +|-
 +| 1
 +| 1
 +|-
 +| 2 (not load-balanced)
 +| 1.46
 +|-
 +| 4 (not load-balanced)
 +| hanged at t = 12 (but this varies)
 +|-
 +| 2 (load-balanced)
 +| 2.1
 +|-
 +| 4 (load-balanced)
 +| hanged at startup
 +|}
 +
 +== Conclusions ==
 +
 +Either Open MPI triggers a bug in Gerris which the other two libraries do not, or Open MPI (or its Ubuntu packaging) have serious problems. Rumours and various posts on Ubuntu Launchpad and other sites suggest that this may be the case. Note also that each Open MPI instance of gerris2D takes about 150 MB of virtual memory in contrast to ~70 MB for MPICH2 and 5 MB for the serial version.
 +
 +The performance of Gerris/MPICH2 is very satisfactory taking into account the small problem size (~5000 elements/CPU).
 +
 +I have switched to MPICH2 on my development system.
== Parameter file == == Parameter file ==

Revision as of 00:00, 5 February 2010

This benchmark uses the the parallel Bénard–von Kármán Vortex Street example. Various implementations of MPI were tested, with and without load-balancing, on the following system:

  • Intel(R) Core(TM)2 Quad CPU Q9400 @2.66GHz, 64-bits
  • Ubuntu 9.10 64-bits
  • Linux popinet 2.6.31-17-generic #54-Ubuntu SMP Thu Dec 10 17:01:44 UTC 2009 x86_64 GNU/Linux
  • Gerris2D version 2010-01-29

MPI versions:

MPICH1 
1.2.7-9.1ubuntu1 (packages mpich-shmem-bin, libmpich-shmem1.0-dev),
MPICH2 
1.2-1ubuntu1.1 (packages mpich2, libmpich2-dev, libmpich2-1.2),
Open MPI 
1.3.2-3ubuntu1.1 (packages libopenmpi1.3, openmpi-common, libopenmpi-dev, openmpi-bin)

Contents

MPICH1

Image:balance-mpich1.png

#CPUs Relative speedup
1 1
2 (load-balanced) 1.33
4 (load-balanced) 1.97

MPICH2

Image:balance-mpich2.png

#CPUs Relative speedup
1 1
2 (not load-balanced) 1.46
4 (not load-balanced) 2.44
2 (load-balanced) 2.1
4 (load-balanced) 3.5

Open MPI

Image:balance-openmpi.png

#CPUs Relative speedup
1 1
2 (not load-balanced) 1.46
4 (not load-balanced) hanged at t = 12 (but this varies)
2 (load-balanced) 2.1
4 (load-balanced) hanged at startup

Conclusions

Either Open MPI triggers a bug in Gerris which the other two libraries do not, or Open MPI (or its Ubuntu packaging) have serious problems. Rumours and various posts on Ubuntu Launchpad and other sites suggest that this may be the case. Note also that each Open MPI instance of gerris2D takes about 150 MB of virtual memory in contrast to ~70 MB for MPICH2 and 5 MB for the serial version.

The performance of Gerris/MPICH2 is very satisfactory taking into account the small problem size (~5000 elements/CPU).

I have switched to MPICH2 on my development system.

Parameter file

Large outputs and movie generation were turned off, the single-CPU parameter file is: 

8 7 GfsSimulation GfsBox GfsGEdge {} {
  Time { end = 15 }
  Solid (x*x + y*y - 0.0625*0.0625)
  RefineSolid 6
  VariableTracer {} T
  Init {} { U = 1 }
  AdaptVorticity { istep = 1 } { maxlevel = 6 cmax = 1e-2 }
  AdaptGradient { istep = 1 } { maxlevel = 6 cmax = 1e-2 } T
  SourceViscosity 0.00078125
  EventBalance { istep = 1 } 0.1
  OutputTime { istep = 10 } stderr
  OutputTime { istep = 1 } balance
  OutputBalance { istep = 1 } balance
  OutputProjectionStats { istep = 10 } stderr
  OutputTiming { start = end } stderr
  OutputSimulation { start = end } end.gfs
}
GfsBox {
  left = Boundary {
    BcDirichlet U 1
    BcDirichlet T { return y < 0. ? 1. : 0.; }
  }
}
GfsBox {}
GfsBox {}
GfsBox {}
GfsBox {}
GfsBox {}
GfsBox {}
GfsBox { right = BoundaryOutflow }
1 2 right
2 3 right
3 4 right
4 5 right
5 6 right
6 7 right
7 8 right
Retrieved from "http://gfs.sourceforge.net/wiki/index.php/Parallel_benchmark_on_multi-core_CPUs"
Personal tools
communication