ProActive performance evaluation with NAS benchmarks and optimization of OO SPMD

1
ProActive performance evaluationwith NAS
benchmarksandoptimization of OO SPMD

• Brian Amedro Vladimir Bodnartchouk

2
Outline

• TimIt A profiling tool for ProActive
• OO SPMD model in ProActive
• Performance evaluation with NAS benchmarks
• Optimizing group communications

3
TimIt A profiling tool for ProActive
A ProActive feature to time and analyze
applications
4
OO SPMD model

• A parallel programming model
• Flexibility and high level of abstraction
• Strongly used in NAS benchmarks implementations

One To All
Scattering
Reduce operation
5
NAS Parallel Benchmarks

• Designed by NASA to evaluate benefits of high
  performance systems
• Strongly based on CFD
• 5 benchmarks (kernels) to test different aspects
  of a system

• Easy to implement thanks to OOSPMD pattern
• Tests performed on Sun 1.5 with RMI for ProActive
  and PGI 6.0 compiler for MPI

6
CG Kernel (Conjugate Gradient)

• 12000 calls
• 570 MB sent
• 1 min 32
• 65 comms

• Floating point operations
• Eigen value computation
• High number of unstructuredcommunications

7
MG Kernel (Multi Grid)

• 600 calls
• 45 MB sent
• 1 min 32
• 80 comms

• Floating point operations
• Solving Poisson problem
• Structured communications

8
IS Kernel (Integer Sort)

• 65 calls
• 22 MB sent
• 4 min 32
• 60 comms

• Keys ranking operations
• Bucket sort
• Large arrays in memory

9
EP Kernel (Embarrassingly Parallel)

• 6 calls
• 246 bytes sent
• 7 min 32
• 2 comms

• Random numbers generation
• Almost no communications

10
FT Kernel (Fourier Transformation)

• 22 calls
• 180 MB sent
• 1 min 32
• 40 comms

• Floating point operations
• Big messages 8 MB per call

11
Optimizing group communications

• Implement efficient group communication
• Minimize the TCP traffic
• Decrease the network congestion

Use clustering techniques to choose the better
algorithm to use
12
Ring all-to-all algorithm

• Best for large size communications
• Takes n-1 steps

A
A
B
D
C
D
D
C
B
A
C
D
B
B
A
C
C
B
A
D
step
1
2
3
13
Recursive doubling all-to-all algorithm

• Best for small size communications
• Takes log(n) steps

A
A
A
D
B
C
D
A
B
C
D
D
D
C
A
B
B
B
B
A
D
C
C
C
step
1
2
14
Conclusion

• TimIt easy and helpful profiling tool
• NAS benchmarks easy to implements with ProActive
  and OO SPMD patternhttp//www-sop.inria.fr/oasis/
  proactive/nas
• Good performances expected with future Sun Java 6
  and usage of Ibis RMI

15
Questions
?
16
MPI / ProActive
MPI ProActive Mpirun deployment MPI_Init a
ctivities creation MPI_Finalize MPI_Comm_Size ge
tMyGroupSize MPI_Comm_rank getMyRank MPI_Send
method call (setter and getter) MPI_Recv MPI_Ba
rrier barrier MPI_Bcast method
call MPI_Scatter method call with a
scatter group as parameter MPI_Gather result
of a group communication MPI_Reduce programmer's
method
Back