Evaluation of High Performance Fortran through Application Kernels

Presented at: HPCN 1997 at Vienna, Austria. April 1997.

Hon W Yau
EPCC & NPAC
Edinburgh & Syracuse University
UK & USA

Email: hwyau@epcc.ed.ac.uk

Outline of this Talk

• Motivation for HPF
• Motivation for the NPAC HPF Applications Suite
• The benchmark configurations
• Tour of the codes:
• Embarrassingly Parallel
• Synchronous-stencil operations
• Synchronous-matrix/vector operations
• Selected results
• Discussion

Why HPF

• Largely a benign extension of Fortran 90
• Use of directives to provide information to the compiler
• Portability with Fortran 90 serial code
• Crucial for continued code development
• Only agreed parallel language standard
• Portability amongst machine and compiler vendors
• Single thread of execution
• Clear distinction between distributed and replicated data
• Consistency of data amongst processors is implied

Why the HPF Applications Suite

• Aims:
• Demonstrate applicability of HPF to common codes
• Measure the maturity of available compilers (features)
• Benchmark available compilers
• Give feedback to the vendors
• Give feedback to the HPF-2.0 Forum meetings
• What it does not aim to do:
• Measure I/O performances
• Consciously test every HPF feature (cf, new ParkBench effort)

Short History of the Project

• Started at NPAC in early 1995:
• Collection of ~ 40 short codes, from various sources; mostly from previous NPAC projects
• TMC, MasPar dialects of data parallel Fortran
• Some message passing variants
• No Fortran 90 versions
• Only DEC HPF compiler was stable
• Unstable releases from PGI and APR
• Fortran 90 compilers were also quite rare...
• Not obvious which HPF features were implemented well by which compiler
• Lots of major rewrite of the codes

Short History of the Project (cont.)

• Today (April 1997):
• Pruned down to 16 functioning & correct codes
• Many codes were research codes, and required corrections
• About another 4 requiring further work
• Some codes still do not compile, or cause run-time errors
• Concentrated on F90 and HPF versions of codes
• Benchmarked and profiled, with measurements of computation and communication code sections
• No specially written message passing versions
• Fortran 90 compilers are now commonly available

Machine and Compiler Configurations

• Cray T3D at EPCC:
• PGI PGHPF v2.1 (June 1996)
• Cray F90 compiler for serial benchmark
• DEC Alphafarm with Gigaswitch at NPAC:
• DEC F90 & PSE v4.0 (February 1996)
• DEC F90 compiler for serial benchmark
• IBM SP2 at Cornell Theory Center:
• PGI PGHPF v2.1-1 (July 1996)
• PGI PGHPF in `F90' run-time mode for serial benchmark
• IBM SP2 at Cornell Theory Center:
• IBM XLHPF v1.1 (March 1996)
• IBM XLF90 F90 compiler for serial benchmark

Methodology

• Minimum elapsed times, over set of >= 8 runs
• Timing of the codes:
• Fortran 90 intrinsic `SYSTEM_CLOCK()' for wall-clock times
• Use of the PGI graphical profiler:
• Sections of code with communications
• Sections of code which are pure computations
• Measured speed-up with respect to single processor HPF total execution time:
• Serial Fortran 90 runs
• HPF with p=2, 4, 8 (and 16) processor runs
• To determine latency & scalability of HPF code

The HPF Application Codes

• Embarrassingly Parallel
• Alternating Direct Implicit
• 2D FFT
• NPAC's NAS EP Benchmark
• 2D Convolution
• Box Muller
• Wavelet Image Transformation
• Hough Image Transformation

HPF Application Kernels (cont.)

• Synchronous Algorithms
• Stencil Operation Codes:
• 2D percolation
• Potts model
• Binary phase quenching of the Cahn-Hilliard equation
• Segmented bitonic sort
• Direct N-body simulation
• Binomial stochastic volatility options pricing
• Matrix-vector Operation Codes:
• Gaussian factorisation
• Cholesky factorisation
• Hopfield Neural Network

Benchmark Results

Embarrassingly Parallel Algorithms (121 KBytes)

Synchronous-Stencil Algorithms (77 KBytes)

Synchronous-Matrix/Vector Operation Algorithms (47 KBytes)

Discussion

• HPF generally good for problems which are:
• Embarrassingly parallel
• Synchronous with stencil operations
• Performance:
• Expect speed-ups between 8 to 4 at 8 processors
• F90 to HPF latency `speed-down' below 2
• Of the compilers tested:
• PGHPF the most mature
• IBM XLHPF also good, considering its youth
• DEC HPF compiler disappointing compared to its F90
• IBM SP2 deals better with small problem sizes than the Cray T3D.

Discussion (cont.)

• Feedback on codes which do not compile or run
• Use of distributed data in common blocks?
• Tuning of the code can be tedious
• Feedback from the compiler can be better:
• Memory copy operations
• Data remappings
• Data locality (better control in HPF-2.0)
• Efficiency of a given line of code (sequentialisation?)
• Efficiencies of parallelised intrinsic functions
• Debugging is currently painful
• Time is right to do large complex HPF codes!

Further Information:

• On the NPAC HPFA Project:
• http://www.npac.syr.edu/hpfa/
• On the NPAC HPFA codes themselves:
• http://www.npac.syr.edu/hpfa/kernels/
• Draft release `HPF for SPMD programming: An Applications Overview':
• http://www.npac.syr.edu/hpfa/Papers/HPFforSPMD/
• To be mirrored on the EPCC web-site:
• http://www.epcc.ed.ac.uk/