SparseBench: a sparse iterative benchmark

Jack Dongarra, Victor Eijkhout
Computer Science Department
University of Tennessee
Knoxville, TN 37996-1301, USA
and
Henk van der Vorst
Universiteit Utrecht
Utrecht, the Netherlands

Click here to see the number of accesses to this library.

For comments and questions, mail to sparsebench@cs.utk.edu.

About the benchmark

SparseBench is a benchmark suite of iterative methods on sparse data. Sparse matrices, such as derived from PDEs, form an important problem area in numerical analysis. Unlike in the case of dense matrices, handling them does not entail much reuse of data. Thus, algorithms for sparse matrices will be more bound by memory-speed than by processor-speed.

This benchmark uses common iterative methods, preconditioners, and storage schemes to evaluate machine performance on typical sparse operations. The benchmark components are:

Instructions

Download the file benchmark.tgz below.
Unpack it by

gunzip benchmark.tgz
tar -xf benchmark.tar  or  tar -x -f benchmark.tar
Go into the benchmark directory
cd SparseBench
and configure for your architecture
configure

Install the software and test your machine by

Test -m <machine name>
where "machine name" is an arbitrary name for your machine. If you run 'Test' more than once, only higher numbers are kept.

Mail the results back to the benchmark reporting authority by

Report -m <machine name>

You are strongly encouraged to read the files README and install.ps below, which are also part of the full tgz file.


Benchmark results

These are preliminary benchmark results, performed mostly on computers owned by the Innovative Computing Labs of the University of Tennessee. All tests report Megaflop rates on code that is compiled straight out of the box.

First we report the highest rate found for any problem. This was typically attained on a fairly small problem size, the implication being that the whole problem fit into cache.

Highest performance ranking
EV6 [a]759
Power3 [a]606
EV6 [b]438
Power3 [b]331
EV56262
PPC G4198
R12000 [a]155
UltraSparcII [a]154
Athlon154
R12000 [b]108
Origin106
UltraSparcII [b]102
PentiumIII96
LX16481
UltraSparcII [c]47
List of machines used
ProcessorMachineOwned byCompiler options
AthlonAthlon 600MHzR. Clint Whaley-O
EV56Dec Alpha, 433 MHzICL, University of Tennessee
EV6 [a]Geophysik, Freie Universitaet Berlin
EV6 [b]DEC Alpha, 500 MHZICL, University of Tennessee
LX164ALPHA, 533MHzICL, University of Tennessee
OriginSGI Origin, single processorNCSA-O
PPC G4Macintosh at 450MHzICL, University of Tennessee
PentiumIIIDell, dual 550MHzICL, University of Tennessee
Power3 [a]IBM quad 375MHz power3ICL, University of Tennessee
Power3 [b]IBM Power3, dual 200MHzICL, University of Tennessee
R12000 [a]SGI Octane, 270 MHzICL, University of Tennessee-O
R12000 [b]SGI IndigoICL, University of Tennessee-O
UltraSparcII [a]Sun Enterprise 450 model 1300, single 296MHzICL, University of Tennessee-O
UltraSparcII [b]Sun Ultra5ICL, University of Tennessee-O
UltraSparcII [c]Sun Enterprise, 248MHzICL, University of Tennessee

Next we filter problem by

and we report the "asymptotic performance" which will be the expected Mflop rate for large problems that overflow the cache. Asymptotic performance is determined by making a y=a+b/x fit through the observations, where x is the data set size in Mbytes.

Asymptotic performance on "gmres" problems
EV6 [a]216
Power3 [a]209
EV6 [b]168
Power3 [b]130
R12000 [a]78
Origin71
EV5660
Athlon44
LX16440
PentiumIII39
PPC G438
UltraSparcII [a]37
R12000 [b]30
UltraSparcII [c]23
UltraSparcII [b]23
Asymptotic performance on "cg" problems
EV6 [a]285
Power3 [a]254
EV6 [b]198
Power3 [b]110
Origin70
UltraSparcII [a]57
R12000 [a]52
PPC G445
LX16445
Athlon43
EV5640
PentiumIII37
UltraSparcII [c]26
UltraSparcII [b]21
R12000 [b]19
Asymptotic performance on "reg" problems
EV6 [a]285
Power3 [a]254
EV6 [b]198
Power3 [b]110
R12000 [a]78
Origin71
UltraSparcII [a]57
EV5655
PPC G445
LX16445
Athlon43
PentiumIII37
R12000 [b]28
UltraSparcII [c]26
UltraSparcII [b]21
Asymptotic performance on "crs" problems
Power3 [a]209
EV6 [a]209
EV6 [b]166
Power3 [b]130
Origin68
R12000 [a]63
EV5660
Athlon44
LX16440
PentiumIII39
UltraSparcII [a]35
R12000 [b]30
UltraSparcII [c]23
UltraSparcII [b]23
PPC G423
Asymptotic performance on "none" problems
Power3 [a]215
EV6 [a]205
EV6 [b]158
Power3 [b]88
R12000 [a]64
Origin63
UltraSparcII [a]40
EV5640
PPC G438
LX16436
Athlon33
PentiumIII27
UltraSparcII [c]26
R12000 [b]24
UltraSparcII [b]21
Asymptotic performance on "ilu" problems
EV6 [a]163
EV6 [b]132
Power3 [a]120
Power3 [b]90
R12000 [a]62
Origin57
EV5639
UltraSparcII [a]34
Athlon34
LX16433
PPC G431
PentiumIII27
R12000 [b]20
UltraSparcII [c]16
UltraSparcII [b]15

#########################################################################

file    readme

file    install.ps
file    install.pdf
for     Installation Guide for the Sparse Iterative Benchmark

file    benchmark.tgz
for     Benchmark of Conjugate Gradient methods, using sparse data storage
,	Sparse benchmark, version 0.9.7, released 17 Nov 2000.
,       Questions/comments to sparsebench@cs.utk.edu
by      Jack Dongarra, Victor Eijkhout, Henk van der Vorst

file    bench.ps
file    bench.pdf
for     Details and results of the Sparse Iterative Benchmark

#########################################################################