Presentation

Compiler and Architecture Seminar, November 11th, 2003
Subsetting SPEC When Measuring Results:
Research vs. Industry
A D’oh… Torus
[email protected]
IBM Labs in Haifa
© 2003 IBM Corporation
IBM Labs in Haifa
Production History
ISCA 2003 Panel, June 9th, 2003, San Diego, CA :
Subsetting SPEC when measuring results: valid or manipulative?
Mis
ulation: Partial and Misleading Use of
SPEC CPU 2000 in Computer Architecture Conferences
IEEE Micro, July/August 2003 (Vol. 23, No. 4)
The Use and Abuse of SPEC: An ISCA Panel
Revenge of the SPEC
2/11
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation
IBM Labs in Haifa
# of papers
Research: Computer Architecture Conferences
ISCA
Micro
HPCA
ISCA
Micro
HPCA-8
Number of papers published:
Papers that used a version of SPEC:
Earliest conference deadline:
SPEC CPU2000 announced:
3/11
Subsetting SPEC When Measuring Results: Research vs. Industry
HPCA
ISCA
HPCA-9
209
140 (66%)
December 2000
December 1999
© 2003 IBM Corporation
IBM Labs in Haifa
Industry: SPEC CPU2000
# of papers
Quotes from www.spec.org
“SPEC CPU2000 is the next-generation industry-standardized
CPU-intensive benchmark suite.”
“These benchmarks measure the performance of the processor,
memory and compiler on the tested system.”
Datapath
4/11
Memory
Multi-Thread
Media
Simulation
Subsetting SPEC When Measuring Results: Research vs. Industry
Compiler
Misc.
© 2003 IBM Corporation
IBM Labs in Haifa
Full Use
# of benchmarks used per paper
# of benchmarks used per paperNo Reason
- ( - )
- ( - )
( )
Reason
- ( - )
- ( - )
( Given
)
%
percent of papers
percent of papers
Research: Partial Use of CPU2000
%
%
%
%
%
%
%
Proc.% Proc.
CINT
CFP
%
Mem.
CINT
Mem.
CFP
Comp.
CINT
Comp.
CFP
# of Papers
Proc.
CINT
Proc.
CFP
Mem.
CINT
Mem.
CFP
Comp.
CINT
Comp.
CFP
Authors omit benchmarks, suites, and explanations!
5/11
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation
IBM Labs in Haifa
C
Research: The
The Chosen
Tower ofFew
Babel
F77
F90
# of paper appearances
# of paper appearances
art
equake
ammp
swim
mesa
applu
C++
gzip
vpr
parser
gcc
mcf
vortex
mgrid
lucas
wupwise
apsi
galgel
sixtrack
facerec
fma d
twolf
bzip
perlbmk
crafty
gap
eon
Average of 8 (CINT) and 6 (CFP) Benchmarks per Paper!
6/11
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation
IBM Labs in Haifa
Research vs. Industry: Top 10 Systems (CINT)
Top Industry SPECint ratios
( / )
7/11
Top Research SPECint ratios
( / )
Intel Pentium
Intel Pentium
AMD Athlon
AMD Athlon
AMD Opteron
HP Itanium
Dell Xeon
SGI Itanium
HP Itanium
AMD Opteron
HP Xeon
Dell Xeon
Fujitsu Pentium
HP Xeon
IBM Opteron
Dell Pentium
Einux Opteron
IBM Opteron
IBM POWER +
Bull Itanium
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation
IBM Labs in Haifa
Research vs. Industry: CFP2000
105 papers used CINT2000, 63 used
CFP2000 (60%)
CFP2000 considered more regular and
predictable
CFP2000 has higher Dcache miss rate
Only 35% of Memory-Hierarchy papers
use a majority of CFP2000
496 CINT2000 and 488 CFP2000
results published by SPEC
Linear Regression model of
SPECratio by the CPU clock rate
Multiple Linear Regression model of
SPECratio by speed and MemoryHierarchy
. %
. %
.
. %
PowerPC
Alpha
. %
. %
8/11
.
CINT
CFP
.
.
CFP CINT
BP BP
CFP CINT
I$
I$
CFP CINT
D$ D$
Linear Regression Multiple LR by MHz
by MHz
and MH
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation
IBM Labs in Haifa
Research vs. Industry: Top 10 Systems (CFP)
Top Industry SPECfp ratios
( / )
HP Itanium
HP Itanium
SGI Itanium
SGI Itanium
Bull Itanium
Bull Itanium
Dell Itanium
ION Itanium
ION Itanium
Dell Itanium
IBM Power +
Sun UltraSparc III
Intel Pentium
Fujitsu SPARC
HP Alpha
HP Alpha
AMD Athlon
IBM POWER +
HP Alpha
9/11
Top Research SPECfp ratios
( / )
C
Subsetting SPEC When Measuring Results: Research vs. Industry
V
NEC Itanium
© 2003 IBM Corporation
IBM Labs in Haifa
Research vs. Industry: Processor Architecture
MIPS
Alpha
Media/SIMD
HP-PA
IAPowerPC
SPARC
AMD
x
%
%
%
%
%
%
Research SPEC
CINT
CINT
Research SPEC
CFP
CFP
Use Architectures of the Future, not the Past!
SimpleScalar – A Two-Edged Sword
10/11
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation
IBM Labs in Haifa
Research
Industry
SPEC: Support research with simulation sized input sets
Vendors: Get simulation tools out to the field
11/11
Subsetting SPEC When Measuring Results: Research vs. Industry
© 2003 IBM Corporation