Linux Benchmark Notes

CPU Benchmarks

Other (non-linux) systems are added for interest.

Linux/Unix nbench

    http://www.tux.org/~mayer/linux/bmark.html

These benchmarks are single-thread/single-processor benchmarks.  Multi-thread/multi-processor performance is NOT measured.

Baseline (LINUX)    : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38.  The table is roughly ordered by comparison of measurements with more weight for INT and FP.
 
CPU/Machine5.720
L2 Cache
OS
Compiler
C-lib
MEM
INT
FP
AMD Opteron244 1.8GHz/333
PC3200
Tyan Tiger K8W (S2875) 2 CPUs
 1M
 Linux 2.4.23-s1
Slackware 9.1
 gcc version 3.2.3
 libc:ld-2.3.2.so
 12.251
 8.857
 17.863
PowerPC 970 (2.2) 2GHz/1GHz
PC3200
Apple Power Mac G5 Dual
 512K
 Mac OS X 10.3.5 (Darwin 7.5.0)
 gcc version 3.3 20030304 (Apple Computer, Inc. build 1495)
 unknown version
 11.415
 8.342
 17.305
AMD Athlon XP 1800+ 1.533GHz/333 PC2700
Asus A7S333
 256K
 Linux 2.4.20r7
Slackware 9.0b
 gcc version 3.2.2
 libc:ld-2.3.1.so
 9.570
 7.863
 14.879
AMD K7 1.33GHz/266
PC2100		 256K Mandrake 8.1 gcc version 2.96 20000731 (Mandrake Linux 8.1 2.96-0.62mdk) unknown version 7.800 6.540 12.673
AMD Athlon Thunderbird 1.33GHz/266
PC2100 
Asus A7M266		 256K Linux 2.4.4p7
Mandrake 7.1		 gcc version 2.95.3 19991030 (prerelease) glibc:libc-2.1.3 7.290 6.601 15.193
PowerPC G4 (3.3) 1GHz/167
PC2700
Apple PowerBook G4 17"
 256K
 Mac OS X 10.3.5 (Darwin 7.5.0)
 gcc version 3.3 20030304 (Apple Computer, Inc. build 1495)
 unknown version
 7.073
 6.383
 7.815
AMD Athlon Thunderbird 1.33GHz/266
PC2100 
Asus A7M266		 256K Linux 2.2.19
Mandrake 7.1		 gcc version 2.95.3 19991030 (prerelease) glibc:libc-2.1.3 6.713 6.050 13.918
AMD K7 1.155GHz
133MHz SDRAM		 256K Mandrake 7.1 gcc version 2.95.3 19991030 (prerelease) unknown version (static) 6.338 5.720 13.133
AMD Athlon Thunderbird 1.13GHz/266
PC2100 
Asus A7M266
 256K Linux 2.4.2
RedHat 7.1		 egcs-2.91.66 19990314/
(egcs-1.1.2 release)		 unknown version (static) 5.965 4.938 8.272
AMD Athlon Thunderbird 1GHz/266
PC2100 
Asus A7M266
 256K Linux 2.4.17
RedHat 7.2		 gcc version 2.96 20000731 (Red Hat Linux 7.1 2.96-98) unknown version (static) 5.334 5.071 9.626
PowerPC G4 667MHz/133
PC133
Apple PowerBook G4 15"
 256K L2
1M L3
 Mac OS X 10.3.5 (Darwin 7.5.0)
 gcc version 3.3 20030304 (Apple Computer, Inc. build 1495)
 unknown version
 4.669
 4.228
 5.483
AMD Athlon 800MHz/200 100MHz SDRAM
Asus K7V
 512K Linux 2.4.17
RedHat 7.2		 gcc version 2.96 20000731 (Red Hat Linux 7.1 2.96-98) unknown version (static) 4.240 4.044 7.867
Intel Pentium III 800MHz/133
PC600 RDRAM 512MB
Dell OptiPlex GX300
 256K
 Linux 2.4.7
RedHat
7.2
 gcc version 2.96 20000731 (Red Hat Linux 7.1 2.96-98)
 unknown version (static)
 3.674
 3.118
 7.184
Intel Pentium III 800MHz
133MHz SDRAM
Asus CUSL2
 256K Linux 2.4.2
RedHat 7.1b		 egcs-2.91.66 19990314/
(egcs-1.1.2 release)		 unknown version (static) 3.495 2.744 6.235
Intel Pentium III 667MHz
133MHz SDRAM 256MB
Dell OptiPlex GX110
 256K
 Linux 2.4.18 Slackware 8.1-b2
 gcc version 2.95.3 20010315 (release)
 unknown version
 2.995
 2.788
 5.557
Sun UltraSPARC II/V9/FP 750MHz 8GB
Fire V880, 4 CPUs
 64K D
32K I
8M L2
 Solaris 8
 cc -xO3 -s Sun WokShop 6 update 1 C 5.2 2000/09/11
 unknown version
 2.737
 2.430
 4.684
Intel Pentium III 600MHz
133MHz SDRAM 128MB
 256K
 Linux 2.54.7-10 RedHat 7.1
 gcc version 2.96 20010731
(release)
 unknown version
 2.681
 2.311
 4.836
Intel Pentium III 600MHz
133MHz SDRAM
Intel CC820		 512K Linux 2.4.2
RedHat 7.0+		 egcs-2.91.66 19990314/
(egcs-1.1.2 release)		 unknown version (static) 2.577 2.100 4.178
Intel Pentium III 600MHz
RDRAM
Intel VC820
 512K Linux 2.4.2
RedHat 7.1b		 egcs-2.91.66 19990314/
(egcs-1.1.2 release)		 unknown version (static) 2.531 2.078 4.180
Sun UltraSPARC II/V9/FP 750MHz 8GB
Fire V880, 4 CPUs
 64K D
32K I
8M L2
 Solaris 8
 gcc version 2.95.2 19991024
(release)
 unknown version
 1.749
 1.948
 3.710
Intel Pentium III 500MHz
100MHz SDRAM
Intel SE440BX-2
 512K Linux 2.2.14
RedHat 6.2+		 egcs-2.91.66 19990314/
(egcs-1.1.2 release)		 unknown version (static) 2.162 1.755 3.540
IBM S/390 zSeries z800 2066
636 BogoMIPS
1.8GB
Linux 2.4.9-37
RedHat 7.2
 gcc version 2.95.3 20010315 (release)
 unkown version
 1.226
 1.642
 1.682
Sun UltraSPARC II/V9/FP
400MHz 8GB
Ultra-Enterprise 450, 8 CPUs
 16K D
16K I
8M L2
 Solaris 8
 gcc version 2.95.2
19991024 (release)
 unknown version
 1.033
 1.332
 2.279
Intel Mobile Pentium III 266MHz
Aaeon PCM-5896		 256K Linux 2.4.4
RedHat 7.1		 egcs-2.91.66 19990314/
(egcs-1.1.2 release)		 unknown version (static) 0.764 0.799 1.277
MIPS R5000 180MHz IP22
SGI Indy/R5000		 32K/32K IRIX 6.5 cc unknown version (static) 0.664 0.716 1.814
MIPS R4400/R4000 200MHz IP22
SGI Indigo2		 16K/16K IRIX 6.5 cc unknown version (static) 0.679 0.599 1.248

These benchmarks are single-thread/single-processor benchmarks.  Multi-thread/multi-processor performance is NOT measured.

Memory

Here's a handy memory reference.

name
 aka
 type
 clock
 bits
 bandwidth
GB/s
 comment
PC66
SDRAM
 66
 64
 0.528
 10ns
PC100
SDRAM
 100
 64
 0.800
PC133
SDRAM
 133
 64
 1.064
PC150
SDRAM
 150
 64
 1.200
PC600/16
RDRAM
 532/266
 16
 1.064
PC700/16
RDRAM
 712/366
 16
 1.424
PC800/16
RDRAM
 800/400
 16
 1.600
PC800/32
RDRAM
 800/400
 32
 3.200
PC1066/16
RDRAM
 1066/533
 16
 2.132
PC1066/32
RDRAM
 1066/533
 32
 4.264
PC1600
 DDR200
 SDRAM
 200/100
 64
 1.600
PC2100
 DDR266
 SDRAM
 266/133
 64
 2.128
PC2400
 DDR300
 SDRAM
 300/150
 64
 2.400
PC2700
 DDR333
 SDRAM
 333/166
 64
 2.664
PC3000
 DDR350
 SDRAM
 350/176
 64
 2.800
PC3200
 DDR400
 SDRAM
 400/200
 64
 3.200

General comments

Benchmarks are controversial.

They only measure the performance of the applications in the benchmark suite.  For real world performance, the enlightened scientist will measure the target application on actual hardware.  But porting and measuring actual end-user applications is laborious, time consuming, and often cost prohibitive.  So benchmark suites include a range of applications, including coverage of memory-intensive, scalar-integer-intensive, floating-point-intensive, and (recently) communications-intensive applications.

Performance is affected by tuning.  Some benchmarks may be "tuned" on one system better than another.  Some manufactures have tuned compilers or drivers specifically for know benchmarks, resulting in misleading performance expectations for other applications.  Target applications can be tuned significantly via compiler optimization options, profiling, focussed tuning of code in hot spots, and adjustment of software architecture for improved workflow and reduced communication overhead.  But tuning your application on a specific platform is laborious, time-consuming, and often cost-prohibitive.  So benchmark suites are used to give an approximation.

Corporations usually present biased and sometimes misleading performance information, e.g., Sun and IBM want to present their systems in the best possible light.  Benchmark information is often unavailable and unverifiable, or selected to present a systems best features and to hide performance problems.  Some performance "measurements" are just extrapolations of  small system measurements onto larger configurations.

Performance can be a complex system engineering problem.  Breaking performance into aspects of integer, floating-point, memory, communication, disk, etc. doesn't necessarily provide relevant measurements.  A systems approach should be taken for improving the performance of your application, especially with regards to partitioning and communications on multi-processor systems.

There is a lack of freely available "standard" benchmarks.  SPEC (Standard Performance Evaluation Corporation) http://www.specbench.org/ charges for their benchmarks.  Most Windows-based benchmarks are not portable to other environments.  Even after recent web searches, I still find myself left with the BYTE benchmarks.  I'm still looking for something that is both freely available and "standard", and I'm willing to consider anything "better."

Summary - for now, we're stuck with nbench for MEM, INT, and FP metrics.
NoBell Home - gjm - last update 10/7/2004