Digital PDFs

EB-N0105-51

March 1993

43 pages

Original

0.3MB

Document:	AlphaAXP Server Family Performance Brief - DEC OSF/1 AXP
Order Number:	EB-N0105-51
Revision:
Pages:	43
Original Filename:	https://web-docs.gsi.de/~kraemer/COLLECTION/ftp.cuhk.edu.hk/pub/doc/digital/alpha-axp-osf1-server.pdf

OCR Text

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP

DEC 10000 Model 610 AXP Server

DEC 7000 Model 610 AXP Server

DEC 4000 Model 610 AXP Server

DEC 3000 Model 500S AXP Server

INSIDE
DEC 3000 Model 400S AXP Server

Digital Equipment Corporation
March 1993
First Edition
EB-N0105-51

Benchmark results for:
• FFT &
• SPEC
Convolution
• AIM
• SLALOM
• Dhrystone
• Basic Real-Time
• LINPACK
Primitives
• CERN
• Rhealstone
• DN&R Labs
• Perfect
CPU2
• Livermore
Loops

First Printing, March 1993
The information in this document is subject to change without notice and should not be construed as a commitment by Digital
Equipment Corporation.
Digital Equipment Corporation assumes no responsibility for any errors that may appear in this document.
Any software described in this document is furnished under a license and may be used or copied only in accordance with the
terms of such license. No responsibility is assumed for the use or reliability of software or equipment that is not supplied by
Digital Equipment Corporation or its affiliated companies.
Restricted Rights: Use, duplication, or disclosure by the U.S. Government is subject to restrictions as set forth in
subparagraph (c)(1)(ii) of the Rights in Technical Data and Computer Software clause at DFARS 252.227 7013.
Copyright  1993 Digital Equipment Corporation
All rights reserved.
Printed in U.S.A.
The following are trademarks of Digital Equipment Corporation: AXP, Alpha AXP, the AXP logo, the AXP signature, and DEC.

The following are third-party trademarks:
AIM is a trademark of AIM Technology.
HP is a registered trademark of Hewlett-Packard Company.
RS 6000 and IBM are trademarks of International Business Machines Corporation.
OSF/1 is a trademark of Open Software Foundation, Inc.
SPEC, SPECratio, SPECmark89, SPECint92, SPECfp92, SPECrate_int92, SPECrate_fp92 are trademarks of the Standard
Performance Evaluation Corporation.
NFS, SUN, and SPARC are trademarks of Sun Microsystems, Inc.
UNIX is a registered trademark of UNIX System Laboratories, Inc.

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Contents
Digital’s Alpha AXP Server Family ........................................................................................ 6
DEC 3000 AXP Servers ................................................................................................................. 6
DEC 4000 AXP Servers ................................................................................................................. 6
DEC 7000 AXP Servers ................................................................................................................. 6
DEC 10000 AXP Servers ............................................................................................................... 6
SPEC Benchmark Suites ............................................................................................................... 10
SPEC SFS Release 1.0 ................................................................................................................ 10
SPEC CINT92 and CFP92 ........................................................................................................... 12
SPEC Homogeneous Capacity Method based on SPEC CINT92 and CFP92 ............................ 16
SPEC Release 1 ........................................................................................................................... 19
AIM Suite III Multiuser Benchmark Suite ..................................................................................... 21
Dhrystone Benchmarks ................................................................................................................ 24
LINPACK 100x100 and 1000x1000 Benchmarks ......................................................................... 26
CERN Benchmark Suite ................................................................................................................ 28
DN&R Labs CPU2 Benchmark ...................................................................................................... 30
Livermore Loops ........................................................................................................................... 31
Fast Fourier Transform and Convolution .................................................................................... 32
Fast Fourier Transform ................................................................................................................. 32
Convolution ................................................................................................................................... 32
SLALOM Benchmark ..................................................................................................................... 33
Basic Real-Time Primitives ........................................................................................................... 34
Rhealstone Benchmark ................................................................................................................. 36
Perfect Benchmarks Suite ............................................................................................................ 40

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Information about Performance ............................................................................................ 41
References ..................................................................................................................................... 42

List of Figures
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9
Figure 10
Figure 11
Figure 12
Figure 13
Figure 14
Figure 15
Figure 16
Figure 17
Figure 18
Figure 19
Figure 20
Figure 21
Figure 22
Figure 23
Figure 24
Figure 25
Figure 26
Figure 27

SPEC SFS Release 1.0 NFS Server Response Time vs. Throughput ..................... 11
SPEC CINT92 Results for Alpha AXP Servers ......................................................... 13
SPEC CINT92 Results for Competitive Systems ...................................................... 13
SPEC CFP92 Results for Alpha AXP Servers .......................................................... 15
SPEC CFP92 Results for Competitive Systems ....................................................... 15
SPECrate_int92 Results for Alpha AXP Servers ...................................................... 17
SPECrate_int92 Results for Competitive Systems ................................................... 17
SPECrate_fp92 Results for Alpha AXP Servers ....................................................... 18
SPECrate_fp92 Results for Competitive Systems .................................................... 18
SPECmark89 Results for Alpha AXP Servers .......................................................... 20
SPECmark89 Results for Competitive Systems ....................................................... 20
Dhrystone V1.1 and V2.1 Results for Alpha AXP Servers ........................................ 25
Dhrystone V1.1 Results for Competitive Systems .................................................... 25
LINPACK Double-Precision Results for Alpha AXP Servers .................................... 27
LINPACK Double-Precision Results for Competitive Systems ................................. 27
CERN Results for Alpha AXP Servers ...................................................................... 29
CERN Results for Competitive Systems ................................................................... 29
DN&R Labs CPU2 Results for Alpha AXP Servers ................................................... 30
DN&R Labs CPU2 Results for Competitive Systems ................................................ 30
Livermore Loops Results for Alpha AXP Servers ..................................................... 31
Basic Real-Time Primitives Process Dispatch Latency Results ................................ 35
Basic Real-Time Primitives Interrupt Response Latency Results ............................. 35
Rhealstone Component–Task-switch Time .............................................................. 36
Rhealstone Component–Preemption Time ............................................................... 37
Rhealstone Component–Semaphore-shuffle Time ................................................... 38
Rhealstone Component–Intertask Message Latency ............................................... 38
Perfect Results for Alpha AXP Servers ..................................................................... 40

List of Tables
Table 1
Table 2
Table 3

Digital’s Alpha AXP Server Family Benchmark Results .............................................. 7
SPEC SFS Release 1.0 Benchmark Suite Results ................................................... 11
AIM Suite III Benchmark Suite Results for Alpha AXP Servers ................................ 22

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Table 4
Table 5
Table 6
Table 7
Table 8

March 1993

AIM Suite III Benchmark Suite Results for Competitive Systems ............................. 23
FFT Benchmark Results for Alpha AXP Servers ...................................................... 32
SLALOM Results ....................................................................................................... 33
Basic Real-Time Primitives Results .......................................................................... 34
Rhealstone Benchmark Results ................................................................................ 39

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Digital’s Alpha AXP Server Family
This document presents the results of industry-standard benchmarks on Digital’s Alpha
AXPTM servers in the DEC OSF/1 AXPTM operating system environment. Alpha AXP
servers are based on Digital’s breakthrough 64-bit RISC architecture.
DEC 3000 AXP Servers
The DEC 3000 Model 400S AXP desktop is an entry-level server or multiuser system that
provides affordable power for small workgroup computing and distributed applications. The
DEC 3000 Model 500S AXP server is a deskside system that provides a blend of
performance and expansion to meet the challenges of the most demanding workgroup
computing. The DEC 3000 Model 400S AXP server runs at a CPU clock speed of 133 MHz,
and the DEC 3000 Model 500S AXP server at 150 MHz.
DEC 4000 AXP Servers
DEC 4000 AXP servers deliver superior computing power in office-sized or rackmountable
packages. Designed for symmetric multiprocessing, multiuser and technical server
applications, DEC 4000 AXP servers meet the needs of technical and scientific markets.
Optimized for speed and availability, DEC 4000 AXP servers will satisfy the commercial
segment as well. The DEC 4000 AXP server runs at a CPU clock speed of 160 MHz and can
be configured with one or two CPU boards.
DEC 7000 AXP Servers
DEC 7000 AXP servers deliver unsurpassed data center performance and price/performance.
DEC 7000 AXP servers offer single or multiprocessing capability for such commercial and
technical applications as transaction processing, general ledger, securities trading, signal
processing, molecular modeling, and imaging. The DEC 7000 AXP server runs at a CPU
clock speed of 182 MHz and can be configured with up to six processors.
DEC 10000 AXP Servers
DEC 10000 AXP servers are Digital’s highest performance Alpha AXP systems. Designed
to meet rigorous demands of compute-intensive, large enterprise applications, DEC 10000
AXP servers can be configured with up to six processors. The DEC 10000 AXP server
provides the fastest enterprise server performance in the industry—at a fraction of the price
of a traditional mainframe or supercomputer. The DEC 10000 AXP server runs at a CPU
clock speed of 200 MHz.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Table 1 Digital’s Alpha AXP Server Family Benchmark Results

DEC 3000
Model
400S AXP

DEC 3000
Model
500S AXP

SPEC SFS Release 1.0
SPECnfs_A93 (ops/sec)
Average Response Time (msec)
SPECnfs_A93 Users

537
26.0
54

601
21.6
60

tbd
tbd
tbd

SPECint92
SPECfp92

74.7
112.5

84.4
127.7

94.6
137.6

103.1
176.0

116.5
193.6

Benchmark

SPECrate_int92
SPECrate_fp92

1,763
2,662

1,997
3,023

DEC 4000
Model
610 AXP

2,198
3,247

DEC 7000 DEC 10000
Model
Model
610 AXP
610 AXP

2,572
4,179

2,765
4,368

SPECmark89

111.1

126.1

137.3

175.6

192.1

AIM Benchmark Suites
Performance Rating
Maximum User Load
Maximum Throughput

70.3
485
688.7

82.9
649
812.9

tbd
tbd
tbd

Perfect Benchmarks Suite
(MFLOPS-geometric mean)

18.4

20.7

23.1

26.4

29.2

Dhrystone
V1.1 (instructions/sec)
V2.1 (instructions/sec)

235,939
238,095

266,487
263,157

297,345
294,117

330,577
333,333

363,743
357,142

LINPACK 64-bit Double-precision
100x100 (MFLOPS)
1000x1000 (MFLOPS)

26.0
91.7

29.6
103.5

35.0
110.1

36.9
137.8

40.5
151.1

CERN Benchmark Suite

18.8

21.3

23.2

DN&R Labs CPU2 (MVUPs)

185.0

209.1

225.6

254.6

279.3

Livermore Loops
(geometric mean)

17.4

19.5

22.3

25.4

27.8

SLALOM (patches)

5,776

6,084

6,496

6,902

7,248

tbd = to be determined

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Benchmark Descriptions and Results:
•
•
•
•
•
•
•
•
•
•
•
•

March 1993

SPEC Benchmark Suites
AIM Benchmark Suites
Dhrystone
LINPACK 100x100 and 1000x1000
CERN Benchmark Suite
DN&R Labs CPU2
Livermore Loops
FFT & Convolution
SLALOM
Basic Real-Time Primitives
Rhealstone Benchmark
Perfect Benchmark Suite

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SPEC Benchmark Suites
SPEC (Standard Performance Evaluation Corporation) was formed to identify and create
objective sets of applications-oriented tests, which can serve as common reference points and
be used to evaluate performance across multiple vendors’ platforms.
SPEC SFS Release 1.0
In March 1993, SPEC announced its SFS (System-level File Server) Release 1.0 Benchmark
Suite. This suite is the first, industry-wide, standard method for measuring and reporting
NFS file server performance.
SFS Release 1.0 contains the 097.LADDIS File Server Benchmark, which supersedes the
nhfsstone benchmark. 097.LADDIS emulates an intense software development environment
where system components important to file serving (i.e., network, file system, disk I/O, and
CPU) are heavily exercised. This benchmark shows a server’s ability to handle NFS
throughput and the speed with which the throughput is processed.
SPEC’s metrics for 097.LADDIS are:
•

SPECnfs_A93 operations/second, which is peak NFS throughput.

•

Average Response Time associated with the peak NFS throughput level, measured in
milliseconds.

•

SPECnfs_A93 Users, an arbitrary, conservative, derived number of users supported
at the reported peak NFS throughput level (SPECnfs_A93 operations/second) at an
associated response time of less than or equal to 50 milliseconds.

Table 2 shows SPEC SFS Release 1.0 results for both Alpha AXP and competitive servers.
Figure 1 shows their NFS server response time versus throughput results at various loads.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Table 2 SPEC SFS Release 1.0 Benchmark Suite Results

System (Network)
DEC 3000/500S AXP (1 FDDI)

Memory
(MB)

Disk
Controllers

Number
of
Disks

256

4 SCSI-2

SPECnfs_A93
(ops/sec)

Average
Response
Time
(msec)

SPECnfs_A93
Users

Number of
File
Systems

601

21.6

DEC 3000/400S AXP (1 FDDI)

128

4 SCSI-2

537

26.0

Auspex NS/5500 (8 Enet)

256

3 SP-III

1,703

49.4

170

Auspex NS/5500 (4 Enet)

224

2 SP-III

933

43.9

Auspex NS/5500 (2 Enet)

208

1 SP-III

466

43.6

HP 9000/H50 (1 FDDI)

576

3 SCSI-2

1,014

47.9

101

HP 9000/755 (1 FDDI)

640

2 SCSI-2

859

36.9

IBM RS 6000/560 (3 Enet)

128

2 SCSI

410

45.8

Average Response Time
(msec)
60

Auspex NS/5500
8 Enet

Auspex NS/5500
4 Enet

IBM RS 6000/550
3 Enet

HP 9000/H50
FDDI

Auspex NS/5500
2 Enet
30

HP 9000/755
FDDI

DEC 3000/400S
FDDI
DEC 3000/500S
FDDI

0
0

200

400

600

800

1000

1200

1400

1600

1800

NFS Throughput
(SPECnfs_A93 NFS Operations/Second)

Figure 1 SPEC SFS Release 1.0 NFS Server Response Time vs. Throughput

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SPEC CINT92 and CFP92
In January 1992, SPEC announced the availability of two new benchmark suites, CINT92
and CFP92. Each suite provides performance indicators for different market segments
because each has different workload characteristics. SPEC CINT92 is a good base indicator
of CPU performance in a commercial environment. SPEC CFP92 may be used to compare
floating-point intensive environments, typically engineering and scientific applications.
SPEC CINT92

CINT92, the integer suite, contains six real-world application benchmarks written in C. The
geometric mean of the suite’s six SPECratios is the SPECint92 figure. CINT92 suite
includes the following application classes:

•

008.espresso–Circuit theory

•

022.li–LISP Interpreter

•

023.eqntott–Logic design

•

026.compress–Data compression

•

072.sc–UNIX spreadsheet

•

085.gcc–GNU C compiler

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP

Figure 2 SPEC CINT92 Results for Alpha AXP Servers
SPECint92

140.0

120.0

116.5
103.1

100.0

94.6
84.4

80.0

74.7

60.0

40.0

20.0

0.0
DEC 3000/400S AXP

DEC 3000/500S AXP

DEC 4000/610 AXP

DEC 7000/610 AXP

DEC 10000/610 AXP

Figure 3 SPEC CINT92 Results for Competitive Systems
SPECint92
140.0

120.0

100.0

80.0

78.2

59.8

60.0

59.2
53.3

45.6

40.0

61.7

36.5
31.4

48.4

47.8

34.8

20.0

0.0
HP 9000/715/50
HP 9000/827S
HP 9000/867S
IBM RS/6000/360/570
IBM RS/6000/580
IBM RS/6000/980B
HP 9000/735/755
HP 9000/847S
HP 9000/887S
IBM RS/6000/370
IBM RS/6000/970B
SUN SPARC 10-41

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SPEC CFP92

CFP92 consists of fourteen real-world applications; two are written in C and twelve in
FORTRAN. Five of the fourteen programs are single precision, and the rest are double
precision. SPECfp92 equals the geometric mean of this suite’s fourteen SPECratios. This
suite contains the following application classes:

•

013.spice2g6–Circuit design

•

015.doduc–Monte Carlo simulation

•

034.mdljdp2−Quantum chemistry

•

039.wave5–Maxwell equations

•

047.tomcatv–Coordinate translation

•

048.ora–Optics ray tracing

•

052.alvinn–Robotics; neural nets

•

056.ear–Human ear modeling

•

077.mdljsp2–Single precision version of 034.mdljdp2

•

078.swm256–Shallow water model

•

089.su2cor–Quantum physics

•

090.hydro2d–Astro physics

•

093.nasa7–NASA math kernels

•

094.fppp–Quantum chemistry

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Figure 4 SPEC CFP92 Results for Alpha AXP Servers
SPECfp92
250.0

200.0

193.6
176.0

150.0

137.6
127.7
112.5

100.0

50.0

0.0
DEC 3000/400S AXP
DEC 4000/610 AXP
DEC 10000/610 AXP
DEC 3000/500S AXP
DEC 7000/610 AXP

Figure 5 SPEC CFP92 Results for Competitive Systems
SPECfp92
250.0

200.0

150.6

150.0

141.6
133.2
124.8

118.2
101.0

97.0

100.0
72.1

65.1

50.0

0.0
HP 9000 715/50

HP 9000/887S
HP 9000/735/755

March 1993

IBM RS/6000/370

IBM RS/6000/360/570

IBM RS/6000/970B

IBM RS/6000/580

Digital Equipment Corporation

SUN SPARC 10-41

IBM RS/6000/980B

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SPEC Homogeneous Capacity Method based on SPEC CINT92 and CFP92
SPEC Homogeneous Capacity Method benchmarks test multiprocessor efficiency.
According to SPEC, "The SPEC Homogeneous Capacity Method provides a fair measure for
the processing capacity of a system — how much work can it perform in a given amount of
time. The "SPECrate" is the resulting new metric, the rate at which a system can complete
the defined tasks....The SPECrate is a capacity measure. It is not a measure of how fast a
system can perform any task; rather it is a measure of how many of those tasks that system
completes within an arbitrary time interval (SPEC Newsletter, June 1992)." The SPECrate is
intended to be a valid and fair comparative metric to use across systems of any number of
processors.
The following formula is used compute the SPECrate:
SPECrate = #CopiesRun * ReferenceFactor * UnitTime / ElapsedExecutionTime
SPECrate_int92 equals the geometric mean of the SPECrates for the six benchmarks in
CINT92. SPECrate_fp92 is the geometric mean of the SPECrates of the fourteen
benchmarks in CFP92.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Figure 6 SPECrate_int92 Results for Alpha AXP Servers

SPECrate_int92
3000
2765
2572
2500
2198
1997

2000
1763
1500

1000

500

0
DEC 3000/400S AXP

DEC 3000/500S AXP

DEC 4000/610 AXP

DEC 7000/610 AXP

DEC 10000/610 AXP

Figure 7 SPECrate_int92 Results for Competitive Systems
SPECrate_int92
3000

2500

2000

1854

1500
1332
1215

1201
1000
816

1263

890

500

0
HP 9000/827S

March 1993

HP 9000/847S

HP 9000/867S

HP 9000/887S

HP 9000/890-1

Digital Equipment Corporation

IBM RS/6000/370

SUN SPARC 10-41

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP

Figure 8 SPECrate_fp92 Results for Alpha AXP Servers
SPECrate_fp92

5000
4368
4179
4000

3247
3023

3000

2662

2000

1000

DEC 3000/400S AXP

DEC 4000/610 AXP
DEC 3000/500S AXP

DEC 10000/610 AXP
DEC 7000/610 AXP

Figure 9 SPECrate_fp92 Results for Competitive Systems

SPECrate_fp92
5000

4000
3490
3000
2612
1937

2000
1335

1503

1483
1180

1000

0
HP 9000/827S

HP 9000/867S
HP 9000/847S

HP 9000/890-1
HP 9000/887S

Digital Equipment Corporation

SUN SPARC 10-41
IBM RS/6000/370

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SPEC Release 1
In October 1989, SPEC introduced SPEC Release 1, a benchmark suite that measures
CPU-intensive, single stream performance of uniprocessor systems. SPEC Release 1
consists of ten portable programs similar to those found in technical environments. Four
programs are written in C and primarily test integer performance. The remaining six
programs are written in FORTRAN and measure floating-point performance.
SPECmark89 is the metric for this suite.
SPECmark89 represents the geometric mean of the ten benchmark SPECratios. The
SPECratio for a benchmark is the quotient derived from dividing the SPEC Reference Time
by a particular machine’s corresponding run time. The SPEC Reference Time is the time
that it takes a DEC VAX 11/780 to run each benchmark (in seconds).

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Figure 10 SPECmark89 Results for Alpha AXP Servers

SPECmark89
250.0

200.0

192.1
175.6

150.0
126.1

137.3

111.1
100.0

50.0

0.0
DEC 3000/400S AXP
DEC 4000/610 AXP
DEC 10000/610 AXP
DEC 3000/500S AXP
DEC 7000/610 AXP

Figure 11 SPECmark89 Results for Competitive Systems

SPECmark89

250.0

200.0

146.8

150.0

136.1
126.2

126.4

117.0

95.7

100.0

69.0

50.0

0.0
HP 9000/715/50

HP 9000/887S
HP 9000/735/755

IBM RS/6000/370
IBM RS/6000/980B
IBM RS/6000/360/570
IBM RS/6000/580

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
AIM Suite III Multiuser Benchmark Suite
Developed by AIM Technology, the AIM Suite III Benchmark Suite was designed to
measure, evaluate, and predict UNIX multiuser system performance of multiple systems. It
uses 33 functional tests, and these tests can be grouped to reflect the computing activities of
various types of applications.
AIM Suite III is designed to stress schedulers and I/O subsystems and includes code that will
exercise TTYs, tape subsystems, printers, and virtual memory management. The benchmark
will run until it reaches either the user-specified maximum number of simulated users or
system capacity.
The 33 subsystem tests, each of which exercises one or more basic functions of the UNIX
system under test, are divided into six categories based on the type of operation involved.
The categories are as follows:
•

RAM

•

Floating Point

•

Pipe

•

Logic

•

Disk

•

Math

Within each of these six categories, the relative frequencies of the subsystem tests are evenly
divided (with the exception of small biases for add-short, add-float, disk reads, and disk
writes).
AIM Suite III contains no application level software. Each simulated user runs a
combination of subsystem tests. The load that all simulated users put on the system is said to
be characteristic of a UNIX time-sharing environment. The mix of subsystem tests can be
varied to simulate environments with differing resource requirements. AIM provides a
default model as a representative workload for UNIX multiuser systems and the competitive
data that AIM Technology publishes is derived from this mix of subsystem tests.
The AIM Performance Rating identifies the maximum performance of the system under
March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
optimum usage of CPU, floating point, and disk caching. At a system’s peak performance,
an increase in the workload will cause a deterioration in performance. AIM Maximum User
Load Rating identifies system capacity under heavy multitasking loads, where disk
performance also becomes a significant factor. Throughput is the total amount of work the
system processes, measured in jobs/minute. Maximum throughput is the point at which the
system is able to process the most jobs per minute. AIM verifies the results of the Suite III
run by licensed vendors and uses the source data for their AIM Performance Report service.
Digital’s Alpha AXP server family’s AIM Suite III benchmark results are shown in the
following table:
Table 3 AIM Suite III Benchmark Suite Results for Alpha AXP Servers

System

Performance
Rating

Maximum User
Loads

Maximum
Throughput
(jobs/minute)

DEC 3000/500S AXP
(192MB memory, 8 disks)

82.9

649

812.9

DEC 3000/400S AXP
(128MB memory, 3 disks)

70.3

485

688.7

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Shown below are Digital Alpha AXP servers’ competitors’ AIM results.
Table 4 AIM Suite III Benchmark Suite Results for Competitive Systems
System

Performance
Rating

Maximum User
Loads

Maximum
Throughput
(jobs/minute)

HP 9000/755
(64MB memory, 2 disks)

71.7

580

703.1

HP 9000/735
(32MB memory, 2 disks)

71.7

422

703.1

HP 9000/887S
(256MB memory, 9 disks)

69.2

620

677.8

HP 9000/750
(64MB memory, 2 disks)

46.7

388

457.6

HP 9000/715/50
(32MB memory,2 disks)

33.7

252

330.0

IBM RS/6000 580
(128MB memory, 4 disks)

62.1

518

609

SUN SPARC 10 Model 41
(64MB memory, 3 disks)

44.3

203

433.8

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Dhrystone Benchmarks
Developed as an Ada program in 1984 by Dr. Reinhold Weicker, the Dhrystone benchmark
was rewritten in C in 1986 by Rick Richardson. Dhrystone measures processor and compiler
efficiency and is representative of systems programming environments. Dhrystones are most
commonly expressed in Dhrystone instructions per second.
Dhrystone V1 and V2 vary considerably. Version 1.1 contains sequences of code segments
that calculate results never used later in the program. These code segments are known as
"dead code." Compilers able to identify the dead code can eliminate these instruction
sequences from the program. These compilers allow a system to complete the program in
less time and result in a higher Dhrystones rating. Dhrystones V2 was modified to execute
all instructions.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Figure 12 Dhrystone V1.1 and V2.1 Results for Alpha AXP Servers
Dhrystones/second
(in thousands)
400.0
V1.1

363.7 357.1

350.0

330.6 333.3

V2.1

297.4 294.1

300.0
266.5 263.2

250.0

235.9 238.1

200.0
150.0
100.0
50.0
0.0
DEC 3000/400S AXP

DEC 3000/500S AXP

DEC 4000/610 AXP

DEC 7000/610 AXP

DEC 10000/610 AXP

Figure 13 Dhrystone V1.1 Results for Competitive Systems
Dhrystones/second
(in thousands)

400.0
V1.1

350.0
300.0
250.0

215.5

200.0

192.4

150.0
100.0

123.8

109.4

92.6

50.0

0.0

HP 9000/715/50

HP 9000/735/755

HP 9000/827S/847S

HP 9000/867S

SUN SPARC 10-41

Note: Hewlett-Packard has reported Dhrystone V2.0 results, but not V2.1. SUN has not reported Dhrystone
V2.0 or V2.1.

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
LINPACK 100x100 and 1000x1000 Benchmarks
LINPACK is a linear equation solver written in FORTRAN. LINPACK programs consist of
floating-point additions and multiplications of matrices. The LINPACK benchmark suite
consists of two benchmarks.
1.

100x100 LINPACK solves a 100x100 matrix of simultaneous linear equations.
Source code changes are not allowed so that the results may be used to evaluate the
compiler’s ability to optimize for the target system.

1000x1000 LINPACK solves a 1000x1000 matrix of simultaneous linear equations.
Vendor optimized algorithms are allowed.

The LINPACK benchmarks measure the execution rate in MFLOPS (millions of
floating-point operations per second). When running, the benchmark depends on
memory-bandwidth and gives little weight to I/O. Therefore, when LINPACK data fit into
system cache, performance may be higher.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Figure 14 LINPACK Double-Precision Results for Alpha AXP Servers
MFLOPS
180.0
100x100

160.0

151.1

140.0

137.8

1000x1000

120.0
110.1
103.5
100.0

91.7

80.0
60.0
40.0

40.5

36.9

35.0

29.6

26.0
20.0
0.0
DEC 3000/400S AXP

DEC 4000/610 AXP

DEC 3000/500S AXP

DEC 7000/610 AXP

DEC 10000/610 AXP

Figure 15 LINPACK Double-Precision Results for Competitive Systems
MFLOPS

180.0
100x100

160.0

140.0
1000x1000

120.0
107.0

105.0

100.0
84.0
80.0

60.0

41.0

40.0

20.0

N/A

22.2

17.1

13.7

13.2

0.0

38.1

N/A

31.0

25.9

N/A

22.4

N/A

7.3

HP 9000/715/50
HP 9000/847S
IBM RS 6000/360/570
IBM RS/6000/970B
SUN SPARC 10-41
HP 9000/735/755
HP 9000/867S
IBM RS 6000/370
IBM RS 6000/580/980B

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
CERN Benchmark Suite
In the late 1970’s, the User Support Group at CERN (the European Laboratory for Particle
Physics) collected from different experimental groups a set of typical programs for event
simulation and reconstruction and created the CERN Benchmark Suite. In 1985, Eric
McIntosh, system analyst, redefined the tests in order to make them more portable and more
representative of the then current workload and FORTRAN 77.
Presently, the CERN Benchmark Suite contains four production tests: two event generators
(CRN3 and CRN4) and two event processors (CRN5 and CRN12). These applications are
basically scalar and are not significantly vectorizable nor numerically intensive.
Additionally, several "kernel" type applications were added to supplement the production
tests to get a feel for compilation times (CRN4C), vectorization (CRN7 and CRN11), and
character manipulation (CRN6).
The CERN Benchmark Suite metric is CPU time. Results are normalized to a DEC VAX
8600, and the geometric mean of the four production tests’ ratios yields the number of CERN
units. CERN units increase with increasing performance.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Figure 16 CERN Results for Alpha AXP Servers
CERN Units

35.00
30.00

29.00
26.00

25.00

23.20

21.30
20.00

18.80

15.00
10.00
5.00
0.00

DEC 3000/400S AXP DEC 3000/500S AXP

DEC 4000/610 AXP

DEC 7000/610 AXP

DEC 10000/610 AXP

Figure 17 CERN Results for Competitive Systems
CERN Units
35.00
30.00
27.20
25.00
20.00
15.30

15.00
10.00

13.50
10.70

9.96

10.70

5.00
0.00
HP 9000/720

HP 9000/735
HP 9000/750

IBM RS 6000/970
IBM RS 6000/550

SUN SPARC 10

Source: "CERN Report" (11/23/92).

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
DN&R Labs CPU2 Benchmark
DN&R Labs CPU2, a benchmark from Digital News & Review magazine, is a floating-point
intensive series of FORTRAN programs and consists of thirty-four separate tests. The
benchmark is most relevant in predicting the performance of engineering and scientific
applications. Performance is expressed as a multiple of MicroVAX II Units of Performance
(MVUPs).
Figure 18 DN&R Labs CPU2 Results for Alpha AXP Servers
MVUPs
300.0

279.3
254.6

250.0
225.6
200.0

209.1
185.0

150.0
100.0
50.0
0.0
DEC 3000/400S AXP
DEC 4000/610
DEC 3000/500S AXP

DEC 7000/610

DEC 10000/610 AXP

Figure 19 DN&R Labs CPU2 Results for Competitive Systems
MVUPs
300.0
250.0
200.0

186.4

150.0
100.0

91.8
48.1

50.0
0.0
HP 9000/715/50

HP 9000/735/755

SUN SPARC 10-41

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Livermore Loops
This benchmark, also known as Livermore FORTRAN Kernels, was developed by the
Lawrence National Laboratory in Livermore, CA. The laboratory developed this benchmark
to evaluate large supercomputer systems. Computational routines were extracted, 24 sections
of code in all, from programs used at the laboratory in the early 1980’s to test scalar and
vector floating performance.
The routines (kernels) are written in FORTRAN and draw from a wide variety of scientific
applications including I/O, graphics, and memory management tasks. These routines also
inhabit a large benchmark driver that runs the routines several times, using different input
data each time. The driver checks on the accuracy and timing of the results.
The results of the 24 routines, one for each kernel, are reported in millions of floating-point
operations per second (MFLOPS). Shown in this report are the calculated geometric means.
Figure 20 Livermore Loops Results for Alpha AXP Servers
MFLOPS

30.0
27.8
25.4

25.0

22.3
19.5

20.0

17.4
15.0

10.0

5.0

0.0

DEC 3000/400S AXP DEC 3000/500S AXP DEC 4000/610 AXP

March 1993

DEC 7000/610 AXP DEC 10000/610 AXP

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Fast Fourier Transform and Convolution
Fast Fourier Transform
Fast Fourier Transform (FFT) is the basis for a mathematical technique called Fourier
Analysis that breaks down a function in space or time into sinusoidal components that have
varying frequencies, amplitudes, and phases. A discrete Fourier Transform decomposes a
collection of data into component sine and cosine representations.
Fourier analysis is widely used in all scientific and engineering work including biology,
seismic engineering, electromagnetic analysis, and image processing. The DXML (Digital
eXtended Math Library) Fourier transform subroutines take advantage of certain
computational techniques to reduce the time required to compute the discrete Fourier
Transform.
Convolution
Convolution is a mathematical operation that is used to complement the signal processing
abilities of the Fourier Transform.
This operation allows for modification of signal sequence by having a function or additional
sequence of numbers selectively weight the chosen sequence. This weighting may be used
to obtain properties from a signal source, to selectively enhance the signal source or to
transform a signal source to another domain. The basis of most signal processing resides in
the application of some type of convolution operation.
Table 5 FFT Benchmark Results for Alpha AXP Servers
Metric
DEC OSF/1 AXP
(MFLOPS)
FFT-1D
SP- Complex FFT
Real FFT
DP- Complex FFT
Real FFT

DEC 3000
DEC 4000
DEC 7000
Model 500 AXP Model 610 AXP Model 610 AXP
91.0
78.4
57.5
56.5

106.6
89.8
69.8
72.5

115.8
94.8
77.1
82.6

Digital Equipment Corporation

DEC 10000
Model 610 AXP
127.2
104.4
80.6
84.6

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SLALOM Benchmark
Developed at Ames Laboratory, U.S. Department of Energy, the SLALOM (Scalable
Language-independent Ames Laboratory One-minute Measurement) benchmark solves a
complete, real problem (optical radiosity on the interior of a box). SLALOM is based on
fixed time rather than fixed problem comparison. It measures input, problem setup, solution,
and output, not just the time to calculate the solution.
SLALOM is very scalable and can be used to compare computers as slow as 104 floatingpoint operations per second to computers running a trillion times faster. You can use the
scalability to compare single processors to massively parallel collections of processors, and
you can study the space of problem size versus ensemble size in fine detail.
The SLALOM benchmark is CPU-intensive and measures, in units called patches, the size of
a complex problem solved by the computer in one minute.
Table 6 SLALOM Results
System

Patches

DEC 10000 Model 610 AXP Server

7,248

DEC 7000 Model 710 AXP Server

6,902

DEC 4000 Model 610 AXP Server

6,496

DEC 3000 Model 500S AXP Server

6,084

DEC 3000 Model 400S AXP Server

5,776

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Basic Real-Time Primitives
Measuring basic real-time primitives such as process dispatch latency and interrupt response
latency enhances our understanding of the responsiveness of the DEC OSF/1 AXP
Real-Time kernel.
Process Dispatch Latency is the time it takes the system to recognize an external event and
switch control of the system from a running, lower-priority process to a higher-priority
process that is blocked waiting for notification of the external event.
Interrupt Response Latency (ISR latency) is defined as the amount of elapsed time from
when the kernel receives an interrupt until execution of the first instruction of the interrupt
service routine.
The DEC 3000 Model 500 AXP’s Basic Real-Time Primitives results appear in the
following table.
Table 7 Basic Real-Time Primitives Results

Metric

Minimum
(µsec)

Maximum
(µsec)

Mean
(µsec)

Process Dispatch Latency

62.5

187.7

68.9

Interrupt Response Latency

7.0

50.4

8.4

Test configuration: DEC 3000 Model 500 AXP, 256 MB memory, DEC OSF/1 AXP RT operating system, Ver 1.2.
Test conditions: Single user mode, no network.

Shown next are histograms of the process dispatch latency times and the interrupt response
latency times of a DEC 3000 Model 500 AXP system running in the DEC OSF/1 AXP
Real-Time operating system environment.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP

Process Dispatch Latency
events
pdl12r10.hist
3
1e+05
3
1e+04
3
1e+03
3
1e+02
3
1e+01
3
1e+00
u seconds
60.00

80.00

100.00

120.00

140.00

160.00

180.00

Figure 21 Basic Real-Time Primitives Process Dispatch Latency Results

Interrupt Response Latency
events
iblk12r10.hist
3
1e+06
3
1e+05
3
1e+04
3
1e+03
3
1e+02
3
1e+01
3
1e+00
u seconds
10.00

20.00

30.00

40.00

50.00

Figure 22 Basic Real-Time Primitives Interrupt Response Latency Results

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Rhealstone Benchmark
The Rhealstone Real-time Benchmark is a definition for a synthetic test that measures the
real-time performance of multitasking systems. It is unique in that "the verbal and graphical
specifications, not the C programs, are the essential core of the benchmark" (Kar 1990). We
implemented the following components of the Rhealstone Benchmark, which measure the
critical features of a real-time system:
1. Task switch time—the average time to switch between two active tasks of equal priority.
2. Preemption time—the average time for a high-priority task to preempt a running
low-priority task.
3. Interrupt latency time—the average delay between the CPU’s receipt of an interrupt
request and the execution of the first instruction in an interrupt service routine.
4. Semaphore shuffle time—the delay within the kernel between a task’s request and its
receipt of a semaphore, which is held by another task (excluding the runtime of the
holding task before it relinquishes the semaphore).
Note: We report the delays associated with sem_wait as well as sem_post plus the two
corresponding context switches.

5. Intertask message latency —the delay within the kernel when a non-zero-length data
message is sent from one task to another.
Deadlock-break time is not applicable to DEC OSF/1 AXP Real-Time.
The following figures show the four components of the Rhealstone benchmark we
implemented.

Task number
Task 3

Task 2
Task 1

Time
1

~
~
Task switch time =~
3
1 =
2 =
(Priority of task 1 = task 2 = task 3)

Figure 23 Rhealstone Component–Task-switch Time

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP

Task number
Task 3
(high)
Task 2
(medium)
Task 1
(low)

Time
2

Preemption time =

Figure 24 Rhealstone Component–Preemption Time

Semaphore
Ownership

Task 1

Task 1
Task 2

Task 2

B
Y

Task 1

Time
1

March 1993

3 +

4 = Semaphore shuffle time

Digital Equipment Corporation

B = Blocked
Y = Yields

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP

Figure 25 Rhealstone Component–Semaphore-shuffle Time

= message latency
Task 2
Task 1
Time

Figure 26 Rhealstone Component–Intertask Message Latency
Source: Figures 23, 24, and 26 largely taken from Kar, Rabindra, P., "Implementing the Rhealstone Real-Time Benchmark"
(Dr. Dobb’s Journal, April 1990).

Rhealstone results measured on a DEC 3000 Model 500 AXP are shown in the following
table.

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Table 8 Rhealstone Benchmark Results
Metric

Mean
(µsec)

Task Switch Time

18.1

Preemption Time

33.1

Interrupt Latency Time

8.4 (Range: 7.0–50.4)

Intertask Message Time

92.8

Semaphore Shuffle Time

152.2

Test configuration: DEC 3000 Model 500 AXP, 256 MB memory, DEC OSF/1 AXP RT operating system, Ver 1.2.
Test conditions: Single user mode, no network.

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Perfect Benchmarks Suite
The Perfect (Performance Evaluation for Cost-effective Transformations) Benchmarks
Suite represents an ongoing effort among several universities, research centers, and
supercomputing firms to produce a benchmark package oriented to supercomputers and
parallel processing. Currently, 13 FORTRAN programs totaling over 50,000 lines of code
make up the Perfect Benchmark Suite. These programs include scientific and engineering
applications representing four types of real applications areas: fluid flow, chemical and
physical, engineering design, and signal processing.
Perfect Benchmark Suite’s results are measured in millions of floating-point operations per
second (MFLOPS). Figure 27 shows the geometric mean of the suite’s results in MFLOPS.

Figure 27 Perfect Results for Alpha AXP Servers

MFLOPS

35.0
30.0

29.2

26.4

25.0
20.0

23.1

20.7
18.4

15.0
10.0
5.0
0.0
DEC 3000/400S AXP

DEC 3000/500S AXP

DEC 4000/610 AXP

DEC 7000/610 AXP

Digital Equipment Corporation

DEC 10000/610 AXP

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
Information about Performance
The performance of the Alpha AXP server family was evaluated using industry-standard
benchmarks. These benchmarks allow comparisons across vendors’ systems. Performance
characterization is one "data point" to be used in conjunction with other purchase criteria
such as features, service, and price.
We chose the competitive systems (shown in the preceding charts and tables) based on
comparable or close CPU performance, coupled with comparable expandability capacity
(primarily memory and disk). Although we do not present price comparisons in this report,
system price was a secondary factor in our competitive choices.
Note: The performance information in this report is for guidance only. System performance is highly
dependent upon application characteristics. Individual work environments must be carefully evaluated and
understood before making estimates of expected performance.

This report simply presents the data, based on specified benchmarks. Competitive
information is based on the most current published data for those particular systems and has
not been independently verified (except as noted). The Alpha AXP performance information
presented in this brief is the latest measured results as of the date published. Digital has an
ongoing program of performance engineering across all products. As system tuning and
software optimizations continue, Digital expects the performance of its servers to increase.
As more benchmark results become available, Digital will publish reports containing the new
and updated benchmark data.
For more information on Digital’s Alpha AXP server family, please contact your local
Digital sales representative.

March 1993

Digital Equipment Corporation

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
References
System and Vendor

Sources

DEC 3000 Model 400S and 500S AXP Servers
DEC 4000 Models 610 AXP Servers
DEC 7000 Models 610 AXP Servers
DEC 10000 Models 610 AXP Server

All benchmarking performed by Digital Equipment Corporation.
All benchmarking performed by Digital Equipment Corporation.
All benchmarking performed by Digital Equipment Corporation.
All benchmarking performed by Digital Equipment Corporation.

Auspex NS/5500

SPEC results reported by Auspex Systems (3/93).

HP 9000 Models 715/50, 735, and 755

SPEC, LINPACK, and Dhrystone benchmark results reported by
Hewlett-Packard in "HP Apollo 9000 Series 700 Workstation Systems"
(11/92). 735 LINPACK results from Dongarra, J., "Performance
of Various Computers Using Standard Linear Equations Software"
(3/6/93).
AIM results reported in UNIX System Price Performance Guide, AIM
Technology (Winter 1992-1993).
DN&R Labs CPU2 results reported in Workstation
Laboratories, Volume 19, Chapters 20 & 21 (1/93).
SPEC benchmark results reported in SPEC Newsletter (12/92).
Dhrystone and LINPACK 100x100 benchmark results reported in "HP
Apollo 9000 Series 700 Performance Brief" (1/92).
LINPACK 1000x1000 benchmark results reported in Dongarra, J.,
"Performance of Various Computers Using Standard Linear Equations
Software" (9/28/92).
DN&R Labs CPU2 results reported in Workstation Laboratories, Inc.,
Volume 16, Chapter 12 (3/1/92).
SPEC benchmark results reported by HP (11/10/92).
Dhrystone and LINPACK benchmark results reported in "HP Apollo 9000
Series 700 Performance Brief" (1/92).
AIM results reported in UNIX System Price Performance Guide, AIM
Technology (Winter 1992-1993).
SPEC benchmark results reported in SPEC Newsletter (3/92).
Dhrystone and LINPACK benchmark results reported in"HP
9000 Series 800 Business Server System Performance" (1992).
AIM results reported in UNIX System Price Performance Guide, AIM
Technology (Winter 1992-1993).
SPEC benchmark results reported in SPEC Newsletter (9/92).
SPEC results reported by HP (3/93).

HP 9000 Model 730

HP 9000 Model 750

HP 9000 Models 827S, 847S, and 867S

HP 9000 Model 887S
HP 9000 Model H50

IBM RS 6000 Models 360 and 370
IBM RS 6000 Model 560
IBM RS 6000 Models 580, 970B, and 980B

SPEC and LINPACK 100x100 results reported by IBM (2/2/93).
SPEC results reported by IBM (3/93).
SPEC benchmark results reported in SPEC Newsletter (9/92) and by IBM
(9/92).
AIM results reported in UNIX System Price Performance Guide, AIM
Technology (Winter 1992-1993).
LINPACK 100x100 benchmark results reported by IBM (9/92) . LINPACK
1000x1000 results reported in Dongarra, J., "Performance of Various
Computers Using Standard Linear Equations Software" (3/6/93).

Digital Equipment Corporation

March 1993

Alpha AXP Server Family
Performance Brief - DEC OSF/1 AXP
SUN SPARC 10 Model 41

SPECmark, LINPACK 100x100 dp, and DN&R Labs CPU2 results reported
in Workstation Laboratories, Volume 19, Chapter 19 (1/93).
AIM results reported in UNIX System Price Performance Guide, AIM
Technology (Fall 1992).
SPEC92 benchmark results reported in SPEC Newsletter (9/92).
LINPACK and Dhrystone benchmark results reported by SUN (11/10/92).

Article
Kar, Rabindra P. "Implementing the Rhealstone Real-Time Benchmark" (Dr. Dobb’s Journal, April 199).

March 1993

Digital Equipment Corporation