"Computer systems technology evolves so rapidly that we must provide new benchmark suites every two to three years to ensure a level playing field," says Kaivalya M. Dixit, SPEC president. "SPEC92 was a great success, but it is time to make the transition to standardized benchmarks that reflect the advances in chip technologies, compilers and applications that have taken place over the last three years; those benchmarks constitute SPEC95."
SPEC95 comprises two sets (or suites) of benchmarks: CINT95 for compute-intensive integer performance and CFP95 for compute-intensive floating point performance. The two suites provide component-level benchmarks that measure the performance of the computer's processor, memory architecture and compiler. SPEC benchmarks are selected from existing application and benchmark source code running across multiple platforms. Each benchmark is tested on different platforms to obtain fair performance results across competing hardware and software systems.
SPEC95 is the third major version of the SPEC benchmark suites, which in 1989 became the first widely accepted standard for comparing computeintensive performance across various architectures. The new release replaces SPEC92, which will be gradually phased out between now and June 1996, when SPEC will stop publishing SPEC92 results and stop selling the benchmark suite. Performance results from SPEC95 cannot be compared to those from SPEC92, since new benchmarks have been added and existing ones changed.
"Compiler writers have learned to optimize performance for individual SPEC92 benchmarks," says Jeff Reilly, SPEC95 release manager. "The best way to avoid these benchmarkspecific optimizations is to develop new benchmark suites." SPEC95 builds on the lessons learned from the SPEC89 and SPEC92 suites, according to Reilly. The new benchmarks were analyzed to ensure that they are as resistant as possible to compiler optimizations that might not translate into realworld performance gains. Improvements to the suites include longer run times and larger problems for benchmarks, more application diversity, greater ease of use, and standard development platforms that will allow SPEC to produce additional releases for other operating systems.
A Sun SPARCstation 10/40 with 128 MB of memory was selected as the SPEC95 reference machine and Sun SC3.0.1 compilers were used to obtain reference timings on the new benchmarks. By definition, the SPECint95 and SPECfp95 numbers for the Sun SPARCstation 10/40 are both "1."
SPEC95 rules permit both baseline and optimized results for CINT95 and CFP95 suites. The baseline rules restrict the number of optimizations that can be used for performance testing. In general, SPEC95 rules are more restrictive in regard to optimizations than the SPEC92 rules. Baseline metrics are mandatory for all reported results; reporting of optimized results is optional.
SPEC95 also allows performance to be measured for both speed d throughput (rate). Speed metrics such as SPECint95 measure how fast a computer completes a single task. Rate metrics such as SPECint_rate95 measure how many tasks a computer can accomplish in a certain amount of time. SPEC95 measures rate performance for single processors, symmetric multi processor systems and cluster systems.
The CINT95 suite, written in C language, contains eight CPUintensive integer benchmarks. It is used to measure and calculate the following metrics:
SPECint95 -- The geometric mean of eight normalized ratios (one for each integer benchmark) when compiled with aggressive optimization for each benchmark.
SPECint_base95 -- The geometric mean of eight normalized ratios when compiled with conservative optimization for each benchmark.
SPECint_rate95 -- The geometric mean of eight normalized throughput ratios when compiled with aggressive optimization for each benchmark.
SPECint_rate_base95 -- The geometric mean of eight normalized throughput ratios when compiled with conservative optimization for each benchmark.
The CFP95 suite, written in FORTRAN language, contains 10 CPU-intensive floating point benchmarks. It is used to measure and calculate the following metrics:
SPECfp95 -- The geometric mean of 10 normalized ratios (one for each floating point benchmark) when compiled with aggressive optimization for each benchmark.
SPECfp_base95 -- The geometric mean of 10 normalized ratios when compiled with conservative optimization for each benchmark.
SPECfp_rate95 -- The geometric mean of 10 normalized throughput ratios when compiled with aggressive optimization for each benchmark.
SPECfp_rate_base95 -- The geometric mean of 10 normalized throughput ratios when compiled with conservative optimization for each benchmark.
Initial results for systems from six vendor companies are included with this release. Additional results will be reported in the next issue of the SPEC Newsletter, scheduled for publication at the end of September. SPEC members are being encouraged to report SPEC95 results on older platforms to provide an historical perspective for the new results.
[Poster's note: I know that the comp.benchmarks readership will probably be interested in a table of these 37 initial results, which were included in the press kit, and will probably appear in press reports soon. However, I don't have a complete table in electronically transferable form. It will probably be posted as soon as it is available. - RW]
SPEC95 (CINT95 and CFP95) is available on CD-ROM from SPEC's administrator, the National Computer Graphics Association (NCGA). The cost is $600 for new customers, $300 for new university customers, $300 for current SPEC92 licensees and $150 for current university licensees. SPEC is a non-profit corporation formed to establish, maintain and endorse a standardized set of relevant benchmarks that can be applied to the newest generation of high-performance computers. Included in its membership are the Open Systems Group (OSG); OSG Associates, consisting of leading universities and research facilities; the HighPerformance Group (HPG); and HPG Associates.
For more information, contact
SPEC, c/o NCGA,
2722 Merrilee Drive, Ste. 200,
Fairfax, VA 22031;
tel: 703-6989604, ext. 325;
Cramblitt & Company
Q2: What is a benchmark?
A2: The definition from Webster's II Dictionary states: "A standard of measurement or evaluation." SPEC is a non profit corporation formed to establish and maintain computer benchmarks for measuring component and system level computer performance.
Q3: What does the "C" in CINT95 and CFP95 stand for?
A3: In its product line, SPEC uses "C" to denote a "component-level" benchmark and "S" to denote a "system level" benchmark. CINT95 and CFP95 are component-level benchmarks.
Q4: What components do CINT95 and CFP95 measure?
A4: Being compute-intensive benchmarks, they emphasize the performance of the computer's processor, the memory architecture and the compiler. It is important to remember the contribution of the latter two components; performance is more than just the processor.
Q5: What component performance is not measured by CINT95 and CFP95?
A5: The CINT95 and CFP95 benchmarks do not stress other computer components such as I/O (disk drives), networking or graphics. It might be possible to configure a system in such a way that one or more of these components impact the performance of CINT95 and CFP95, but that is not the intent of the suites.
Q6: What is included in the SPEC95 package?
A6: SPEC provides the following in its SPEC95 package: SPEC95 tools for compiling, running and validating the benchmarks, compiled for a variety of operating systems; source code for the SPEC95 tools, to allow the tools to be built for systems not covered by the precompiled tools; source code for the benchmarks; tools for generating performance reports; run and reporting rules defining how the benchmarks should be used to produce standard results; and SPEC95 documentation.
The initial offering of SPEC95 will have tools for most UNIX operating systems. Additional products for other operating systems (Windows NT, OpenVMS, etc.) will be released as later products if SPEC detects enough demand. All of this will be shipped on a single CD-ROM disk.
Q7: What does the SPEC95 user have to provide?
A7: The user must have a computer system running a UNIX environment with a compiler installed and a CD-ROM drive. Approximately 150 MB will be needed on a hard drive to install SPEC95. It is also assumed that the system has at least 64 MB of RAM to ensure that the benchmarks remain compute-intensive (SPEC is assuming this will be the standard amount of desktop memory during the life of this suite).
Q8: What are the basic steps in running the benchmarks?
A8: Installation and use are covered in detail in the SPEC95 User Documentation. The basic steps are as follows: Install SPEC95 from media. Run the installation scripts specifying your operating system. Compile the tools, if executables are not provided in SPEC95. Determine what metric you wish to run. Create a configuration file for that metric. In this file, you specify compiler flags and other system dependent information. Run the SPEC tools to build (compile), run and validate the benchmarks. If the above steps are successful, generate a report based on the run times and metric equations.
Q9: What source code is provided? What exactly makes up these suites?
A9: CINT95 and CFP95 are based on compute- intensive applications provided as source code. CINT95 contains eight applications written in C that are used as benchmarks:
099.go -- Artificial intelligence; plays the game of "Go"
124.m88ksim -- Moto 88K chip simulator; runs test program
126.gcc -- New version of GCC; builds SPARC code
129.compress -- Compresses and decompresses file in memory
130.li -- LISP interpreter
132.ijpeg -- Graphic compression and decompression
134.perl -- Manipulates strings (anagrams) and prime numbers in Perl
147.vortex -- A database program
CFP95 contains 10 applications written in FORTRAN that are used as benchmarks:
101.tomcatv -- A mesh-generation program
102.swim -- Shallow water model with 1024 x 1024 grid
103.su2cor -- Quantum physics; Monte Carlo simulation
104.hydro2d -- Astrophysics; Hydrodynamical Navier Stokes equations
107.mgrid -- Multi-grid solver in 3D potential field
110.applu -- Parabolic/elliptic partial differential equations
125.turb3d -- Simulates isotropic,homogeneous turbulence in a cube
141.apsi -- Solves problems regarding temperature, wind, velocity and distribution of pollutants
145.fpppp -- Quantum chemistry
146.wave5 -- Plasma physics; electromagnetic particle simulation
Q10: What metrics can be measured?
A10: The CINT95 and CFP95 suites can be used to measure and calculate the following metrics:
SPECint95: The geometric mean of eight normalized ratios (one for each integer benchmark) when compiled with aggressive optimization for each benchmark.
SPECint_base95: The geometric mean of eight normalized ratios when compiled with conservative optimization for each benchmark.
SPECint_rate95: The geometric mean of eight normalized throughput ratios when compiled with aggressive optimization for each benchmark.
SPECint_rate_base95: The geometric mean of eight normalized throughput ratios when compiled with conservative optimization for each benchmark.
SPECfp95: The geometric mean of ten normalized ratios (one for each floating point benchmark) when compiled with aggressive optimization for each benchmark.
SPECfp_base95: The geometric mean of ten normalized ratios when compiled with conservative optimization for each benchmark.
SPECfp_rate95: The geometric mean of ten normalized throughput ratios when compiled with aggressive optimization for each benchmark.
SPECfp_rate_base95: The geometric mean of ten normalized throughput ratios when compiled with conservative optimization for each benchmark.
The ratio for each of the benchmarks is calculated using a SPEC-determined reference time and the run time of the benchmark.
Q11: What is the difference between a "base" metric and a "non-base"
A11: In order to provide comparisons across different computer hardware, SPEC had to provide the benchmarks as source code. Thus, in order to run the benchmarks, they must be compiled. There was agreement that the benchmarks should be compiled the way users compile programs. But how do users compile programs? On one side, people might experiment with many different compilers and compiler flags to achieve the best performance. On the other side, people might just compile with the basic options suggested by the compiler vendor. SPEC recognizes that it cannot exactly match how everyone uses compilers, but two reference points are possible. The base metrics (e.g., SPECint_base95) are required for all reported results and have set guidelines for compilation (e.g., the same flags must be used in the same order for all benchmarks). The non-base metrics (e.g., SPECint95) are optional and have less strict requirements (e.g., different compiler options may be used on each benchmark).
A full description of the distinctions can be found in the SPEC95 Run and Reporting Rules available with SPEC95.
Q12: What is the difference between a "rate" and a "nonrate" metric?
A12: There are several different ways to measure computer performance. One way is to measure how fast the computer completes a single task; this is a speed measure. Another way is to measure how many tasks a computer can accomplish in a certain amount of time; this is called a throughput, capacity or rate measure.
The SPEC speed metrics (e.g., SPECint95) are used for comparing the ability of a computer to complete single tasks. The SPEC rate metrics (e.g., SPECint_rate95) measure the throughput or rate of a machine carrying out a number of tasks.
Q13: Why and/or when should I use SPEC95?
A13: Typically, the best measure of a system is your own application with your own workload. Unfortunately, it is often very difficult and expensive to get a wide base of reliable, repeatable and comparable measurements on different systems for your own application with your own workload. This might be due to time, money or other constraints.
Benchmarks exist to act as a reference point for comparison. It's the same reason that EPA gas mileage exists, although probably no driver in America gets exactly the EPA gas mileage. If you understand what benchmarks measure, they're useful. It's important to know that CINT95 and CFP95 are CPU-focused and not system-focused benchmarks. These CPU benchmarks focus on only one portion of those factors that contribute to applications performance. A graphics or network performance bottleneck within an application, for example, will not be reflected in these benchmarks. Understanding your own needs helps determine the relevance of the benchmarks.
Q14: Which SPEC95 metric should be used to determine performance?
A14: It depends on your needs. SPEC provides the benchmarks and results as tools for you to use. You need to determine how you use a computer or what your performance requirements are and then choose the appropriate SPEC benchmark or metrics.
A single user running a compute-intensive integer program, for example, might only be interested in SPECint95 or SPECint_base95. On the other hand, a person who maintains a machine used by multiple scientists running floating point simulations might be more concerned with SPECfp_rate95 or SPECfp_rate_base95.
Q15: SPEC92 is already an available product. Why create SPEC95
and will it show anything different from SPEC92?
A15: Technology is always improving. As the technology improves, the benchmarks need to improve as well. SPEC needed to address the following issues:
Run-time -- Several of the SPEC92 benchmarks were running in less than a minute on leading-edge processors/systems. Given the SPEC measurement tools, small changes or fluctuations in the measurements were having significant impacts on the percentage improvements being seen. SPEC chose to make the SPEC95 benchmarks longer to take into account future performance and prevent this from being an issue for the life of the suite.
Application size -- Many comments received by SPEC indicated that applications had grown in complexity and size and that SPEC92 was becoming less representative of what runs on current systems. For SPEC95, SPEC selected programs with larger resource requirements to provide a mix with some of the smaller programs.
Application type -- SPEC felt that there were additional application areas that should be included in SPEC95 to increase variety and representation within the suites. Areas such as imaging and database have been added. Portability - SPEC found that computeintensive performance was important beyond the UNIX workstation arena where SPEC was founded. It was important, therefore, that the benchmarks and the tools running the benchmarks be as independent of the operating system as possible. While the first release of SPEC95 will be geared toward UNIX, SPEC has consciously chosen programs and tools that are dependent only upon POSIX or ANSI standard development environments. SPEC will produce additional releases for other operating systems (such as Microsoft Windows/NT) based on demand.
Moving target -- The initial hope for benchmarks is that improvements in the benchmark performance will be generally applicable to other situations. As competition develops, however, improvements in the test performance can become specific to that test only. SPEC95 provides updated benchmarks so that general improvements will be encouraged and testspecific optimizations become less effective.
Education -- As the computer industry grows, benchmark results are being quoted more often. With the release of new benchmark suites, SPEC has a fresh opportunity to discuss and clarify how and why the suite was developed.
Q15: What happens to SPEC92 after SPEC95 is released?
A15: SPEC will begin the process of making SPEC92 obsolete. The results published by SPEC will be marked as obsolete and by June 1996, SPEC will stop publishing SPEC92 results and stop selling the SPEC92 suites.
Q16: Is there a way to translate SPEC92 results to SPEC95 results
or vice versa?
A16: There is no formula for converting from SPEC92 results to SPEC95 results; they are different products. There might be a high correlation between SPEC92 and SPEC95 results (i.e., machines with higher SPEC92 results might have higher SPEC95 results), but there is no universal formula for all systems.
SPEC is strongly encouraging SPEC licensees to publish SPEC95 numbers on older platforms to provide a historical perspective.
Q17: What criteria was used to select the benchmarks?
A17: In the process of selecting applications to use as benchmarks, SPEC considered the following criteria: portability to all SPEC hardware architectures (32 and 64 bit including Alpha, Intel Architecture, PA RISC, Rxx00, SPARC, etc.); portability to various operating systems, particularly UNIX, NT and OpenVMS; benchmarks should not include measurable I/O; benchmarks should not include networking or graphics; benchmarks should run in 64-MB RAM without swapping (SPEC is assuming this will be a minimal memory requirement for the life of SPEC95 and the emphasis is on computeintensive performance and not disk activity); benchmarks should run at least five minutes on a Digital Equipment Corp. 200-MHz Alpha system; and no more than five percent of benchmarking time should be spent processing code not provided by SPEC.
Q18: Weren't some of the SPEC95 benchmarks in SPEC92?
How are they different?
A18: Although some of the benchmarks from SPEC92 are included in SPEC95, they all have been given different workloads or modified to improve their coding style or use of resources. The revised benchmarks have been assigned different identifying numbers to distinguish them from versions in previous suites and to indicate they are not comparable with their predecessors.
Q19: Why were some of the benchmarks not carried over from SPEC92?
A19: Some benchmarks were not carried over because it was not possible to create a longer running workload or to create a more robust workload, or the benchmarks were too susceptible to benchmark- specific compiler optimization.
Q20: Why does SPEC use a reference machine for determining performance
metrics? What machine is used for SPEC95 benchmark suites?
A20: SPEC uses a reference machine to normalize the performance metrics used in the SPEC95 suites. Each benchmark is run and measured on this machine to establish a reference time for that benchmark. These times are then used in the SPEC calculations. SPEC uses the SPARCstation 10/40 (40-MHz SuperSPARC with no L2 cache) as the reference machine. It takes approximately 48 hours to run a SPECconforming execution of CINT95 and CFP95 on this machine.
Q21: How long does it take to run the SPEC95 benchmark suites?
A21: This depends on the suite and the machine that is running the benchmarks. As mentioned above, on the reference machine it takes two days for a SPEC conforming run (at least three iterations of each benchmark to ensure that results can be reproduced).
Q22: What if the tools cannot be run or built on a system?
Can they be run manually?
A22: To generate SPEC-compliant results, the tools used must be approved by SPEC. If several attempts at using the SPEC tools are not successful for the operating system for which you purchased SPEC95, you should contact SPEC for technical support. SPEC will work with you to correct the problem and/or investigate SPECcompliant alternatives.
Q23: What if I don't want to run the benchmarks? Is there any place
where results are available?
A23: Here are the alternatives: Every quarter, SPEC publishes the SPEC Newsletter, which contains results submitted to SPEC by SPEC members and licensees. Subscription information is available from SPEC. SPEC provides information to the Performance Database Server found at: http://performance.netlib.org/performance/html/spec.html This typically lags three months behind the SPEC Newsletter. SPEC is working on establishing its own Internet presence, although details are not yet available.
Q24: How do I contact SPEC?
A24: Here is the contact information for SPEC:
2722 Merrilee Drive, Ste. 200
Fairfax, VA 22031
Tel: 703-698-9604, ext. 325
Questions and answers were prepared by Kaivalya Dixit of IBM and Jeff Reilly of Intel Corp. Dixit is president of SPEC and Reilly is release manager for SPEC95.