ABSTRACT
This document provides guidelines required to build,
run, and report on the SPECaccel benchmarks.
1. Philosophy
1.1 A SPECaccel Result Is An Observation
1.2 A Published SPECaccel Result Is a Declaration of Expected Performance
1.3 A SPECaccel Result is a Claim About Maturity of Performance Methods
1.4 Peak and Base Builds and Runs
1.5 Estimates
1.6 About SPEC
1.7 Compliance and Compatibility
1.8 Usage of the Philosophy Section
1.9 Name of the Suite
2. Building the SPECaccel Suite
2.1 Build Procedures
2.2 General Optimization Rules
2.3 Base Optimization Rules
2.4 Peak Optimization Rules
3. Running the SPECaccel Suite
3.1 System Configuration
3.2 Controlling Benchmark Jobs
3.3 Run-time Environment
3.4 Continuous Run Requirement
3.5 Base, peak, and basepeak
3.6 Run-Time Dynamic Optimization
4. Results Disclosure
4.1 Rules regarding availability dates and systems not yet shipped
4.2 Configuration Disclosure
4.3 Test Results Disclosure
4.4 Required Disclosures
4.5 Research and Academic usage of the SPECaccel Suite
4.6 Fair Use
5. Run Rule Exceptions
(Click on an item number above, to go to the detailed contents about that item.)
|
1. Philosophy 1.1. A SPECaccel Result Is An Observation 1.1.1. Test Methods 1.1.2. Conditions of Observation 1.1.3. Assumptions About the Tester 1.1.4. Intentions of the benchmark suite 1.2. A Published SPECaccel Result Is a Declaration of Expected Performance 1.2.1. Reproducibility 1.2.2. Obtaining Components 1.2.2.1 Hardware, Operating System and Compilers 1.3. A SPECaccel Result is a Claim About Maturity of Performance Methods 1.4. Peak and Base Builds and Runs 1.5. Estimates 1.6. About SPEC 1.6.1. Publication on SPEC's web site is encouraged 1.6.2. Publication on SPEC's web site is not required 1.6.3. SPEC May Require New Tests 1.6.4. SPEC May Adapt the Suite 1.7. Compliance and Compatibility 1.7.1. Target Languages 1.7.2. Supported Operating Systems 1.7.3. API Specifications 1.8. Usage of the Philosophy Section 1.9. Name of the Suite 2. Building the SPECaccel Suite 2.1. Build Procedures 2.1.1 SPEC's tools must be used 2.1.2 The runaccel build environment 2.1.3 Continuous Build requirement 2.1.4 Changes to the runaccel build environment 2.1.5 Cross-compilation allowed 2.1.6 Individual builds allowed 2.1.7 Tester's assertion of equivalence between build types 2.2 General Optimization Rules 2.2.1 Must not use names 2.2.2 Limitations on library substitutions 2.2.3 Limitations on size changes 2.2.4 Portability Flags 2.2.5 Source Code Changes 2.3 Base Optimization Rules 2.3.1 Safety and Standards Conformance 2.3.2 Same for all benchmarks of a given language 2.3.3 Assertion flags must NOT be used in base 2.3.4 Floating point reordering allowed 2.3.5 Base build environment 2.3.6 Portability Switches for Data Models 2.3.7 Cross-module optimization 2.3.8 Alignment switches are allowed 2.4 Peak Optimization Rules 2.4.1 Individual Benchmark Tuning 2.4.2 Peak build environment 2.4.3 Assertion flags may be used in peak 2.4.4 Parallel Models and Methods 2.4.5 Directive Modifications |
3. Running the SPECaccel Suite 3.1. System Configuration 3.1.1 Operating System State 3.1.2 File Systems and File Servers 3.2. Controlling Benchmark Jobs 3.2.1 Number of runs in a reportable result 3.2.2 Device selection 3.2.3 The submit directive 3.3. Run-time Environment 3.4. Continuous Run Requirement 3.5. Base, Peak, and Basepeak 3.6. Run-Time Dynamic Optimization 3.6.1 Definitions and Background 3.6.2 RDO Is Allowed, Subject to Certain Conditions 3.6.3 RDO Disclosure and Resources 3.6.4 RDO Settings Cannot Be Changed At Run-time 3.6.5 RDO and safety in base 3.6.6 RDO carry-over by program is not allowed 4. Results Disclosure 4.1. Rules regarding availability dates and systems not yet shipped 4.1.1 Pre-production software can be used 4.1.2 Software Component Names 4.1.3 Specifying Dates 4.1.4 If dates are not met 4.1.5 Performance changes for pre-production systems 4.2. Configuration Disclosure 4.2.1 Identification of System, Manufacturer and Tester 4.2.1.1 Identification of Equivalent Systems 4.2.2 Host Configuration 4.2.3 Device Configuration 4.2.4 Software Configuration 4.2.5 Tuning Configuration 4.2.6 Description of Portability and Tuning Options ("Flags File") 4.2.7 Configuration Disclosure for User Built Systems 4.3. Test Results Disclosure 4.3.1 SPECaccel Performance Metrics 4.3.2 SPECaccel Energy Metrics 4.3.3 Metric Selection 4.3.4 Estimates are allowed 4.3.5 Performance changes for production systems 4.4. Required Disclosures 4.5. Research and Academic usage of the SPECaccel Suite 4.6. Fair Use 5. Run Rule Exceptions |
This section is an overview of the purpose, definitions, methods, and assumptions for the SPECaccel run rules. The purpose of the SPEC ACCEL benchmark and its run rules is to further the cause of fair and objective benchmarking of high-performance computing systems. The rules help ensure that published results are meaningful, comparable to other results, and reproducible. SPEC believes that the user community benefits from an objective series of tests which serve as a common reference.
Per the SPEC license agreement, all SPECaccel results disclosed in public -- whether in writing or in verbal form -- must adhere to the SPECaccel Run and Reporting Rules, or be clearly described as estimates.
A published SPECaccel result means three things:
A published SPECaccel result is an empirical report of performance observed when carrying out certain computation- and communication-intensive tasks. The result is accomplished on a specific Accelerator device for the entire benchmark run.
SPEC supplies the SPECaccel benchmarks in the form of source code, which
testers are not allowed to modify except under certain very restricted
circumstances. The SPECaccel
benchmarks are work-based benchmarks. The amount of work is fixed,
and the time used to perform the work is measured.
The tester supplies compilers, and the System Under Test (SUT). In addition, the tester provides a config file, where appropriate optimization flags are set, as well as, where they are needed, portability flags. SPEC provides example config files in the config subtree as well as documentation on how to create a config file in Docs/config.html. SPEC supplies tools which automatically:
The report that certain performance has been observed is meaningful only if the conditions of observation are stated. SPEC therefore requires that a published result include a description of all performance-relevant conditions.
It is assumed that the tester:
The person who actually carries out the test is, therefore, the first and the most important audience for these run rules. The rules attempt to help the tester by trying to be clear about what is and what is not allowed.
This benchmark suite is intended to measure the performance of hardware Accelerator devices and their supporting software tool chains using computationally-intensive parallel applications. The suite is comprised of scientific applications used in High Performance Computing (HPC). The suite has been ported using several Accelerator programming models and methods, with the user selecting the most appropriate for their test.
A published SPECaccel result is a declaration that the observed level of performance can be obtained by others. Such declarations are widely used by vendors in their marketing literature, and are expected to be meaningful to ordinary customers.
It is expected that later testers can obtain a copy of the SPECaccel suite, obtain the components described in the original result, and reproduce the claimed performance within a small range to allow for run-to-run variation. See section 4.3.5 for specific requirements.
Therefore, it is expected that the components used in a published result can in fact be obtained, with the level of quality commonly expected for products sold to ordinary customers. Different components are subject to different standards, described below:
Subcomponents are required to:
The judgment of whether a component meets the above list may sometimes pose difficulty, and various references are given in these rules to provide guidelines for such judgment. But by way of introduction, imagine a vendor-internal version of a compiler, designated only by an internal code name, unavailable to customers, which frequently generates incorrect code. Such a compiler would fail to provide a suitable environment for general programming, and would not be ready for use in a SPECaccel result.
A published SPECaccel result carries an implicit claim that the performance methods it employs are more than just "prototype" or "experimental" or "research" methods; it is a claim that there is a certain level of maturity and general applicability in its methods. Unless clearly described as an estimate, a published SPEC result is a claim that the performance methods employed (whether hardware or software, compiler, or other):
SPEC is aware of the importance of optimizations in producing the best performance. SPEC is also aware that it is sometimes hard to draw an exact line between legitimate optimizations that happen to benefit SPEC benchmarks, versus optimizations that exclusively target the SPEC benchmarks. However, with the list above, SPEC wants to increase awareness of implementers and end users to issues of unwanted benchmark-specific optimizations that would be incompatible with SPEC's goal of fair benchmarking.
The tester must describe the performance methods that are used in terms that a performance-aware user can follow, so that users can understand how the performance was obtained and can determine whether the methods may be applicable to their own applications. The tester must be able to make a credible public claim that a class of applications in the real world may benefit from these methods.
"Peak" metrics may be produced by building each benchmark in the suite with a set of options individually selected for that benchmark, and running them with environment settings individually selected for that benchmark. The optimizations selected must adhere to the set of general benchmark optimization rules described in section 2.2 below. This may also be referred to as "aggressive compilation".
"Base" metrics must be produced by building all the benchmarks in the suite with a common set of options, and running them with environment settings common to all the benchmarks in the suite. In addition to the general benchmark optimization rules (section 2.2), base optimizations must adhere to a stricter set of rules described in section 2.3.
These additional rules serve to form a "baseline" of performance that can be obtained with a single set of compiler switches, single-pass make process, and a high degree of portability, safety, and performance.
Rules for building the benchmarks are described in section 2.
SPECaccel metrics may be estimated. All estimates must be clearly designated as such.
This philosophy section has described how a "result" has certain characteristics: e.g. a result is an empirical report of performance, includes a full disclosure of performance-relevant conditions, can be reproduced, uses mature performance methods. By contrast, estimates may fail to provide one or even all of these characteristics.
Nevertheless, estimates have long been seen as valuable for SPEC benchmarks. Estimates are set at inception of a new chip design and are tracked carefully through analytic, simulation, and HDL (Hardware Description Language) models. They are validated against prototype hardware and, eventually, production hardware. With chip designs taking years, and requiring very large investments, estimates are central to corporate roadmaps. Such roadmaps may compare SPECaccel estimates for several generations of processors, and, explicitly or by implication, contrast one company's products and plans with another's.
SPEC wants the accelerator benchmarks to be useful, and part of that usefulness is allowing the metrics to be estimated.
The key philosophical point is simply that estimates must be clearly distinguished from results.
SPEC encourages the review of SPECaccel results by the relevant subcommittee, and subsequent publication on SPEC's web site (http://www.spec.org/accel2023). SPEC uses a peer-review process prior to publication, in order to improve consistency in the understanding, application, and interpretation of these run rules.
Review by SPEC is not required. Testers may publish rule-compliant results independently. No matter where published, all results publicly disclosed must adhere to the SPEC Run and Reporting Rules, or be clearly marked as estimates. SPEC may take action if the rules are not followed.
In cases where it appears that the run rules have not been followed, SPEC may investigate such a claim and require that a result be regenerated, or may require that the tester correct the deficiency (e.g. make the optimization more general purpose or correct problems with code generation).
The SPEC High Performance Group reserves the right to adapt the SPECaccel suite as it deems necessary to preserve its goal of fair benchmarking. Such adaptations might include (but are not limited to) removing benchmarks, modifying codes or workloads, adapting metrics, republishing old results adapted to a new metric, or requiring retesting by the original tester.
The SPECaccel benchmarks are written in Fortran 2003 and C99. If benchmarks fail due to non-compliance with the appropriate Language Standard, the SPEC HPG Committee will be inclined to accept performance-neutral source-code changes.
The SPECaccel benchmarks have been tested on Linux/UNIX
systems. It is not our intent to exclude them from
working on other platforms. The burden of porting the benchmarks and
tools to other operating systems may fall on you, however, if
you decide to submit results. Section 5
provides for exceptional cases where the standard run- and
submission- rules cannot be followed.
The SPECaccel suite is written to comply with OpenACC 3.0 and OpenMP 5.2. If benchmarks fail due to non-compliance with the API specifications, the SPEC HPG Committee will be inclined to accept performance-neutral source-code changes. In cases where the implementation of the API is non-compliant or imposes some fundamental limitation, the SPEC HPG Committee is inclined to advocate fixing the implementation of the API rather than accept changes to the benchmark source.
This philosophy section is intended to introduce concepts of fair benchmarking. It is understood that in some cases, this section uses terms that may require judgment, or which may lack specificity. For more specific requirements, please see the sections below.
In case of a conflict between this philosophy section and a run rule in one of the sections below, normally the run rule found below takes priority.
Nevertheless, there are several conditions under which questions should be resolved by reference to this section: (a) self-conflict: if rules below are found to impose incompatible requirements; (b) ambiguity: if they are unclear or silent with respect to a question that affects how a result is obtained, published, or interpreted; (c) obsolescence: if the rules below are made obsolete by changing technical circumstances or by directives from superior entities within SPEC.
When questions arise as to interpretation of the run rules:
The registered mark is SPECaccel, so the benchmark should always be referred to as such, and not just as ACCEL. It must also be used only as an adjective, as in "the SPECaccel 2023 benchmark suite", never just as "SPECaccel".
SPEC has adopted a set of rules defining how the SPECaccel benchmark suite must be built and run to produce peak and base metrics.
With the release of the SPECaccel suite, a set of tools based on GNU Make and Perl5 are supplied to build and run the benchmarks. To produce publication-quality results, these SPEC tools must be used. This helps ensure reproducibility of results by requiring that all individual benchmarks in the suite are run in the same way and that a configuration file is available that defines the optimizations used.
The primary tool is called runaccel. It is described in the runaccel documentation in the Docs subdirectory of the SPEC root directory.
SPEC supplies pre-compiled versions of the tools for a variety of platforms. If a new platform is used, please see tools-build[.html] in the Docs directories for information on how to build the tools, and how to obtain approval for them. SPEC's approval is required for the tools build, so a log must be generated during the build.
For more complex ways of compilation, SPEC has provided hooks in the tools so that such compilation and execution is possible (see the tools documentation for details). If, for some reason, building and running with the SPEC tools does not work for your environment, the test sponsor may ask for permission to use performance-neutral alternatives (see section 5).
When runaccel is used to build the SPECaccel benchmarks, it must be used in generally available, documented, and supported environments (see section 1), and any aspects of the environment that contribute to performance must be disclosed to SPEC (see section 4).
On occasion, it may be possible to improve run time performance by environmental choices at build time. For example, one might install a performance monitor, turn on an operating system feature such as BIGPAGES, or set an environment variable that causes the cc driver to invoke a faster version of the linker.
It is difficult to draw a precise line between environment settings that are reasonable versus settings that are not. Some settings are obviously not relevant to performance (such as hostname), and SPEC makes no attempt to regulate such settings. But for settings that do have a performance effect, for the sake of clarity, SPEC has chosen that:
As described in section 1, it is expected that testers can reproduce other testers' results. In particular, it must be possible for a new tester to compile both the base and peak benchmarks for an entire suite in one execution of runaccel, with appropriate command line arguments and an appropriate configuration file, and obtain executable binaries that are (from a performance point of view) equivalent to the binaries used by the original tester.
The simplest and least error-prone way to meet this requirement is for the original tester to take production hardware, production software, a SPEC config file, and the SPEC tools and actually build the benchmarks in a single invocation of runaccel on the System Under Test (SUT). But SPEC realizes that there is a cost to benchmarking and would like to address this, for example through the rules that follow regarding cross-compilation and individual builds. However, in all cases, the tester is taken to assert that the compiled executables will exhibit the same performance as if they all had been compiled with a single invocation of runaccel (see 2.1.7).
SPECaccel base binaries must be built using the environment rules of section 2.1.2, and must not rely upon any changes to the environment during the build.
Note 1: Base cross compilations using multiple hosts are allowed (2.1.5), but the performance of the resulting binaries must not depend upon environmental differences among the hosts. It must be possible to build performance-equivalent base binaries with one set of switches (2.3.1), in one execution of runaccel (2.1.3), on one host, with one environment (2.1.2).
For a peak build, the environment may be changed, subject to the following constraints:
Note 2: Peak cross compilations using multiple hosts are allowed (2.1.5), but the performance of the resulting binaries must not depend upon environmental differences among the hosts. It must be possible to build performance-equivalent peak binaries with one config file, in one execution of runaccel (2.1.3), in the same execution of runaccel that built the base binaries, on one host, starting from the environment used for the base build (2.1.2), and changing that environment only through config file hooks (2.1.4).
It is permitted to use cross-compilation, that is, a building process where the benchmark executables are built on a host (or hosts) that differ(s) from the SUT. The runaccel tool must be used on all systems (typically with -a build on the host(s) and -a validate on the SUT).
If all systems belong to the same product family and if the software used to build the executables is available on all systems, this does not need to be documented. In the case of a true cross compilation, (e.g. if the software used to build the benchmark executables is not available on the SUT, or the host system provides performance gains via specialized tuning or hardware not on the SUT), the host system(s) and software used for the benchmark building process must be documented in the Notes section. See section 4.
It is permitted to use more than one host in a cross-compilation. If more than one host is used in a cross-compilation, they must be sufficiently equivalent so as not to violate rule 2.1.3. That is, it must be possible to build the entire suite on a single host and obtain binaries that are equivalent to the binaries produced using multiple hosts.
The purpose of allowing multiple hosts is so that testers can save time when recompiling many programs. Multiple hosts must NOT be used in order to gain performance advantages due to environmental differences among the hosts. In fact, the tester must exercise great care to ensure that any environment differences are performance neutral among the hosts, for example by ensuring that each has the same version of the operating system, the same performance software, the same compilers, and the same libraries. The tester must exercise due diligence to ensure that differences that appear to be performance neutral - such as differing MHz or differing memory amounts on the build hosts - are in fact truly neutral.
Multiple hosts must NOT be used in order to work around system or compiler incompatibilities (e.g. compiling the C benchmarks on a different OS version than the Fortran benchmarks in order to meet the different compilers' respective OS requirements), since that would violate the Continuous Build rule (2.1.3).
It is permitted to build the benchmarks with multiple invocations of runaccel, for example during a tuning effort. But, the executables must be built using a consistent set of software. If a change to the software environment is introduced (for example, installing a new version of the C compiler which is expected to improve the performance of one of the floating point benchmarks), then all affected benchmarks must be rebuilt (in this example, all the C benchmark).
The previous 4 rules may appear to contradict each other (2.1.3 through 2.1.6), but the key word in 2.1.3 is the word "possible".
Consider the following sequence of events:
In this example, the tester is taken to be asserting that the above sequence of events produces binaries that are, from a performance point of view, equivalent to binaries that it would have been possible to build in a single invocation of the tools.
- A tester has built a complete set of SPECaccel executable images ("binaries") on her usual host system.
- A hot new SUT arrives for a limited period of time. It has no compilers installed.
- A SPECaccel tree is installed on the SUT, along with the binaries and config file generated on the usual host.
- It is learned that performance could be improved if the peak version of 999.sluggard were compiled with -O5 instead of -O4.
- On the host system, the tester edits the config file to change to -O5 for 999.sluggard, and issues the command:
runaccel -c myconfig -D -a build -T peak sluggard- The tester copies the new binary and config file to the SUT
- A complete run is started by issuing the command:
runaccel -c myconfig -a validate all- Performance is as expected, and the results are published at SPEC (including the config file).
If there is some optimization that can only be applied to individual benchmark builds, but which it is not possible to apply in a continuous build, the optimization must not be used.
Rule 2.1.7 is intended to provide some guidance about the kinds of practices that are reasonable, but the ultimate responsibility for result reproducibility lies with the tester. If the tester is uncertain whether a cross-compile or an individual benchmark build is equivalent to a full build on the SUT, then a full build on the SUT is required (or, in the case of a true cross-compilation which is documented as such, then a single runaccel -a build is required on a single host.) Although full builds add to the cost of benchmarking, in some instances a full build in a single runaccel may be the only way to ensure that results will be reproducible.
The following rules apply to compiler flag selection and source code changes for SPECaccel Peak and Base Metrics. Additional rules for Base Metrics follow in section 2.3.
Benchmark source file, variable, and subroutine names must not be used within optimization flags or compiler/build options.
Identifiers used in preprocessor directives to select alternative source code are also forbidden, except for a rule-compliant library substitution (2.2.2), an approved portability flag (2.2.4), or a specifically provided SPEC-approved alternate source (src.alt).
For example, if a benchmark source code uses one of:
#ifdef IDENTIFIER #ifndef IDENTIFIER #if defined IDENTIFIER #if !defined IDENTIFIER
to provide alternative source code under the control of a compiler option such as -DIDENTIFIER, such a switch may not be used unless it meets the criteria of 2.2.2 or 2.2.4.
Flags which substitute pre-computed (e.g. library-based) routines for routines defined in the benchmark on the basis of the routine's name must not be used. Exceptions are:
Note: This rule does not forbid flags that select alternative implementations of library functions defined in an ANSI/ISO language standard. For example, such flags might select an optimized library of these functions, or allow them to be inlined.
Flags that change a data type size to a size different from the default size of the compilation system are not allowed to be used in base builds. Exceptions are:
which are acceptable as portability flags in base builds, and may be used as (or to facilitate) optimizations in peak builds.
Pointer size changes, in particular, may be used to make the benchmark binaries execute within a 32-bit or 64-bit address space. Section 2.3.6 states the restrictions on using mixtures of 32-bit and 64-bit binaries in base measurements.
Rule 2.3.2 requires that all benchmarks use the same flags in base. Portability flags are an exception to this rule: they may differ from one benchmark to another, even in base. Such flags are subject to two major requirements:
A proposed portability flag will normally not be approved unless it is essential in order to successfully build and run the benchmark.
A proposed portability flag will normally be approved if one of the following conditions holds:
If more than one solution can be used for a problem, the subcommittee will review attributes such as precedent from previously published results, performance neutrality, standards compliance, amount of code affected, impact on the expressed original intent of the program, and good coding practices (in rough order of priority).
If a benchmark is discovered to violate the relevant standard, that may or may not be reason for the subcommittee to grant a portability flag. If the justification for a portability flag is standards compliance, the tester must include a specific reference to the offending source code module and line number, and a specific reference to the relevant sections of the appropriate standard. The tester should also address impact on the other attributes mentioned in the previous paragraph.
If a given portability problem (within a given language) occurs in multiple places within a suite, then, in base, the same method(s) must be applied to solve all instances of the problem.
Source code changes to the OpenACC and OpenMP directives for peak are allowed as defined in section 2.4.5). For all other SPECaccel metrics, source code modifications are not allowed except for performance-neutral changes to allow for standard compliance (see section 1.8).
In addition to the rules listed in section 2.2 above, the selection of optimizations to be used to produce SPECaccel Base Metrics includes the following:
The optimizations used are expected to be safe, and it is expected that system or compiler vendors would endorse the general use of these optimizations by customers who seek to achieve good application performance.
The requirements that optimizations be safe, and that they generate correct code for a class of programs larger than the suite itself (rule 1.4), are normally interpreted as requiring that the system, as used in base, implement the language correctly. "The language" is defined by the appropriate ANSI/ISO standard (C99, Fortran 2003).
The principle of standards conformance is not automatically applied, because SPEC has historically allowed certain exceptions:
Otherwise, a deviation from the standard that is not performance neutral, and that gives the particular implementation a SPECaccel performance advantage over standard-conforming implementations, is considered an indication that the requirements about "safe" and "correct code" optimizations are probably not met. Such a deviation may be a reason for SPEC to find a result not rule-conforming.
If an optimization causes any SPECaccel benchmark to fail to validate, and if the relevant portion of this benchmark's code is within the language standard, then the failure is taken as additional evidence that an optimization is not safe.
Regarding C++: Note that for C++ applications, the standard calls for support of both run-time type information (RTTI) and exception handling. The compiler, as used in base, must enable these.
For example, a compiler enables exception handling by default; it can be turned off with --noexcept. The switch --noexcept is not allowed in base.
For example, a compiler defaults to no run time type information, but allows it to be turned on via --rtti. The switch --rtti must be used in base.
Regarding accuracy: Because language standards generally do not set specific requirements for accuracy, SPEC has also chosen not to do so. Nevertheless:
In cases where the class of appropriate applications appears to be so narrowly drawn as to constitute a "benchmark special", that may be a reason for SPEC to find a result non-conforming.
In base, the same compiler must be used for all modules of a given language within a benchmark suite. Except for portability flags (see 2.2.4 above), all flags or options that affect the transformation process from SPEC-supplied source to completed executable must be the same, including but not limited to:
All flags must be applied in the same order for all compiles of a given language.
Note that the SPEC tools provide methods to set flags on a per-language basis.
For example, if a tester sets:
default=base: COPTIMIZE = -O4 FOPTIMIZE = -O5
then the C benchmarks will be compiled with -O4 and the Fortran benchmarks with -O5. (This is legal: there is no requirement to compile C codes with the same optimization level as Fortran codes).
Regarding benchmarks that have been written in more than one language:
In a mixed-language benchmark, the tools automatically compile each source module with the options that have been set for its language.
Continuing the example just above, a benchmark that uses both C and Fortran would have its C modules compiled with -O4 and its Fortran modules with -O5. This, too, is legal.
In order to link an executable for a mixed-language benchmark, the tools need to decide which link options to apply (e.g. those defined in CLD/CLDOPT vs. those in FLD/FLDOPT vs. those in CXXLD/CXXLDOPT). This decision is based on benchmark classifications that were determined during development of the SPECaccel suite. For reasons of link time library inclusion, the classifications were not made based on percentage of code nor on the language of the main routine; rather, the classifications had been set to either F (for mixed Fortran/C benchmarks) or CXX (for benchmarks that include C++).
Link options must be consistent in a base build. For example, if FLD is set to /usr/opt/advanced/ld for pure Fortran benchmarks, the same setting must be used for any mixed language benchmarks that have been classified, for purpose of linking, as Fortran.
Inter-module optimization and mixed-language benchmarks:
For mixed-language benchmarks, if the compilers have an incompatible inter-module optimization format, flags that require inter-module format compatibility may be dropped from base optimization of mixed-language benchmarks. The same flags must be dropped from all benchmarks that use the same combination of languages. All other base optimization flags for a given language must be retained for the modules of that language.
For example, suppose that a suite has exactly two benchmarks that employ both C and Fortran, namely 997.CFmix1 and 998.CFmix2. A tester uses a C compiler and Fortran compiler that are sufficiently compatible to be able to allow their object modules to be linked together - but not sufficiently compatible to allow inter-module optimization. The C compiler spells its intermodule optimization switch -ifo, and the Fortran compiler spells its switch --intermodule_optimize. In this case, the following would be legal:
default=base: COPTIMIZE = -fast -O4 -ur=8 -ifo FOPTIMIZE = --prefetch:all --optimize:5 --intermodule_optimize FLD=/usr/opt/advanced/ld FLDOPT=--nocompress --lazyload --intermodule_optimize 997.CFmix1,998.CFmix2=base: COPTIMIZE = -fast -O4 -ur=8 FOPTIMIZE = --prefetch:all --optimize:5 FLD=/usr/opt/advanced/ld FLDOPT=--nocompress --lazyload
Following the precedence rules as explained in config.html, the above section specifiers set default tuning for the C and Fortran benchmarks, but the tuning is modified for the two mixed-language benchmarks to remove switches that would have attempted inter-module optimization.
An assertion flag is one that supplies semantic information that the compilation system did not derive from the source statements of the benchmark.
With an assertion flag, the programmer asserts to the compiler that the program has certain nice properties that allow the compiler to apply more aggressive optimization techniques (for example, that there is no aliasing via C pointers). The problem is that there can be legal programs (possibly strange, but still standard-conforming programs) where such a property does not hold. These programs could crash or give incorrect results if an assertion flag is used. This is the reason why such flags are sometimes also called "unsafe flags". Assertion flags should never be applied to a production program without previous careful checks; therefore they must not be used for base.
Exception: a tester is free to turn on a flag that asserts that the benchmark source code complies to the relevant standard (e.g. -ansi_alias). Note, however, that if such a flag is used, it must be applied to all compiles of the given language (C, C++, or Fortran), while still passing SPEC's validation tools with correct answers for all the affected programs.
Base results may use flags which affect the numerical accuracy or sensitivity by reordering floating-point operations based on algebraic identities, provided of course that the result validates.
The system environment must not be manipulated during a build of the base binaries. For example, suppose that an environment variable called BIGPAGES can be set to yes or no, and the default is no. The tester must not change the choice during the build of the base binaries. See section 2.1.4.
Normally, it is expected that the data model (such as pointer sizes, sizes of int, etc) will be consistent in base for all compilations of a given language. In particular, compilers provide switches like -m64 to select between data declarations of different sizes. If one of these flags is used in base, then normally it should be set for all benchmarks of the given language in the suite for base.
If for some reason it is not practical to use a consistent data model in base, the following rules apply:
If no consistent combination of benchmark code switches and rule 1 can be found to work, possibly due to mixed-language benchmarks in the suite, the tester could describe the problem to SPEC and request that SPEC allow use of an inconsistent data model in base. SPEC would consider such a request using the same process outlined in rule 2.2.4, including consideration of the technical arguments as to the nature of the data model problem and consideration of the practicality of technical alternatives, if any. SPEC might or might not grant the portability flag. SPEC might also choose to fix source code limitations, if any, that are causing difficulty.
Frequently, performance may be improved via optimizations that work across source modules, for example -ifo, -xcrossfile, or -IPA. Some compilers may require the simultaneous presentation of all source files for inter-file optimization, as in:
cc -ifo -o a.out file1.c file2.c
Other compilers may be able to do cross-module optimization even with separate compilation, as in:
cc -ifo -c -o file1.o file1.c cc -ifo -c -o file2.o file2.c cc -ifo -o a.out file1.o file2.o
Tthe SPEC tools operate in the latter mode, the former may not be used.
Switches that cause data to be aligned on natural boundaries may be used in base.
The general rules listed in section 2.2 apply to compiler flag selection for SPECaccel Peak Metrics. Peak Metrics do not need to follow the additional rules listed in section 2.3.
In peak, different compilers may be used for each benchmark. Options that may be tuned per benchmark include:
The system environment may be manipulated during a build of the peak binaries. For example, suppose that an environment variable called BIGPAGES can be set to yes or no, and the default is no. The tester may change the choice during the build/run of the peak binaries via setting the "ENV_" option (See: envars) in the configuration file's section for a particular benchmark.
Assertion flags, where one that supplies semantic information that the compilation system did not derive from the source statements of the benchmark, may be used in peak provided that the benchmark validates. However, such flags must still adhere to the general rules as described in the rules listed in section 2.2 above.
The parallel model, OpenACC or OpenMP, and method for OpenMP may be different for each benchmark in peak.
Submitters may modify a benchmark's OpenACC or OpenMP directives in peak runs for architecture tuning with the following restrictions:
The alternate source code package must be submitted to SPEC/HPG for review and approval before they can be used in a compliant results. The srcalt must be made available for other SPECaccel license holders to use.
The operating system state (multi-user, single-user, init level N) may be selected by the tester. This state along with any changes in the default configuration of daemon processes or system tuning parameters must be documented in the notes section of the results disclosure. (For Microsoft Windows, system state is normally "Default"; a list of services that are shut down should be provided, if any, e.g. print spooler shut down).
The SPECaccel suite requires that a single file system be used to contain the installed directory tree. Additional file systems may be used to store temporary and run directories. A single shared run-directory must be used for each benchmark in a base run. Peak runs are allowed to replicate run directories, and the directories and file systems can be arranged differently for different benchmarks.
SPEC allows any type of file system (disk-based, memory-based, NFS, DFS, FAT, NTFS etc.) to be used. The type and arrangement of the directories and file systems must be disclosed in reported results.
A reportable run consists of either two or three runs of the suite. The reportable result will be the median of the three runs, or the less performing of the two.
The config file directive submit is the preferred means to assign work to processors. The tester may, if desired:
The submit directive can be used to change the run time environment (see section 3.3).
In peak, different benchmarks may use different submit directives.
Run-time environment settings are treated similarly to compilation options in the SPECaccel suite. The rules are as follows:
The usage rules are as follows, from highest precedence to lowest:
submit = export PROFILE_DIR=...; ....Settings are hard to discern if too many details are packed into the submit line. One advantage, however, is that changing the settings does not cause anything to rebuild.
env_vars = 1 ENV_PROFILE_DIR = ...The settings are quite clear from reading the text of the config file. A disadvantage is that the settings also apply to the build phase, and changing a setting will cause the affected benchmarks to rebuild.
export PROFILE_DIR=... runaccel ...The settings are invisible to the automatic report generation and must be carefully documented.
All benchmark executions, including the validation steps, contributing to a particular submittable report must occur continuously, that is, in one execution of runaccel. For a reportable run, the runaccel tool will run all three workloads (test, train, and ref), and will ensure that the correct answer is obtained for all three. (Note: the execution and validation of test and train is not part of the timing of the benchmark - it is only an additional test for correct operation of the binary.)
If a submittable report will contain both base and peak measurements, a single runaccel invocation must be used for the runs. When both base and peak are run, the tools run the base executables first, followed by the peak executables.
It is permitted to publish base results as peak. This can be accomplished in various ways, all of which are allowed:
Note: It is permitted but not required to compile in the same runaccel invocation as the execution. See Section 2.1.5 regarding cross compilation.
As used in these run rules, the term "run-time dynamic optimization" (RDO) refers broadly to any method by which a system adapts to improve performance of an executing program based upon observation of its behavior as it runs. This is an intentionally broad definition, intended to include techniques such as:
RDO may be under control of hardware, software, or both.
Understood this broadly, RDO is already commonly in use, and usage can be expected to increase. SPEC believes that RDO is useful, and does not wish to prevent its development. Furthermore, SPEC views at least some RDO techniques as appropriate for base, on the grounds that some techniques may require no special settings or user intervention; the system simply learns about the workload and adapts.
However, benchmarking a system that includes RDO presents a challenge. A central idea of SPEC benchmarking is to create tests that are repeatable: if you run a benchmark suite multiple times, it is expected that results will be similar, although there will be a small degree of run-to-run variation. But an adaptive system may recognize the program that it is asked to run, and "carry over" lessons learned in the previous execution; therefore, it might complete a benchmark more quickly each time it is run. Furthermore, unlike in real life, the programs in the benchmark suite are presented with the same inputs each time they are run: value prediction is too easy if the inputs never change. In the extreme case, an adaptive system could be imagined that notices which program is about to run, notices what the inputs are, and which reduces the entire execution to a print statement. In the interest of benchmarking that is both repeatable and representative of real-life usage, it is therefore necessary to place limits on RDO carry-over.
Run time dynamic optimization is allowed, subject to the usual provisions that the techniques must be generally available, documented, and supported. It is also subject to the conditions listed in the rules immediately following.
Section 4.2 applies to run-time dynamic optimization: any settings which the tester has set to non-default values must be disclosed. Resources consumed by RDO must be included in the description of the hardware configuration as used by the benchmark suite.
For example, suppose that a system can be described as a 64-core system. After experimenting for a while, the tester decides that the optimum performance is achieved by dedicating 4 cores to the run-time dynamic optimizer, and running the benchmarks with only 60 threads. The system under test is still correctly described as a 64-core system, even though only 60 cores were used to run SPEC code.
Run time dynamic optimization is subject to Section 3.4: settings cannot be changed at run-time. But Note 2 of rule 3.4 also applies to RDO: for example, in peak it would be acceptable to compile a subset of the benchmarks with a flag that suggests to the run-time dynamic optimizer that code rearrangement should be attempted. Of course, Section 2.2.1 also would apply: such a flag could not tell RDO which routines to rearrange.
If run-time dynamic optimization is effectively enabled for base (after taking into account the system state at run-time and any compilation flags that interact with the run-time state), then RDO must comply with 2.3.1, the safety rule. It is understood that the safety rule has sometimes required judgment, including deliberation by SPEC in order to determine its applicability. The following is intended as guidance for the tester and for SPEC:
SPEC has an interest in preventing carry-over of information from run to run. Specifically, no information may be carried over which identifies the specific program or executable image. Here are some examples of behavior that is, and is not, allowed.
It doesn't matter whether the information is intentionally stored, or just "left over"; if it's about a specific program, it's not allowed:
If information is left over from a previous run that is not associated with a specific program, that is allowed:
Any form of RDO that uses memory about a specific program is forbidden:
The system is allowed to respond to the currently running program, and to the overall workload:
SPEC requires a full disclosure of results and configuration details sufficient to reproduce the results. For results published on its web site, SPEC also requires that base results be published whenever peak results are published. If peak results are published outside of the SPEC web site in a publicly available medium, the tester must supply base results on request. Publication of results under non-disclosure or company internal use or company confidential are not "publicly" available.
A full disclosure of results must include:
A full disclosure of results must include sufficient information to allow a result to be independently reproduced. If a tester is aware that a configuration choice affects performance, then s/he must document it in the full disclosure.
Note: this rule is not meant to imply that the tester must describe irrelevant details or provide massively redundant information.
For example, if the SuperHero Model 1
comes with a write-through cache, and the SuperHero Model 2 comes with
a write-back cache, then specifying the model number is sufficient, and
no additional steps need to be taken to document the cache protocol.
But if the Model 3 is available with both write-through and write-back
caches, then a full disclosure must specify which cache is used.
For information on how to publish a result on SPEC's web site, contact the SPEC office. Contact information is maintained at the SPEC web site, http://www.spec.org/.
If a tester publishes results for a hardware or software configuration that has not yet shipped,
Note 1: "Generally available" is defined in the SPEC High Performance Group Policy document, which can be found at http://www.spec.org/hpg/policy.html.
Note 2: It is acceptable to test larger configurations than customers are currently ordering, provided that the larger configurations can be ordered and the company is prepared to ship them.
A "pre-production", "alpha", "beta", or other pre-release version of a compiler (or other software) can be used in a test, provided that the performance-related features of the software are committed for inclusion in the final product.
The tester must practice due diligence to ensure that the tests do not use an uncommitted prototype with no particular shipment plans. An example of due diligence would be a memo from the compiler Project Leader which asserts that the tester's version accurately represents the planned product, and that the product will ship on date X.
The final, production version of all components must be generally available within 90 days after first public release of the result.
When specifying a software component name in the results disclosure, the component name that should be used is the name that customers are expected to be able to use to order the component, as best as can be determined by the tester. It is understood that sometimes this may not be known with full accuracy; for example, the tester may believe that the component will be called "TurboUnix V5.1.1" and later find out that it has been renamed "TurboUnix V5.2", or even "Nirvana 1.0". In such cases, an editorial request can be made to update the result after publication.
Some testers may wish to also specify the exact identifier of the version actually used in the test (for example, "build 20070604"). Such additional identifiers may aid in later result reproduction, but are not required; the key point is to include the name that customers will be able to use to order the component.
The configuration disclosure includes fields for both "Hardware Availability" and "Software Availability". In both cases, the date which must be used is the date of the component which is the last of the respective type to become generally available. The date is specified Mmm-YYYY as in the following examples: Jan-2007, Feb-2007. The Month is abbreviated to three letters with the first letter capitalized. A hyphen separates the Month and Year fields. The Year field is specified with four digits.
Since all components must be available within 90 days of the first public release of the result, the first day of the specified Month (and Year) must fall within this 90 day window.
If a software or hardware date changes, but still falls within 90 days of first publication, a result page may be updated on request to SPEC.
If a software or hardware date changes to more than 90 days after first publication, the result is considered Non Compliant. For procedures regarding Non Compliant results, see the SPEC High Performance Group Policy Document, http://www.spec.org/hpg/policy.html.
SPEC is aware that performance results for pre-production systems may sometimes be subject to change, for example when a last-minute bugfix reduces the final performance.
For results measured on pre-production systems, if the tester becomes aware of something that will reduce production system performance by more than 5% on an overall metric, the tester is required to republish the result, and the original result shall be considered Non Compliant.
Analogous rules apply to performance changes across post-production upgrades (Section 4.3.5).
The SPEC website posts a set of reports, each consisting of a measured result and the description of the system being measured. A SPECaccel report is broken down into separate sections as follows:
The measured results are captured in the submission file by the runaccel script when a run is made. The other description information is read from the config file and captured in the preamble of the submission file. This preamble can be edited after the run is made, if the information in the config file was incorrect or incomplete. The config file document describes how the fields are to be written. The following subsections describe the information to provide for each piece of the disclosure report, giving examples in some cases.
Details are:
SPEC recommends that measurements be done on the actual systems for which results are claimed. Nevertheless, SPEC recognizes that there is a cost of benchmarking, and that multiple publications from a single measurement may sometimes be appropriate. For example, two systems badged as "Model A" versus "Model B" may differ only in the badge itself; in this situation, differences are sometimes described as only "paint deep", and a tester may wish to perform only a single test (i.e. the runaccel tool is invoked only once, and multiple rawfiles are prepared with differing system descriptions).
Although paint is usually not a performance-relevant difference, for other differences it can be difficult to draw a precise line as to when two similar systems should no longer be considered equivalent. For example, what if Model A and B come from different vendors? Use differing firmware, power supplies, or line voltage? Support additional types or numbers of disks, or other devices?
For the SPECaccel suite, a single measurement may be published as multiple equivalent results provided that all of the following requirements are met:
Performance differences from factors such as those listed in the paragraph above (paint, vendor, firmware, and so forth) are within normal run-to-run variation.
The CPU is the same.
The motherboards are the same:
same motherboard manufacturer
same electrical devices (for example, IO support chips, memory slots, PCI slots)
same physical shape.
The memory systems are the same:
same caches
same number of memory modules
memory modules are run at the same speed
memory modules comply with same specifications, where applicable (for example, the same labels as determined by the JEDEC DDR3 DIMM Label Specification).
As tested, all hardware components are supported on both systems.
For example, the Model A and Model B meet the requirements listed above, including a motherboard with the same number of DIMM slots. The Model A can be fully populated with 96 DIMMs. Due to space and thermal considerations, the Model B can only be half-populated; i.e. it is not supported with more than 48 DIMMs. If the actual sytem under test is the Model A, the tester must fill only the DIMM slots that are allowed to be filled for both systems.
Disclosures must reference each other, and must state which system was used for the actual measurement. For example:
This result was measured on the Acme Model A. The Acme Model A and the Bugle Model B are equivalent.
No power measurement is made. A single power measurement may not be published as multiple equivalent systems.
When a single measurement is used for multiple systems, SPEC may ask for a review of the differences between the systems, may ask for substantiation of the requirements above, and/or may require that additional documentation be included in the publications.
The host will consist of one node. The following configuration information needs to be supplied.
The current fields are:
Regarding the fields in the above list that mention the word "enabled": if a chip, core, or thread is available for use during the test, then it must be counted. If one of these resources is disabled - for example by a firmware setting prior to boot - then it need not be counted, but the tester must exercise due diligence to ensure that disabled resources are truly disabled, and not silently giving help to the result.
Regarding the field (hw_ncoresperchip), the tester must count the cores irrespective of whether they are enabled.
Example: In the following tests, the host is a Turboblaster Model 32-64-256, which contains 32 chips. Each chip has 4 cores. Each core can run 4 hardware threads.
hw_ncores: 128 hw_nchips: 32 hw_nthreadspercore: 4
hw_ncores: 32 hw_nchips: 32 hw_nthreadspercore: 1
Note: if resources are disabled, the method(s) used for such disabling must be documented and supported.
A TurboCabinet contains 4 chips.
System State: Run level 4 (multi-user with display manager)
On other systems:
Note: some Unix (and Unix-like) systems have deprecated the concept of "run levels", preferring other terminology for state description. In such cases, the system state field should use the vocabulary recommended by the operating system vendor.
Hardware description information may be gathered automatically from sysinfo or manually added by the tester.
This section describes the configuration of the device. The following information needs to be supplied. Details are:
This section describes the compiler invocation and running of the benchmarks. The following configuration information needs to be supplied.
Scripted Installations and Pre-configured Software: In order to reduce the cost of benchmarking, test systems are sometimes installed using automatic scripting, or installed as preconfigured system images. A tester might use a set of scripts that configure the corporate-required customizations for IT Standards, or might install by copying a disk image that includes best practices of the performance community. SPEC understands that there is a cost to benchmarking, and does not forbid such installations, with the proviso that the tester is responsible to disclose how end users can achieve the claimed performance (using appropriate fields above).
Example: the Corporate Standard Jumpstart Installation Script has 73 documented customizations and 278 undocumented customizations, 34 of which no one remembers. Of the various customizations, 17 are performance relevant for the SPECaccel benchmark - and 4 of these are in the category "no one remembers". The tester is nevertheless responsible for finding and documenting all 17. Therefore to remove doubt, the tester prudently decides that it is less error-prone and more straightforward to simply start from customer media, rather than the Corporate Jumpstart.
Any additional notes such as listing any use of SPEC-approved alternate sources or tool changes.
For example, suppose the tester uses a pre-release compiler with:
f90 -O4 --newcodegen --loopunroll:outerloop:alldisable
but the tester knows that the new code generator will be automatically applied in the final product, and that the spelling of the unroll switch will be simpler than the spelling used here. The recommended spelling for customers who wish to achieve the effect of the above command will be:
f90 -O4 -no-outer-unroll
In this case, the flags report will include the actual spelling used by the tester, but a note should be added to document the spelling that will be recommended for customers.
The SPECaccel suite provides benchmarks in source code form, which are compiled under control of SPEC's toolset. The SPEC tools automatically detect the use of compilation and linkage flags in the config file and document them in the benchmark configuration section of the final report. Both portability and optimization flags will be captured in the report subsection.
The SPEC tools require a flag description file which provides information about the syntax of the flags and their meanings. A result will be marked "invalid" unless it has an associated flag description file. A description of how to write one may be found at www.spec.org/accel2023/Docs .
The level of detail in the description of a flag is expected to be sufficient so that an interested technical reader can form a preliminary judgment of whether he or she would also want to apply the option.
It is acceptable, and even common practice, for testers to build on each others' flags files, copying all or part of flags files posted by others into their own flags files; but doing so does not relieve an individual tester of the responsibility to ensure that the description is accurate.
Although these descriptions have historically been called "flags files", they must also include descriptions of other performance-relevant options that have been selected, including but not limited to environment variables, kernel options, file system tuning options, BIOS options, and options for any other performance-relevant software packages.
SPECaccel results are for systems, not just for chips: it is required that a user be able to obtain the system described in the result page and reproduce the result (within a small range for run-to-run variation).
Nevertheless, SPEC recognizes that chip and motherboard suppliers have a legitimate interest in accelerator benchmarking. For those suppliers, the performance-relevant hardware components typically are the accelerator, cpu chip, motherboard, and memory; but users would not be able to reproduce a result using only those four. To actually run the benchmarks, the user has to supply other components, such as a case, power supply, and disk; perhaps also a specialized CPU cooler, extra fans, a disk controller, graphics card, network adapter, BIOS, and configuration software.
Such systems are sometimes referred to as "white box", "home built", "kit built", or by various informal terms. For SPEC purposes, the key point is that the user has to do extra work in order to reproduce the performance of the tested components; therefore, this document refers to such systems as "user built".
For user built systems, the configuration disclosure must supply a parts list sufficient to reproduce the result. As of the listed availability dates in the disclosure, the user should be able to obtain the items described in the disclosure, spread them out on an anti-static work area, and, by following the instructions supplied with the components, plus any special instructions in the SPEC disclosure, build a working system that reproduces the result. It is acceptable to describe components using a generic name (e.g. "Any ATX case"), but the recipe must also give specific model names or part numbers that the user could order (e.g. "such as a Mimble Company ATX3 case").
Component settings that are listed in the disclosure must be within the supported ranges for those components. For example, if the memory timings are manipulated in the BIOS, the selected timings must be supported for the chosen type of memory.
Components for a user built system may be divided into two kinds: performance-relevant (for the SPECaccel suite), and non-performance-relevant. For example, SPECaccel benchmark scores are affected by memory speed, and motherboards often support more than one choice for memory; therefore, the choice of memory type is performance-relevant. By contrast, the motherboard needs to be mounted in a case. Which case is chosen in not normally performance-relevant; it simply has to be the correct size (e.g. ATX, microATX, etc).
Example:
hw_cpu_name = Frooble 1500 hw_memory = 2 GB (2x 1GB Mumble Inc Z12 DDR2 1066) sw_other = SnailBios 17 notes_plat_000 = notes_plat_005 = The BIOS is the Mumble Inc SnailBios Version 17, notes_plat_010 = which is required in order to set memory timings notes_plat_015 = manually to DDR2-800 5-5-5-15. The 2 DIMMs were notes_plat_020 = configured in dual-channel mode. notes_plat_025 = notes_plat_030 = A standard ATX case is required, along with a 500W notes_plat_035 = (minimum) ATX power supply [4-pin (+12V), 8-pin (+12V) notes_plat_040 = and 24-pin are required]. An AGP or PCI graphics notes_plat_045 = adapter is required in order to configure the system. notes_plat_050 = notes_plat_055 = The Frooble 1500 CPU chip is available in a retail box, notes_plat_060 = part 12-34567, with appropriate heatsinks and fan assembly. notes_plat_065 = notes_plat_070 = As tested, the system used a Mimble Company ATX3 case, notes_plat_075 = a Frimble Ltd PS500 power supply, and a Frumble notes_plat_080 = Corporation PCIe Z19 graphics adapter. notes_plat_085 =
Additional notes:
Note 1: Regarding graphics adapters:
Note 2: Regarding power modes: Sometimes CPU chips are capable of running with differing performance characteristics according to how much power the user would like to spend. If non-default power choices are made for a user built system, those choices must be documented in the notes section.
Note 3: Regarding cooling systems: Sometimes CPU chips are capable of running with degraded performance if the cooling system (fans, heatsinks, etc.) is inadequate. When describing user built systems, the notes section must describe how to provide cooling that allows the chip to achieve the measured performance.
The actual test results consist of the elapsed times and ratios for the individual benchmarks and the overall SPEC metric produced by running the benchmarks via the SPEC tools. The required use of the SPEC tools ensures that the results generated are based on benchmarks built, run, and validated according to the SPEC run rules.
Below is a list of the measurement components for each SPECaccel suite and metric:
These are calculated as follows:
All runs of a specific benchmark when using the SPEC tools are required to have validated correctly. The benchmark executables must have been built according to the rules described in section 2 above.
Publication of peak performance results are considered optional by SPEC, so the tester may choose to publish only base performance results. Since by definition base performance results adhere to all the rules that apply to peak performance results, the tester may choose to refer to these results by either the base or peak metric names (e.g. SPECaccel 2023_base or SPECaccel 2023_peak).
It is permitted to publish base-only results. Alternatively, the use of the flag basepeak is permitted, as described in section 3.6.
SPECaccel metrics may be estimated. All estimates must be clearly identified as such. It is acceptable to estimate a single metric (for example, SPECaccel 2023_base, or the elapsed seconds for the benchmark component). Note that it is permitted to estimate a peak metric without being required to provide a corresponding estimate for base.
SPEC requires that every use of an estimated number be clearly marked with "est." or "estimated" next to each estimated number, rather than burying a footnote at the bottom of a page.
For example, say that the JumboFast will achieve estimated performance of:
Model 1 SPECaccel 2023_base 50 est.
SPECaccel 2023_peak 60 est.
Model 2 SPECaccel 2023_base 70 est.
SPECaccel 2023_peak 80 est.
If estimates are used in graphs, the word "estimated" or "est." must be plainly visible within the graph, for example in the title, the scale, the legend, or next to each individual result that is estimated.
Note: the term "plainly visible" in this rule is not defined; it is intended as a call for responsible design of graphical elements. Nevertheless, for the sake of giving at least rough guidance, here are two examples of the right way and wrong way to mark estimated results in graphs:
Licensees are encouraged to give a rationale or methodology for any estimates, together with other information that may help the reader assess the accuracy of the estimate. For example:
Those who publish estimates are encouraged to publish actual SPEC ACCEL metrics as soon as possible.
As mentioned previously, performance may sometimes change for pre-production systems; but this is also true of production systems (that is, systems that have already begun shipping). For example, a later revision to the firmware, or a mandatory OS bugfix, might reduce performance.
For production systems, if the tester becomes aware of something that reduces performance by more than 5% on an overall metric (for example, SPECaccel 2023_peak), the tester is encouraged but not required to republish the result. In such cases, the original result is not considered Non Compliant. The tester is also encouraged, but not required, to include a reference to the change that makes the results different (e.g. "with OS patch 20020604-02").
If a SPECaccel licensee publicly discloses an SPECaccel result (for example in a press release, academic paper, magazine article, or public web site), and does not clearly mark the result as an estimate, any SPEC member may request that the rawfile(s) from the run(s) be sent to SPEC. The rawfiles must be made available to all interested members no later than 10 working days after the request. The rawfile is expected to be complete, including configuration information (section 4.2 above).
A required disclosure is considered public information as soon as it is provided, including the configuration description.
For example, Company A claims a result of 1000 SPECaccel 2023_peak. A rawfile is requested, and supplied. Company B notices that the result was achieved by stringing together 50 chips in single-user mode. Company B is free to use this information in public (e.g. it could compare the Company A system vs. a Company B system that scores 999 using only 25 chips in multi-user mode).
Review of the result: Any SPEC member may request that a required disclosure be reviewed by the SPEC HPG subcommittee. At the conclusion of the review period, if the tester does not wish to have the result posted on the SPEC result pages, the result will not be posted. Nevertheless, as described above, the details of the disclosure are public information.
When public claims are made about SPECaccel results, whether by vendors or by academic researchers, SPEC reserves the right to take action if the rawfile is not made available, or shows different performance than the tester's claim, or has other rule violations.
SPEC encourages use of the SPECaccel suite in academic and research environments. It is understood that experiments in such environments may be conducted in a less formal fashion than that demanded of testers who publish on the SPEC web site. For example, a research environment may use early prototype hardware that simply cannot be expected to stay up for the length of time required to meet the continuous run requirement (see section 3.5), or may use research compilers that are unsupported and are not generally available (see section 1).
Nevertheless, SPEC would like to encourage researchers to obey as many of the run rules as practical, even for informal research. SPEC respectfully suggests that following the rules will improve the clarity, reproducibility, and comparability of research results.
Where the rules cannot be followed, SPEC requires that the deviations from the rules be clearly disclosed, and that any SPEC metrics (such as SPECaccel 2023_base) be clearly marked as estimated.
It is especially important to clearly distinguish results that do not comply with the run rules when the areas of non-compliance are major, such as not using the reference workload, or only being able to correctly validate a subset of the benchmarks.
Consistency and fairness are guiding principles for SPEC. to Help assure that these principles are met, any organization or individual who makes public use of SPEC benchmark results must do so in accordance with the SPEC Fair Use Rules, as posted at http://www.spec.org/fairuse.html.
If for some reason, the test sponsor cannot run the benchmarks as specified in these rules, the test sponsor can seek SPEC approval for performance-neutral alternatives. No publication may be done without such approval. The SPEC High Performance Group (HPG) maintains a Policies and Procedures document that defines the procedures for such exceptions.
Copyright © 2014-2023 Standard Performance Evaluation Corporation
All Rights Reserved