Dhrystone Benchmark: Rationale for Version 2 and Measurement Rules [published in SIGPLAN Notices 23,8 (Aug. 1988), 49-62] Reinhold P. Weicker Siemens AG, E STE 35 [now: Siemens AG, AUT E 51] Postfach 3220 D-8520 Erlangen Germany (West) 1. Why a Version 2 of Dhrystone? The Dhrystone benchmark program [1] has become a popular benchmark for CPU/compiler performance measurement, in particular in the area of minicomputers, workstations, PC's and microprocesors. It apparently satisfies a need for an easy-to-use integer benchmark; it gives a first performance indication which is more meaningful than MIPS numbers which, in their literal meaning (million instructions per second), cannot be used across different instruction sets (e.g. RISC vs. CISC). With the increasing use of the benchmark, it seems necessary to reconsider the benchmark and to check whether it can still fulfill this function. Version 2 of Dhrystone is the result of such a re-evaluation, it has been made for two reasons: o Dhrystone has been published in Ada [1], and Versions in Ada, Pascal and C have been distributed by Reinhold Weicker via floppy disk. However, the version that was used most often for benchmarking has been the version made by Rick Richardson by another translation from the Ada version into the C programming language, this has been the version distributed via the UNIX network Usenet [2]. There is an obvious need for a common C version of Dhrystone, since C is at present the most popular system programming language for the class of systems (microcomputers, minicomputers, workstations) where Dhrystone is used most. There should be, as far as possible, only one C version of Dhrystone such that results can be compared without restrictions. In the past, the C versions distributed by Rick Richardson (Version 1.1) and by Reinhold Weicker had small (though not significant) differences. Together with the new C version, the Ada and Pascal versions have been updated as well. o As far as it is possible without changes to the Dhrystone statistics, optimizing compilers should be prevented from removing significant statements. It has turned out in the past that optimizing compilers suppressed code generation for too many statements (by "dead code removal" or "dead variable elimination"). This has lead to the danger that benchmarking results obtained by a naive application of Dhrystone - without inspection of the code that was generated - could become meaningless. The overall policiy for version 2 has been that the distribution of statements, operand types and operand locality described in [1] should remain unchanged as much as possible. (Very few changes were necessary; their impact should be negligible.) Also, the order of statements should remain unchanged. Although I am aware of some critical remarks on the benchmark - I agree with several of them - and know some suggestions for improvement, I didn't want to change the benchmark into something different from what has become known as "Dhrystone"; the confusion generated by such a change would probably outweight the benefits. If I were to write a new benchmark program, I wouldn't give it the name "Dhrystone" since this denotes the program published in [1]. However, I do recognize the need for a larger number of representative programs that can be used as benchmarks; users should always be encouraged to use more than just one benchmark. The new versions (version 2.1 for C, Pascal and Ada) will be distributed as widely as possible. (Version 2.1 differs from version 2.0 distributed via the UNIX Network Usenet in March 1988 only in a few corrections for minor deficiencies found by users of version 2.0.) Readers who want to use the benchmark for their own measurements can obtain a copy in machine-readable form on floppy disk (MS-DOS or XENIX format) from the author. 2. Overall Characteristics of Version 2 In general, version 2 follows - in the parts that are significant for performance measurement, i.e. within the measurement loop - the published (Ada) version and the C versions previously distributed. Where the versions distributed by Rick Richardson [2] and Reinhold Weicker have been different, it follows the version distributed by Reinhold Weicker. (However, the differences have been so small that their impact on execution time in all likelihood has been negligible.) The initialization and UNIX instrumentation part - which had been omitted in [1] - follows mostly the ideas of Rick Richardson [2]. However, any changes in the initialization part and in the printing of the result have no impact on performance measurement since they are outside the measaurement loop. As a concession to older compilers, names have been made unique within the first 8 characters for the C version. The original publication of Dhrystone did not contain any statements for time measurement since they are necessarily system-dependent. However, it turned out that it is not enough just to inclose the main procedure of Dhrystone in a loop and to measure the execution time. If the variables that are computed are not used somehow, there is the danger that the compiler considers them as "dead variables" and suppresses code generation for a part of the statements. Therefore in version 2 all variables of "main" are printed at the end of the program. This also permits some plausibility control for correct execution of the benchmark. At several places in the benchmark, code has been added, but only in branches that are not executed. The intention is that optimizing compilers should be prevented from moving code out of the measurement loop, or from removing code altogether. Statements that are executed have been changed in very few places only. In these cases, only the role of some operands has been changed, and it was made sure that the numbers defining the "Dhrystone distribution" (distribution of statements, operand types and locality) still hold as much as possible. Except for sophisticated optimizing compilers, execution times for version 2.1 should be the same as for previous versions. Because of the self-imposed limitation that the order and distribution of the executed statements should not be changed, there are still cases where optimizing compilers may not generate code for some statements. To a certain degree, this is unavoidable for small synthetic benchmarks. Users of the benchmark are advised to check code listings whether code is generated for all statements of Dhrystone. Contrary to the suggestion in the published paper and its realization in the versions previously distributed, no attempt has been made to subtract the time for the measurement loop overhead. (This calculation has proven difficult to implement in a correct way, and its omission makes the program simpler.) However, since the loop check is now part of the benchmark, this does have an impact - though a very minor one - on the distribution statistics which have been updated for this version. 3. Discussion of Individual Changes In this section, all changes are described that affect the measurement loop and that are not just renamings of variables. All remarks refer to the C version; the other language versions have been updated similarly. In addition to adding the measurement loop and the printout statements, changes have been made at the following places: o In procedure "main", three statements have been added in the non-executed "then" part of the statement if (Enum_Loc == Func_1 (Ch_Index, 'C')) they are strcpy (Str_2_Loc, "DHRYSTONE PROGRAM, 3'RD STRING"); Int_2_Loc = Run_Index; Int_Glob = Run_Index; The string assignment prevents movement of the preceding assignment to Str_2_Loc (5'th statement of "main") out of the measurement loop (This probably will not happen for the C version, but it did happen with another language and compiler.) The assignment to Int_2_Loc prevents value propagation for Int_2_Loc, and the assignment to Int_Glob makes the value of Int_Glob possibly dependent from the value of Run_Index. o In the three arithmetic computations at the end of the measurement loop in "main ", the role of some variables has been exchanged, to prevent the division from just cancelling out the multiplication as it was in [1]. A very smart compiler might have recognized this and suppressed code generation for the division. o For Proc_2, no code has been changed, but the values of the actual parameter have changed due to changes in "main". o In Proc_4, the second assignment has been changed from Bool_Loc = Bool_Loc | Bool_Glob; to Bool_Glob = Bool_Loc | Bool_Glob; It now assigns a value to a global variable instead of a local variable (Bool_Loc); Bool_Loc would be a "dead variable" which is not used afterwards. o In Func_1, the statement Ch_1_Glob = Ch_1_Loc; was added in the non-executed "else" part of the "if" statement, to prevent the suppression of code generation for the assignment to Ch_1_Loc. o In Func_2, the second character comparison statement has been changed to if (Ch_Loc == 'R') ('R' instead of 'X') because a comparison with 'X' is implied in the preceding "if" statement. Also in Func_2, the statement Int_Glob = Int_Loc; has been added in the non-executed part of the last "if" statement, in order to prevent Int_Loc from becoming a dead variable. o In Func_3, a non-executed "else" part has been added to the "if" statement. While the program would not be incorrect without this "else" part, it is considered bad programming practice if a function can be left without a return value. To compensate for this change, the (non-executed) "else" part in the "if" statement of Proc_3 was removed. The distribution statistics have been changed only by the addition of the measurement loop iteration (1 additional statement, 4 additional local integer operands) and by the change in Proc_4 (one operand changed from local to global). The distribution statistics in the comment headers have been updated accordingly. 4. String Operations The string operations (string assignment and string comparison) have not been changed, to keep the program consistent with the original version. There has been some concern that the string operations are over-represented in the program, and that execution time is dominated by these operations. This was true in particular when optimizing compilers removed too much code in the main part of the program, this should have been mitigated in version 2. It should be noted that this is a language-dependent issue: Dhrystone was first published in Ada, and with Ada or Pascal semantics, the time spent in the string operations is, at least in all implementations known to me, considerably smaller. In Ada and Pascal, assignment and comparison of strings are operators defined in the language, and the upper bounds of the strings occuring in Dhrystone are part of the type information known at compilation time. The compilers can therefore generate efficient inline code. In C, string assignemt and comparisons are not part of the language, so the string operations must be expressed in terms of the C library functions "strcpy" and "strcmp". (ANSI C allows an implementation to use inline code for these functions.) In addition to the overhead caused by additional function calls, these functions are defined for null-terminated strings where the length of the strings is not known at compilation time; the function has to check every byte for the termination condition (the null byte). Obviously, a C library which includes efficiently coded "strcpy" and "strcmp" functions helps to obtain good Dhrystone results. However, I don't think that this is unfair since string functions do occur quite frequently in real programs (editors, command interpreters, etc.). If the strings functions are implemented efficiently, this helps real programs as well as benchmark programs. I admit that the string comparison in Dhrystone terminates later (after scanning 20 characters) than most string comparisons in real programs. For consistency with the original benchmark, I didn't change the program despite this weakness. 5. Intended Use of Dhrystone When Dhrystone is used, the following "ground rules" apply: o Separate compilation (Ada and C versions) As mentioned in [1], Dhrystone was written to reflect actual programming practice in systems programming. The division into several compilation units (5 in the Ada version, 2 in the C version) is intended, as is the distribution of inter-module and intra-module subprogram calls. Although on many systems there will be no difference in execution time to a Dhrystone version where all compilation units are merged into one file, the rule is that separate compilation should be used. The intention is that real programming practice, where programs consist of several independently compiled units, should be reflected. This also has implies that the compiler, while compiling one unit, has no information about the use of variables, register allocation etc. occuring in other compilation units. Although in real life compilation units will probably be larger, the intention is that these effects of separate compilation are modeled in Dhrystone. A few language systems have post-linkage optimization available (e.g., final register allocation is performed after linkage). This is a borderline case: Post-linkage optimization involves additional program preparation time (although not as much as compilation in one unit) which may prevent its general use in practical programming. I think that since it defeats the intentions given above, it should not be used for Dhrystone. Unfortunately, ISO/ANSI Pascal does not contain language features for separate compilation. Although most commercial Pascal compilers provide separate compilation in some way, we cannot use it for Dhrystone since such a version would not be portable. Therefore, no attempt has been made to provide a Pascal version with several compilation units. o No procedure merging Although Dhrystone contains some very short procedures where execution would benefit from procedure merging (inlining, macro expansion of procedures), procedure merging is not to be used. The reason is that the percentage of procedure and function calls is part of the "Dhrystone distribution" of statements contained in [1]. This restriction does not hold for the string functions of the C version since ANSI C allows an implementation to use inline code for these functions. o Other optimizations are allowed, but they should be indicated It is often hard to draw an exact line between "normal code generation" and "optimization" in compilers: Some compilers perform operations by default that are invoked in other compilers only when optimization is explicitly requested. Also, we cannot avoid that in benchmarking people try to achieve results that look as good as possible. Therefore, optimizations performed by compilers - other than those listed above - are not forbidden when Dhrystone execution times are measured. Dhrystone is not intended to be non-optimizable but is intended to be similarly optimizable as normal programs. For example, there are several places in Dhrystone where performance benefits from optimizations like common subexpression elimination, value propagation etc., but normal programs usually also benefit from these optimizations. Therefore, no effort was made to artificially prevent such optimizations. However, measurement reports should indicate which compiler optimization levels have been used, and reporting results with different levels of compiler optimization for the same hardware is encouraged. o Default results are those without "register" declarations (C version) When Dhrystone results are quoted without additional qualification, they should be understood as results obtained without use of the "register" attribute. Good compilers should be able to make good use of registers even without explicit register declarations ([3], p. 193). Of course, for experimental purposes, post-linkage optimization, procedure merging and/or compilation in one unit can be done to determine their effects. However, Dhrystone numbers obtained under these conditions should be explicitly marked as such; "normal" Dhrystone results should be understood as results obtained following the ground rules listed above. In any case, for serious performance evaluation, users are advised to ask for code listings and to check them carefully. In this way, when results for different systems are compared, the reader can get a feeling how much performance difference is due to compiler optimization and how much is due to hardware speed. 6. Acknowledgements The C version 2.1 of Dhrystone has been developed in cooperation with Rick Richardson (Tinton Falls, NJ), it incorporates many ideas from the "Version 1.1" distributed previously by him over the UNIX network Usenet. Through his activity with Usenet, Rick Richardson has made a very valuable contribution to the dissemination of the benchmark. I also thank Chaim Benedelac (National Semiconductor), David Ditzel (SUN), Earl Killian and John Mashey (MIPS), Alan Smith and Rafael Saavedra-Barrera (UC at Berkeley) for their help with comments on earlier versions of the benchmark. 7. Bibliography [1] Reinhold P. Weicker: Dhrystone: A Synthetic Systems Programming Benchmark. Communications of the ACM 27, 10 (Oct. 1984), 1013-1030 [2] Rick Richardson: Dhrystone 1.1 Benchmark Summary (and Program Text) Informal Distribution via "Usenet", Last Version Known to me: Sept. 21, 1987 [3] Brian W. Kernighan and Dennis M. Ritchie: The C Programming Language. Prentice-Hall, Englewood Cliffs (NJ) 1978