Dynamic temporal optimization framework

[0001] The present invention relates to temporal profiling and memory access optimization of computer programs, and particularly for dynamic optimization during program execution.

[0002] With processor speed increasing much more rapidly than memory access speed, there is a growing performance gap between processor and memory in computers. More particularly, processor speed continues to adhere to Moore's law (approximately doubling every 18 months). By comparison memory access speed has been increasing at the relatively glacial rate of 10% per year. Consequently, there is a rapidly growing processor-memory performance gap. Computer architects have tried to mitigate the performance impact of this imbalance with small high-speed cache memories that store recently accessed data. This solution is effective only if most of the data referenced by a program is available in the cache. Unfortunately, many general-purpose programs, which use dynamic, pointer-based data structures, often suffer from high cache miss rates, and therefore are limited by memory system performance.

[0003] Due to the increasing processor-memory performance gap, memory system optimizations have the potential to significantly improve program performance. One such optimization involves prefetching data ahead of its use by the program, which has the potential of alleviating the processor-memory performance gap by overlapping long latency memory accesses with useful computation. Successful prefetching is accurate (i.e., correctly anticipates the data objects that will be accessed in the future) and timely (fetching the data early enough so that it is available in the cache when required). For example, T. Mowry, M. Lam and A Gupta, Design And Analysis Of A Compiler Algorithm For Prefetching, Architectural Support For Programming Languages And Operating Systems (ASP-LOS) (1992) describe an automatic prefetching technique for scientific codes that access dense arrays in tightly nested loops, which relies on static compiler analyses to predict the program's data accesses and insert prefetch instructions at appropriate program points. However, the reference pattern of general-purpose programs, which use dynamic, pointer-based data structures, is much more complex, and the same techniques do not apply.

[0004] An alternative to static analyses for predicting data access patterns is to perform program data reference profiling. Recent research has shown that programs possess a small number of hot data streams, which are data reference sequences that frequently repeat in the same order, and these account for around 90% of a program's data references and more than 80% of cache misses. (See, e.g., T. M. Chilimbi, Efficient Representations And Abstractions For Quantifying And Exploiting Data Reference Locality, Proceedings Of The ACM SIGPLAN '01 Conference On Programming Language Design And Implementation (June 2001); and S. Rubin, R. Bodik and T. Chilimbi, An Efficient Profile-Analysis Framework For Data-Layout Optimizations, Principles Of Programming Languages, POPL'02 (January 2002).) These hot data streams can be prefetched accurately since they repeat frequently in the same order and thus are predictable. They are long enough (15-20 object references on average) so that they can be prefetched ahead of use in a timely manner.

[0005] In prior work, Chilimbi instrumented a program to collect the trace of its data memory references; then used a compression technique called Sequitur to process the trace off-line and extract hot data streams. (See, T. M. Chilimbi, Efficient Representations And Abstractions For Quantifying And Exploiting Data Reference Locality, Proceedings Of The ACM SIGPLAN '01 Conference On Programming Language Design And Implementation (June 2001).) Chilimbi further demonstrated that these hot data streams are fairly stable across program inputs and could serve as the basis for an off-line static prefetching scheme. (See, T. M. Chilimbi, On The Stability Of Temporal Data Reference Profiles, International Conference On Parallel Architectures And Compilation Techniques (PACT) (2001).) However, this off-line static prefetching scheme may not be appropriate for programs with distinct phase behavior.

[0006] Dynamic optimization uses profile information from the current execution of a program to decide what and how to optimize. This can provide an advantage over static and even feedback-directed optimization, such as in the case of the programs with distinct phase behavior. On the other hand, dynamic optimization must be more concerned with the profiling overhead, since the slow-down from profiling has to be recovered by the speed-up from optimization.

[0007] One common way to reduce the overhead of profiling is through use of sampling: instead of recording all the information that may be useful for optimization, sample a small, but representative fraction of it. In a typical example, sampling counts the frequency of individual events such as calls or loads. (See, J. Anderson et al., Continuous Profiling: Where Have All The Cycles Gone, ACM Transactions On Computer Systems (TOCS) (1997).) Other dynamic optimizations exploit causality between two or more events. One example is prefetching with Markov-predictors using pairs of data accesses. (See, D. Joseph and D. Grunwald, Prefetching Using Markov Predictors, International Symposium On Computer Architecture (ISCA) (1997).) Some recent transparent native code optimizers focus on single-entry, multiple-exit code regions. (See, e.g., V. Bala, E. Duesterwald and S. Banerjia, Dynamo: A Transparent Dynamic Optimization System, Programming Languages Design And Implementation (PLDI) (2000); and D. Deaver, R. Gorton and N. Rubin, Wiggins/Redstone: An On-Line Program Specializer, Hot Chips (1999).) Another example provides cache-conscious data placement during generational garbage collection to lay out sequences of data objects. (See, T. Chilimbi, B. Davidson and J. Larus, Cache-Conscious Structure Definition, Programming Languages Design And Implementation (PLDI) (1999); and T. Chilimbi and J. Larus, Using Generational Garbage Collection To Implement Cache-Conscious Data Placement, International Symposium On Memory Management (ISMM) (1998).) However, for lack of low-overhead temporal profilers, these systems usually employ event profilers. But, as Ball and Larus point out, event (node or edge) profiling may misidentify frequencies of event sequences. (See, T. Ball and J. Larus, Efficient Path Profiling, International Symposium On Microarchitecture (MICRO) (1996).)

[0008] The sequence of all events occurring during execution of a program is generally referred to as the trace. A burst on the other hand is a subsequence of the trace. Arnold and Ryder present a framework that samples bursts. (See, M. Arnold and B. Ryder, A Framework For Reducing The Cost Of Instrumented Code, Programming Languages Design And Implementation (PLDI) (2001).) In their framework, the code of each procedure is duplicated. (Id., at FIG. 2.) Both versions of the code contain the original instructions, but only one version is instrumented to also collect profile information. The other version only contains checks at procedure entries and loop back-edges that decrement a counter nCheck, which is initialized to nCheck₀. Most of the time, the (non-instrumented) checking code is executed. Only when the nCheck counter reaches zero, a single intraprocedural acyclic path of the instrumented code is executed and nCheck is reset to nCheck₀.

[0009] A limitation of the Arnold-Ryder framework is that it stays in the instrumented code only for the time between two checks. Since it has checks at every procedure entry and loop back-edge, the framework captures a burst of only one acyclic intraprocedural path's worth of trace. In other words, only the burst between the procedure entry check and a next loop back-edge is captured. This limitation can fail to profile many longer hot data stream bursts, and thus fail to optimize such hot data streams. Consider for example the code fragment:

[0010] for (i0; i<n; i)

[0011] if ( . . . ) f( );

[0012] else g( );

[0013] Because the Arnold-Ryder framework ends burst profiling at loop back-edges, the framework would be unable to distinguish the traces fgfgfgfg and ffffgggg. For optimizing single-entry multiple-exit regions of programs, this profiling limitation may make the difference between executing optimized code most of the time or not.

[0014] Another limitation of the Arnold-Ryder framework is that the overhead of the framework can still be too high for dynamic optimization of machine executable code binaries. The Arnold-Ryder framework was implemented for a Java virtual machine execution environment, where the program is a set of Java class files. These Java programs typically have a higher execution overhead, so that the overhead of the instrumentation checks is smaller compared to a relatively slow executing program. The overhead of the Arnold-Ryder framework's instrumentation checks may make dynamic optimization with the framework impractical in other settings for programs with lower execution overhead (such as statically compiled machine code programs).

[0015] A further problem is that the overhead of hot data stream detection has been overly high for use in dynamic optimization systems, such as the Arnold-Ryder framework.

[0016] Techniques described herein provide low-overhead temporal profiling and analysis, such as for use in dynamic memory access optimization.

[0017] In accordance with one technique described herein, temporal profiling of longer bursts in a program trace is achieved by incorporating symmetric checking code and instrumented code counters in a temporal profiling framework employing non-instrumented (checking) code and instrumented code versions of a program. Rather than immediately transitioning back to the checking code at a next proximate check in the instrumented code as in the prior Arnold-Ryder framework, a counter also is placed on checks in the instrumented code. After transitioning to the instrumented code, a count of plural checks in the instrumented code is made before returning to the checking code. This permits the instrumented code to profile longer continuous bursts sampled out of the program trace.

[0018] In accordance with further techniques, the overhead of temporal profiling is reduced by intelligently eliminating checks. In the prior Arnold-Ryder framework, checks were placed at all procedure entries and loop back-edges in the code to ensure that the program can never loop or recurse for an unbounded amount of time without executing a check. The techniques intelligently eliminate checks from procedure entries and loop back-edges. In one implementation, the intelligent check elimination performs a static call graph analysis of the program to determine where checks should be placed on procedure entries to avoid unbounded execution without checking. Based on the call graph analysis, the intelligent check elimination places checks at entries to root procedures, procedures whose address is taken, and procedures with recursion from below. On the other hand, the intelligent check elimination does not place checks on leaf procedures (that call no other code in the program) in the call graph. Further, the intelligent check elimination eliminates checks at loop back-edges of tight inner loops, and at k-boring loops (loops with no calls and at most k profiling events of interest, since these are easy for a compiler to statically optimize). Other techniques to reduce checks also can be employed. This reduction in temporal profiling overhead can make dynamic optimization practical for faster executing programs (e.g., binary code), as well as improving efficiency of dynamic optimization of just-in-time compiled (JITed) code and interpreted programs.

[0019] In accordance with another technique, an improved hot data stream detection more quickly identifies hot data streams from profiled bursts of a program, which can make dynamic prefetching practical for dynamic optimization of programs. In one implementation, the improved hot data stream detection constructs a parse tree of the profiled bursts, then forms a Sequitur grammar from the parse tree. The improved hot stream detection then traverses the grammar tree in reverse postorder numbering order. At each grammar element, the improved hot stream detection calculates a regularity magnitude or heat of the element based on a length of the burst sequence represented by the element multiplied by its number of cold uses (i.e., number of times the element occurs in the complete parse tree, not counting occurrences as sub-trees of another hot element). The improved hot stream detection identifies elements as representing hot data streams if their heat exceeds a heat threshold.

[0020] Additional features and advantages of the invention will be made apparent from the following detailed description that proceeds with reference to the accompanying drawings.

[0021] FIG. 1 is a data flow diagram of a dynamic optimizer utilizing a low overhead, long burst temporal profiling framework and fast hot data stream detection to dynamically optimize a program with dynamic hot data stream prefetching.

[0022] FIG. 2 is a block diagram of a program modified according to the prior Arnold-Ryder framework for burst profiling.

[0023] FIG. 3 is a block diagram of a program modified according to an improved framework for longer burst profiling in the dynamic optimizer of FIG. 1.

[0024] FIG. 4 is a program code listing for a check to control transitions between checking and instrumented code versions in the improved framework of FIG. 3 for longer burst profiling.

[0025] FIG. 5 is a call graph of an example program to be modified according to an improved framework for low-overhead burst profiling.

[0026] FIG. 6 is an illustration of an analysis of the call graph of FIG. 5 for modifying the example program according to the improved framework for low-overhead burst profiling.

[0027] FIG. 7 is a data flow diagram illustrating processing for dynamic optimization of a program image in the dynamic optimizer of FIG. 1.

[0028] FIG. 8 is a timeline showing phases of the low-overhead, long burst temporal profiling by the dynamic optimizer of FIG. 1.

[0029] FIG. 9 is an illustration of grammar analysis of an exemplary data reference sequence in bursts profiled with the low-overhead, long burst temporal profiling forming part of the processing by the dynamic optimizer shown in FIG. 7.

[0030] FIG. 10 is a program code listing for fast hot data stream detection in the processing by the dynamic optimizer shown in FIG. 7.

[0031] FIG. 11 is an illustration of the fast hot data stream detection performed according to the program code listing of FIG. 10 on the grammar of the exemplary data reference sequence from FIG. 9.

[0032] FIG. 12 is a table listing results of the fast hot data stream detection illustrated in FIG. 11.

[0033] FIG. 13 is a block diagram of a suitable computing device environment for devices in the network device architecture of FIG. 1.

[0034] The following description is directed to techniques for low-overhead, long burst temporal profiling and fast hot data stream detection, which can be utilized in dynamic optimization of computer programs. More particularly, these technique are described in their particular application to a dynamic optimization involving hot data stream prefetching to optimize a program's memory accesses. However, the techniques can be applied in contexts other than the described hot data stream prefetching dynamic optimization.

[0035] 1. Overview of Dynamic Optimizer

[0036] With reference to FIG. 1, an exemplary dynamic optimizer 100 utilizes techniques described more fully herein below for low-overhead, long burst temporal profiling and fast hot data stream detection in a process of dynamically optimizing a computer program. The exemplary dynamic optimizer 120 includes a program editing tool 122 to build a program image 130 in accordance with a low-overhead temporal profiling framework described below, including inserting instrumentation and checking code for profiling long burst samples of a trace of the program's execution. In the exemplary dynamic optimizer, the program editing tool 122 inserts the instrumentation and checking code for the low-overhead temporal profiling framework by editing an executable or binary version 115 of the program to be optimized, after compiling and linking by a conventional compiler from the program's source code version. For example, the source code 105 of the program to be optimized may be initially written by a programmer in a high level programming language, such as C or C. Such program source code is then compiled using an appropriate conventional compiler 110, such as a C/C compiler available in the Microsoft Visual Studio development platform, to produce the machine-executable program binary 115. The executable editing tool for the instrumentation insertion 122 can be the Vulcan executable editing tool for x86 computer platform program binaries, which is described in detail by A. Srivastava, A. Edwards, and H. Vo, Vulcan: Binary Transformation In A Distributed Environment, Technical Report MSR-TR-2001-50, Microsoft Research (2001). This has the advantage that the dynamic optimizer does not require access to the source code, and can employed to optimize programs where only an executable binary version is available. In other embodiments, the profiling framework can be built into the program image 130 as part of the process of compiling the program from source code or an intermediate language form, such as for use with programs written in Java, or intermediate code representations for the Microsoft .Net platform. In such other embodiments, the compiler that inserts instrumentation and checks embodies the tool 122.

[0037] The temporal profiling framework provided in the program image 130 produces profiled burst data 135 representing sampled bursts of the program's execution trace. The exemplary dynamic optimizer 120 includes a hot data stream analyzer 140 and hot stream prefetching code injection tool 142. The hot data stream analyzer 140 implements fast hot data stream detection described herein below that process the profiled burst data to identify hot data streams, which are frequently recurring sequences of data accesses by the program. The hot stream prefetching code injection tool 142 then dynamically modifies the program image 130 to perform prefetching so as to optimize cache utilization and data accesses by the program, based on the identified hot data streams.

[0038] 2. Temporal Profiling Framework

[0039] The program image 130 (FIG. 1) is structured according to a low-overhead, long burst temporal profiling framework 300 illustrated in FIG. 3, which is an improvement on the prior Arnold-Ryder framework 200 (FIG. 2).

[0040] In the prior Arnold-Ryder framework 200, the code of each procedure from an original program version (e.g., original procedure 210 with code blocks 212-213) is duplicated. Both duplicate versions of the code in the framework 200 contain the original instructions, but only one version is instrumented to also collect profile information (referred to herein as the instrumented code 220). The other version (referred to herein as the checking code 230) only contains checks 240-241 at procedure entries and loop back-edges that decrement a counter nCheck, which is initialized to nCheck₀. Most of the time, the (non-instrumented) checking code 230 is executed. Only when the nCheck counter reaches zero, a single intraprocedural acyclic path of the instrumented code 220 is executed and nCheck is reset to nCheck₀. All back-edges 250 in the instrumented code 220 transition back to the checking code 230.

[0041] While executing in the instrumented code 220, the Arnold-Ryder framework 200 profiles a burst out of the program execution trace, which begins at a check (e.g., procedure entry check 240 or back-edge check 241) and extends to the next check. In other words, the profiling captures one intraprocedural acyclic path. The profile of the program captured during execution of this path can be, for example, the data accesses made by the program.

[0042] Profiling Longer Bursts

[0043] The improved framework 300 extends the prior Arnold-Ryder framework 200 (FIG. 2) so that profiled bursts can extend over multiple checks, possibly crossing procedure boundaries. This way, the improved framework can obtain interprocedural, context-sensitive and flow-sensitive profiling information.

[0044] As in the Arnold-Ryder framework 200, the improved framework 300 is structured to include duplicate non-instrumented (checking code) 330 and instrumented code 320 versions of at least some original procedures 310 of the program. Further, checks 340-341 are placed at procedure entry and loop back-edges.

[0045] The extension in the improved framework 300 adds a second profiling phase counter (labeled nlnstr) to make execution flow in the instrumented code 320 symmetric with the checking code 330. Further, the loop back-edges 350 from the instrumented code 320 do not transition directly back to the procedure entry as in the prior Arnold-Ryder framework 200, but instead go to a back-edge check 341.

[0046] The program logic or code 400 for the checks 340-341 is shown in FIG. 4. Initially, the value of the checking phase counter (nCheck) is set to its initial value, nCheck₀. While in the checking code 400, the framework 300 decrements the checking phase counter (nCheck) (statement 410) at every check 340-341. The framework 300 continues to execute in the checking code (statement 420) as long as the value of the checking phase counter has not yet reached zero. For example, from the entry and back-edge checks 340-341, the framework 300 takes the paths 360-361 to the checking code 330.

[0047] When the checking phase counter (nCheck) reaches zero, the framework 300 initializes the profiling phase counter (nlnstr) to an initial value, nlnstr₀, and transitions to the instrumented code 320 (statement 430). In general, the checking phase counter's initial value is selected to be much greater than that of the profiling phase counter (i.e., nlnstr₀<<nCheck₀), which determines the sampling rate of the framework ( rnlnstr₀(nCheck₀nlnstr₀)).

[0048] While executing in the instrumented code, the framework 300 decrements the profiling phase counter (nlnstr) at every check 340-341 (statement 440). The framework 300 continues to execute in the instrumented code (statement 450) as long as the value of the profiling phase counter has not yet reached zero. For example, from the entry and back-edge checks 340-341, the framework 300 takes the paths 370-371 to the instrumented code 320. When the profiling phase counter reaches zero, the framework again initializes the checking phase counter to the initial value, nCheck₀, and returns to the checking code 330 (statement 460).

[0049] The check code 400 is structured so that in the common case where the framework is executing in the checking code and is to continue executing the checking code (checking phase), the check consists of a decrement of the checking phase counter and a conditional branch.

[0050] Compared to the prior Arnold-Ryder framework 200, the improved framework 300 profiles longer bursts of the program trace and provides more precise profiles. For example, consider the following code fragment:

[0051] for (i0; i<n; i)

[0052] if ( . . . ) f( );

[0053] else g( );

[0054] In this example code fragment, the Arnold-Ryder framework returns to the checking code upon the back-edge path from each execution of the procedures, f( ) and g( ). Accordingly, the Arnold-Ryder framework profiles only on acyclic intraprocedural path of the program trace, and would be unable to distinguish the traces, fgfgfgfg and fffgggg. The improved framework 300 profiles longer bursts across procedure boundaries. In the dynamic optimizer 120 (FIG. 1), this can make a difference between executing optimized code most of the time or not.

[0055] Low-overhead Temporal Profiling

[0056] For the dynamic optimization to effectively enhance the performance of the program, the overhead imposed by the temporal profiling framework desirably is relatively small compared to the overall program execution, so that performance gains are achieved from dynamically optimizing the program. The overhead of the temporal profiling framework can be particularly significant in the exemplary dynamic optimizer 120 in which the program image 130 is built from editing an executable program binary 115, to which the compiler 110 has already applied many static optimizations. In such case, the overhead of the prior Arnold-Ryder framework may be too high for effective dynamic optimization. The prior Arnold-Ryder framework has checks at all procedure entries and loop back-edges to insure that the program can never loop or recurse for an unbounded amount of time without executing a check. Otherwise, sampling could miss too much profiling information (when the program spends an unbounded amount of time in the checking code), or the overhead could become too high (when the program spends an unbounded amount of time in the instrumented code).

[0057] The low-overhead temporal profiling framework described herein decreases the overhead of the burst sampling by intelligently eliminating some checks (i.e., placing checks at fewer than all procedure entries and loop back-edges), while still ensuring that the program does not spend an unbounded amount of time without executing a check.

[0058] Eliminating Checks at Procedure Entries

[0059] In the low-overhead temporal profiling framework, the instrumentation tool 122 places checks at an approximated minimum set of procedure entries so that the program cannot recurse for an unbounded amount of time without executing a check. The instrumentation tool 122 performs a static call graph analysis of the program 115 to determine this approximate minimum set (CN) of nodes in the program's call graph, such that every cycle in the call graph contains at least one node of the set.

[0060] In the dynamic optimizer 120, the instrumentation tool 122 selects this set (CN) of procedures f at which to place procedure entry checks, according to the criteria represented in the following expression: $C=\{fN|\begin{array}{c}\mathrm{is\_leaf}\left(f\right)\\ \left(\mathrm{is\_root}\left(f\right)\mathrm{addr\_taken}\left(f\right)\mathrm{recursion\_from}\mathrm{\_below}\left(f\right)\right)\}\end{array}$

[0061] In accordance with this criteria, the instrumentation tool 122 does not place any check on any entry to a leaf procedure (i.e., a procedure that calls nothing), since such leaf procedures cannot be part of a recursive cycle. Otherwise, the instrumentation tool 122 places a check on entries to all root procedures (i.e., procedures that are only called from outside the program), so as to ensure that execution starts in the correct version of the code. Also, the tool places a check on entry to every procedure whose address is taken, since such procedures may be part of recursion with indirect calls. Further, the tool places a check on entry to every procedure with recursion from below. A procedure f has recursion from below, iff it is called by g in the same strongly connected component as f that is at least as far away from the roots. The distance of a procedure f from the roots is the minimum length of the shortest path from a root to f.

[0062] The recursion_from_below heuristic in this criteria guarantees that there is no recursive cycle without a check and breaks the ties to determine where in the cycle to put the check (similarly to back-edges in loops). The tool breaks ties so that checks are as far up in the call-stack as possible. This should reduce the number of dynamic checks.

[0063] For example, FIG. 5 illustrates a call graph 500 of an exemplary program being structured by the tool 122 according to the low-overhead temporal profiling framework. In this call graph 500, the only root is procedure main 510, and the only leaf procedure is delete-digram 520. The only non-trivial strongly connected component in the call graph 500 is the component 650 (of procedures check, match, substitute 530-532).

[0064] FIG. 6 illustrates an analysis 600 of the call graph 500 by the tool 122 to determine the set of procedures for entry check placement. For this analysis, the tool 122 begins with a breadth-first search of the call graph. The tool calculates the distances (e.g., from 0 to 4 in this example) of each procedure from the root procedure (main 510), and determines that only the procedure check 530 has recursion from below, since it is called from the procedure substitute 532 which is further away from the root procedure main 510. The tool 122 thus determines that for this example with call graph 500, only the procedures main 510 and check 530 meet the above criteria for placing an entry check (i.e., the above expression evaluates to the minimum set Cmain,check for this call graph). Accordingly, by placing a check on entry to every procedure in this minimum set Cmain,check, the program cannot recurse indefinitely without executing checks.

[0065] Eliminating Checks at Loop Back-Edges

[0066] In the low-overhead temporal profiling framework, the instrumentation tool 122 also places checks at fewer than all loop back-edges in the program. In particular, the instrumentation tool 122 eliminates checks for some tight inner loops. This is because a dynamic optimizer that complements a static optimizer may often find the profiling information from tight inner loops to be of little interest because static optimization excels at optimizing such loops. At the same time, checks at the back-edges of tight inner loops can become extremely expensive (i.e., create excessive overhead relative to potential optimization performance gain). With the dynamic optimizer 100 that prefetches data into cache memory based on hot data streams, loops that compare or copy arrays preferably should not have checks. Such loops typically are easy to optimize statically, the check on the back-edge is almost as expensive as the loop body, and the loop body contains too little work to overlap with the prefetch.

[0067] More particularly, the instrumentation tool 122 eliminates checks on loop back-edges of loops meeting a k-boring loops criteria. According to this criteria, k-boring loops are defined as loops with no calls and at most a number (k) of profiling events of interest. The instrumentation tool 122 does not instrument either version of the code of a k-boring loop, and does not place a check on its back-edge. Since the loop is not included in the instrumented code 320 (FIG. 1) version, the program image 130 does not spend an unbounded amount of time executing in instrumented code. The program image may spend an unbounded amount of time executing such a loop in uninstrumented code (checking code 330 of FIG. 1) without executing a check. But, if the k-boring loop hypothesis holds (i.e., there is little or no gain from optimizing such loops with hot data stream prefetching), the dynamic optimizer 120 does not miss interesting profiling information. Experiments have shown that the quality of the profile actually improved when instrumenting of back-edge checks were eliminated from 4-boring loops (i.e., k4) in an experimental program image, where the quality of the profile is measured by the ability to detect hot data streams. Accordingly, eliminating k-boring loop from profiling helps focus sampling on more interesting events (for optimizing with hot data stream prefetching).

[0068] In alternative implementations, the instrumentation tool 122 may eliminate additional checks on loop back-edges. For example, the instrumentation tool may eliminate back-edge checks from a loop that has only a small, fixed number of iterations. Further, if a check is always executing within a loop body, the loop does not need a check on the loop's back-edge. In yet further alternative implementations, the instrumentation tool 122 can combine the loop counter with the profiling phase counter; if the counters are linearly related, the program image can execute checks for the loop via a predicate on the loop counter, rather than updating the profiling counter each iteration of the loop.

[0069] 3. Hot Data Stream Prefetching

[0070] With reference now to FIG. 7, the temporal profiling 710 using the above-described low-overhead, long burst temporal profiling framework 300 (FIG. 3) is a first phase in an overall dynamic optimization process 700 based on hot data stream prefetching. The dynamic optimization process 700 operates in three phasesprofiling 710, analysis and optimization 720, and hibernation 730. First, the profiling phase collects (740) a temporal data reference profile 135 from a running program with low-overhead, which is accomplished using the program image 130 (FIG. 1) structured according to the improved temporal profiling framework 300. As described in more detail below, a grammar analysis using the Sequitur compression process 750 incrementally builds an online grammar representation 900 of the traced data references.

[0071] Once sufficient data references have been traced, profiling is turned off, and the analysis and optimization phase 720 commences. First, a fast hot data stream detection 140 extracts hot data streams 760 from the Sequitur grammar representation 900. Then, a prefetching engine 142 builds a stream prefix matching deterministic finite state machine (DFSM) 770 for these hot data streams, and dynamically injects checks at appropriate program points to detect and prefetch these hot data streams in the program image 130. This dynamic prefetching based on a DFSM is described in more detail in co-pending Chilimbi, T., DYNAMIC PREFETCHING OF HOT DATA STREAMS, patent application Ser. No. ______ filed on ______, which is hereby incorporated herein by reference.

[0072] Finally, the process enters the hibernation phase where no profiling or analysis is performed, and the program continues to execute (780) as optimized with the added prefetch instructions. At the end of the hibernation phase, the program image 130 is de-optimized (790) to remove the inserted checks and prefetch instructions, and control returns to the profiling phase 710. For long-running programs, this profiling 710, analysis and optimization 720 and hibernate 730 cycle may repeat multiple times.

[0073] FIG. 8 shows a timeline 800 for the three phase profiling, analysis and optimization, and hibernation cycle operation of the dynamic optimizer 100 (FIG. 1). As discussed above, the low-overhead, long burst temporal profiling framework uses the checking phase and profiling phase counters (nCheck, nlnstr) to control its overhead and sampling rate of profiling, by transitioning between a checking phase 810 in which the program image 130 (FIG. 1) executes in its non-instrumented checking code 330 (FIG. 3) and a profiling phase 820 in which it executes in its instrumented code 320 (FIG. 3). The time periods for these checking and profiling phase are parameterized by the nCheck₀ and nlnstr₀ counter initialization values. For example, setting nCheck₀ to 9900 and nlnstr₀ to 100 results in a sampling rate of profiling of 100/100001% and a burst length of 100 dynamic checks. The time spent for one iteration of the checking and profiling phase (nCheck₀nlnstr₀) is referred to as a burst period 850.

[0074] For dynamic optimization, the above-described low-overhead temporal profiling framework 300 (FIG. 3) is further extended to alternate between two additional phases, awake 830 and hibernating 840, which are controlled via two additional (awake and hibernating) counters. The temporal profiling framework starts out in the awake phase 830, and continues operating in the awake phase for a number (nAwake₀) of burst-periods, yielding (nAwake₀nlnstr₀) checks (860) worth of traced data references (or bursts). Then, as described above and illustrated in FIG. 7, the dynamic optimizer 100 performs the optimizations, and then the profiler hibernates while the optimized program executes. This is done by setting nCheck₀ to (nCheck₀nlnstr₀1) and nlnstr₀ to 1 for the next nHibernate₀ burst-periods (which causes the check code 400 in FIG. 4 to keep the program image executing in the non-instrumented checking code 330), where nHibernate₀ is much greater than nAwake₀. When the hibernating phase 840 is over, the profiling framework is woken up by resetting nCheck₀ and nlnstr₀ to their original values.

[0075] While the profiling framework is hibernating, the program image traces next to no data references and hence incurs only the basic overhead of executing the checks 400 (FIG. 4). With the values of nCheck₀ and nlnstr₀ set as described above during hibernation, the burst-periods correspond to the same time (in executed checks 860) in both awake and hibernating phases. This facilitates control over the relative length of the awake and hibernating phases by appropriately setting the initial value parameters nAwake₀ and nHibernate₀ of the awake and hibernating counters relative to each other.

[0076] Fast Hot Data Stream Detection

[0077] When the temporal profiling framework 300 executes in the instrumented code 320 (FIG. 3), the temporal profiling instrumentation produces data reference bursts or temporal data reference sequences 135 (FIGS. 1 and 7). A data reference r is a load or store operation on a particular address, represented in the exemplary dynamic optimizer 120 as a data pair (r.pc,r.addr). The pc value (i.e., r.pc), is the value of the program counter, which indicates the address in the executing program of the data load or store instruction being executed. The addr value (i.e., r.addr), is the memory location accessed by the load or store operation. The profiled burst is a temporal sequence or stream of these data references.

[0078] During the profiling phase 710 (FIG. 7) as discussed above, this data reference sequence is incrementally processed into a compressed Sequitur grammar representation 900 using the Sequitur grammar analysis processing, as described in T. M. Chilimbi, Efficient Representations And Abstractions For Quantifying And Exploiting Data Reference Locality, Proceedings Of The ACM SIGPLAN '01 Conference On Programming Language Design And Implementation (June 2001). FIG. 9 illustrates an example of a grammar 900 produced from an input data reference sequence (input string 910). The grammar 900 represents a hierarchical structure (a directed acyclic graph 920) of the data references.

[0079] More particularly, each observed data reference (r.pc,r.addr) is conceptually represented as a symbol in a grammar, and the concatenation of the profiled bursts is a string w of symbols (910). The Sequitur grammar analysis constructs a context-free grammar for the language w consisting of exactly one word, the string w. The Sequitur grammar analysis runs in time O(w.length). It is incremental (one symbol can be appended at a time), and deterministic. Thus, the grammar analysis can be performed as the profiled data is sampled during the profiling phase 710 (FIG. 7). The grammar 900 is a compressed representation of the input burst 910. Further, it is unambiguous and acyclic in the sense that no non-terminal directly or indirectly defines itself.

[0080] In the Sequitur grammar 910, the terminal nodes (denoted in small case letters) represent individual data references (r.pc,r.addr), which may be repeated in the profiled burst. The intermediate nodes (denoted in capital letters) represent temporal sequences of the data references. For example, the grammar 910 produced from the example input string 910 shows that the string S consists of the sequence AaBB. A, in turn, consists of the data references a and b. The intermediate node B represents a sequence with two occurrences of the intermediate node C, which is a sequence of the intermediate node A and data reference c.

[0081] After construction of the grammar 900 in the profiling phase 710, the dynamic optimizer 100 performs a fast hot data stream detection 140 (FIGS. 1 and 7) to identify frequently recurring data reference subsequences (the hot data streams) in the profiled bursts. For the fast hot data stream detection, the exemplary dynamic optimizer performs analysis of the grammar as represented in a hot data stream detection code 1000 shown in FIG. 10. The purposes of the fast hot data stream analysis is to identify hot data streams, which are a data reference subsequence in the profiled bursts, whose regularity magnitude exceeds a predetermined heat threshold, H. The regularity magnitude, given a data reference subsequence v, is defined as v.heatv.length*v.frequency, where v.frequency is the number of non-overlapping occurrences of v in the profiled bursts.

[0082] The analysis in code 1000 is based on the observation that each non-terminal node (A) of a Sequitur grammar generates a language L(A)w_A with just one word w_A.

[0083] For the fast hot data stream detection analysis, the regularity magnitude of a non-terminal A is defined instead as A.heatw_A.length*A.coldUses, where A.coldUses is the number of times A occurs in the (unique) parse tree of the complete grammar, not counting occurrences in sub-trees belonging to hot non-terminals other than A. A non-terminal A is hot iff minLen<A.length<maxlen and H<A.heat, where H is the predetermined heat threshold. The result of the analysis is the set w_AA is a hot non-terminal of hot data streams.

[0084] FIGS. 11 and 12 show an example 1100 of the analysis in the code 1000 (FIG. 10) for the input data reference sequence 910 and grammar 900 in FIG. 9. As a result of the Sequitur grammar analysis 750 (FIG. 7), the input data reference sequence has been parsed (as shown by parse tree 1110) and sub-sequences grouped under intermediate (non-terminal) nodes into the Sequitur grammar (1120). Further, the Sequitur grammar analysis also yields the length of the subsequence represented in each non-terminal node of the grammar 1120. Accordingly, the information shown in the first three columns (the non-terminal nodes, their children, and their lengths) of the table 1200 is provided to the fast hot data stream detection analysis. As shown in FIGS. 11 and 12, a non-terminal node is considered the child of another non-terminal node if it is listed on the right-hand side of the grammar rule of the other non-terminal node in Sequitur grammar 900 (FIG. 9).

[0085] In the fast hot data stream analysis code 1000, the analyzer 140 (FIG. 1) first executes instructions (1010) to perform a reverse post-order numbering of the non-terminal nodes in the grammar. For the example grammar, the results in numbering the nodes S, A, B, and C as 0, 3, 1 and 2, respectively, as shown in the index column of the table 1200 (FIG. 12) and illustrated in the reverse postorder numbering tree 1130 (FIG. 11). This results in the non-terminal nodes being numbered such that whenever a non-terminal node (e.g., node C) is a child of another non-terminal node (e.g., B), the number of the child node is greater (e.g., B.index<C.index). This property guarantees that the analysis does not visit a non-terminal node before having visited all its predecessors.

[0086] The analyzer 140 next determines at instructions 1020 in code 1000 how often each non-terminal node occurs in the parse-tree 1110 (FIG. 11), which is represented in the use column of the table 1200 (FIG. 12). Each of the non-terminal nodes is now associated with two values, its number of hot uses and its length, which are depicted conceptually in the uses:length tree 1140 (FIG. 11).

[0087] Finally, the analyzer 140 finds the number of cold uses for each non-terminal node, which are the number of hot uses not attributable in the cold uses of a hot predecessor node. More specifically, the analyzer finds hot non-terminal nodes such that a non-terminal node is only considered hot if it accounts for enough of the trace on its own, where it is not part of the expansion of the other hot non-terminals. In the example grammar with a heat threshold (H8) and length restrictions (minLen2, maxLen7), only the non-terminal node B is considered as hot, since its heat (cold useslength2612) exceeds the heat threshold (12>8). All uses of the non-terminal node C are completely subsumed its the predecessor hot non-terminal node B and therefore is not considered hot (its heatcold useslength030). The non-terminal node A has a single use apart from as a subsequence of the hot non-terminal node B, but this single use is not sufficient to exceed the heat threshold (A's cold useslength133<8). The single hot non-terminal node B represents the hot data stream w_Babcabc, which accounts for 12/1580% of all data references in this example burst.

[0088] 4. Computing Environment

[0089] FIG. 13 illustrates a generalized example of a suitable computing environment 1300 in which the described techniques can be implemented. The computing environment 1300 is not intended to suggest any limitation as to scope of use or functionality of the invention, as the present invention may be implemented in diverse general-purpose or special-purpose computing environments.

[0090] With reference to FIG. 13, the computing environment 1300 includes at least one processing unit 1310 and memory 1320. In FIG. 13, this most basic configuration 1330 is included within a dashed line. The processing unit 1310 executes computer-executable instructions and may be a real or a virtual processor. In a multi-processing system, multiple processing units execute computer-executable instructions to increase processing power. The memory 1320 may be volatile memory (e.g., registers, cache, RAM), non-volatile memory (e.g., ROM, EEPROM, flash memory, etc.), or some combination of the two. The memory 1320 stores software 1380 implementing the dynamic optimizer 100 (FIG. 1).

[0091] A computing environment may have additional features. For example, the computing environment 1300 includes storage 1340, one or more input devices 1350, one or more output devices 1360, and one or more communication connections 1370. An interconnection mechanism (not shown) such as a bus, controller, or network interconnects the components of the computing environment 1300. Typically, operating system software (not shown) provides an operating environment for other software executing in the computing environment 1300, and coordinates activities of the components of the computing environment 1300.

[0092] The storage 1340 may be removable or non-removable, and includes magnetic disks, magnetic tapes or cassettes, CD-ROMs, CD-RWs, DVDs, or any other medium which can be used to store information and which can be accessed within the computing environment 1300. The storage 1340 stores instructions for the dynamic optimizer software 1380.

[0093] The input device(s) 1350 may be a touch input device such as a keyboard, mouse, pen, or trackball, a voice input device, a scanning device, or another device that provides input to the computing environment 1300. For audio, the input device(s) 1350 may be a sound card or similar device that accepts audio input in analog or digital form, or a CD-ROM reader that provides audio samples to the computing environment. The output device(s) 1360 may be a display, printer, speaker, CD-writer, or another device that provides output from the computing environment 1300.

[0094] The communication connection(s) 1370 enable communication over a communication medium to another computing entity. The communication medium conveys information such as computer-executable instructions, audio/video or other media information, or other data in a modulated data signal. A modulated data signal is a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media include wired or wireless techniques implemented with an electrical, optical, RF, infrared, acoustic, or other carrier.

[0095] The device connectivity and messaging techniques herein can be described in the general context of computer-readable media. Computer-readable media are any available media that can be accessed within a computing environment. By way of example, and not limitation, with the computing environment 1300, computer-readable media include memory 1320, storage 1340, communication media, and combinations of any of the above.

[0096] The techniques herein can be described in the general context of computer-executable instructions, such as those included in program modules, being executed in a computing environment on a target real or virtual processor. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The functionality of the program modules may be combined or split between program modules as desired in various embodiments. Computer-executable instructions for program modules may be executed within a local or distributed computing environment.

[0097] For the sake of presentation, the detailed description uses terms like determine, generate, adjust, and apply to describe computer operations in a computing environment. These terms are high-level abstractions for operations performed by a computer, and should not be confused with acts performed by a human being. The actual computer operations corresponding to these terms vary depending on implementation.

[0098] In view of the many possible embodiments to which the principles of our invention may be applied, we claim as our invention all such embodiments as may come within the scope and spirit of the following claims and equivalents thereto.