J. Anderson, L. M. Berc, J. Dean, S. Ghemawat, M. R. Henzinger, S. Leung, R. L. Sites, M. T. Vandevoorde, C. A. Waldspurger, and W. E. Weihl, "Continuous profiling: Where have all the cycles gone?," in Proceedings of the 16th ACM Symposium of Operating Systems Principles, pp. 1--14, October 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....that transparently and automatically profile and reoptimize programs. However, even when the cost of profiling is minimized, instrumentation based profiling can still incur significant overhead [6] Low overhead methods of transparent profiling have been developed based on statistical sampling [7] [8] 9] These approaches suffer from three primary drawbacks. First, the entire profile of each application must be continuously maintained at run time on the production system. Second, the profile represents only average behavior over an extended period of time. And third, the latency of ....

J. Anderson, L. M. Berc, J. Dean, S. Ghemawat, M. R. Henzinger, S. Leung, R. L. Sites, M. T. Vandevoorde, C. A. Waldspurger, and W. E. Weihl, "Continuous profiling: Where have all the cycles gone?," in Proceedings of the 16th ACM Symposium of Operating Systems Principles, pp. 1--14, October 1997.

.... by incorporating structural information about the application under study such as its call graph [4] or by pruning and prioritizing the space in which the search takes place based on the application s behavior during previous runs [13] We have developed a new technique that uses sampling [1,2,8] to augment an automated search for performance problems to attack the problem of code size scalability. Our hybrid approach, called Deep Start, substantially improves a tool s search effectiveness by locating performance problems more quickly and by locating performance problems hidden from a ....

....Deep Start 0.4 Callgraph 5 22 01 15 the prof and gprof [8] profiling tools for performing flat and call graph based profiles, respectively. Anderson and Lazowska [2] addressed the shortcomings of prof and gprof for parallel applications running on shared memory multiprocessors. Anderson et al. [1] used program counter sampling in DCPI to obtain low level information about instructions executing on in order Alpha processors. Recognizing the limitations of the DCPI approach for out oforder processors, Dean et al. [6] designed hardware support for obtaining accurate instruction profile ....

J. M. Anderson, L. M. Berc, J. Dean, S. Ghemawat, M. R. Henzinger, S. A. Leung, R. L. Sites, M. T. Vandevoorde, C. A. Waldspurger, and W. E. Weihl, "Continuous Profiling: Where Have All the Cycles Gone?" 16th ACM Symposium on Operating Systems Principles, Saint Malo, France, October 1997.

....ideas and algorithms for our feedback directed optimizations were taken from many different sources, and we discuss their relationship to prior work in Section 3. 3 Optimization Profiles are collected by either instrumenting a binary with pixie [2] or using the statistical sampling profiler DCPI [14]. Both produce a database of basic block execution counts. The compiler reads the profile database and annotates basic blocks in its IL with execution counts. It computes call edges counts and To appear in 2nd ACM Workshop on Feedback Directed Optimization, Haifa, Israel, November 15, 1999 2 ....

....maximize performance [20] and feedback directed optimization is just one part of it. The main obstacle that we have found is ease of use. Generating profiles requires additional steps, which leads to opportunities for user error. Compaq has worked on making the process as transparent as possible [23,9,14], and we expect feedback directed optimization to be even more widely used in the future. Acknowledgements Gene Albert developed the system for managing profiles, wrote new optimizations and improved others. Michael Adler, David Blickstein, Peter Craig, Caroline Davidson, Neil Faiman, Kent ....

J. Anderson et al., "Continuous Profiling: Where Have All the Cycles Gone?" Proceedings of the Sixteenth ACM Symposium on Operating System Principles, Saint-Malo, France (October 1997): 1-14.

....of performance statistics, classify this data appropriately, and perform off line analysis to construct a characterization of the system under normal behavior and detect anomalous behavior. Continuous on line profiling of operating systems with fine grain performance based feedback is described in [33], which also reports impressive performance improvements due to the continuous 8 profiling methodology adopted. In contrast, we focus exclusively on the communication subsystem, with the primary goal of parameterizing it via on line profiling and making this information available to the admission ....

Jennifer Anderson, Lance M. Berc, et al., "Continuous profiling: Where have all the cycles gone?," in Proc. ACM Symp. on Operating Systems Principles, October 1997.

....ideas and algorithms for our profile based optimizations were taken from many different sources, and we discuss their relationship to prior work in Section 3. 3. Optimizations Profiles are collected by either instrumenting a binary with pixie [2] or using the statistical sampling profiler DCPI [14]. Both produce a database of basic block execution counts. The compiler reads the profile database and annotates basic blocks in its IL with execution counts. It computes call edges counts and estimates flow edge counts from this information. This is the only profile information that is used. ....

....performance [20] and profile based optimization is just one part of it. The main obstacle that we have found is ease of use. Generating profiles requires additional steps, which leads to opportunities for user error. Compaq has worked on making the process as transparent as possible [23] 9] [14], and we expect profile based optimization to be even more widely used in the future. Acknowledgements Gene Albert developed the system for managing profiles, wrote new optimizations and improved others. Michael Adler, David Blickstein, Peter Craig, Caroline Davidson, Neil Faiman, Kent Glossop, ....

J. Anderson, L. M. Berc, J. Dean, S. Ghemawat, M. R. Henzinger, S.-T. Leung, R. L. Sites, M. T. Vandevoorde, C. A. Waldspurger, and W. E. Weihl, "Continuous Profiling: Where Have All the Cycles Gone?" Proceedings of the Sixteenth ACM Symposium on Operating System Principles, SaintMalo, France (October 1997): 1-14.

....same machine. For example, a server application may be invoked to run in a low energy cache mode during low activity periods, and run in a high performance cache mode with more cache ways enabled during high load periods. The operating system or a continuous profiling and optimization system [17, 18] could effectively control the PDT by changing the number of ways enabled under different loading conditions. In Section 4, we evaluate the energy savings that is achievable with different PDT values. In the next subsection, we address the issue of how to properly handle access to data stored in a ....

.... mgrid 2 3 3 fpppp 2 2 2 wave5 2 2 3 stereo 1 1 1 compress 0 0 0 ijpeg 0 0 1 li 0 1 1 32KB turb3d 0 0 1 2 way mgrid 0 1 1 fpppp 0 0 0 wave5 0 0 1 stereo 0 0 0 compress 0 0 0 ijpeg 1 1 1 li 0 1 1 64KB turb3d 0 0 1 2 way mgrid 1 1 1 fpppp 0 0 1 wave5 0 1 1 stereo 0 0 0 and DCPI [17] leverage the on chip performance counters found on modern processors to provide additional insight on cache behavior. Rather than simply count events, ProfileMe and DCPI can be used to gather detailed information on a per instruction basis, allowing for a more accurate assessment of cache ....

J. Anderson et al., "Continuous profiling: Where have all the cycles gone?," Proceedings of the 16th Symposium on Operating Systems Principles, October 1997.

....blending methods. 1 Introduction Many recent systems use profiles to guide optimization and translation towards a variety of goals. These systems have been striking in their ability to instrument with extremely low performance overhead (e.g. Com paq s Continuous Profiling Infrastructure [1]) their innovative application of profiling to binary translation (e.g. the FX 32 x86 to Alpha translation system [7] and their exhibition of speedups in dynamic optimization (e.g. Hewlett Packard s Dynamo dynamic optimization project [2] All of these recent systems blur the traditional ....

....distribution. Next, we use the relative entropy metric from Section 2 to choose the value of ; we look for a value of such that is equidistant from both and . We could do this directly, but there is an easier way. To do this, we build a function that computes then binary search the interval [0,1] for , the value that minimizes this function. Then is our synthesized profile distribution. Table 4 tabulates part of the search space when our algorithm combines prof2 with prof5. The first column depicts values of in the interval ; the second through fifth columns show the corresponding ....

[Article contains additional citation context not shown here]

J. Anderson, L. Berc, J. Dean, S. Ghemawat, M. Henzinger, S. Leung, R. Sites, M. Vandevoorde, C. Wald- spurger, and W. Weihl. "Continuous Profiling: Where Have All the Cycles Gone?" ACM Transactions on Computing Systems, 15(4):357-390, Nov. 1997.

....on a set of sample inputs. Consequently, they cannot be adaptively optimized in order to account for changes in program behavior or to take advantage of variations in the production system. More recently, low overhead methods of profiling have been developed based on statistical sampling [1] [2] 3] 4] The basic approach, however, remains the same: profile information for the program s entire execution is averaged into a large database and later fed back into a static compiler. This approach is undesirable for three reasons. 1) The entire profile of each application must be ....

J. Anderson, L. M. Berc, J. Dean, S. Ghemawat, M. R. Henzinger, S. Leung, R. L. Sites, M. T. Vandevoorde, C. A. Waldspurger, and W. E. Weihl, "Continuous profiling: Where have all the cycles gone?," in Proc. of the 16th ACM Symposium of Operating Systems Principles, pp. 1--14, October 1997.

....(16.3 52.8 [3] Submitted to PLDI 00 Please do not distribute 2 A profiling system that is transparent to the user and collects profile data with minimal overhead would be a significant improvement. This line of thought helped to spur the development of profiling systems like DCPI [1] and Morph [17] which achieve extremely low overheads through the use of statistical sampling, and the recent investigations into hardware based approaches to profile gathering [6,10] Though the overhead in these approaches is nearly unnoticeable (Anderson et al. 1] report overhead of 1 3 and ....

.... profiling systems like DCPI [1] and Morph [17] which achieve extremely low overheads through the use of statistical sampling, and the recent investigations into hardware based approaches to profile gathering [6,10] Though the overhead in these approaches is nearly unnoticeable (Anderson et al. [1] report overhead of 1 3 and Conte et al. 6] report an overhead of 0.4 4.6 ) researchers are still investigating the usefulness of statistical and hardware generated profiles in code optimization. This paper introduces a new technique for program profiling called ephemeral instrumentation. The ....

[Article contains additional citation context not shown here]

J. Anderson, et al., "Continuous Profiling: Where Have All the Cycles Gone?," Proc. of 16th ACM Symposium on Operating Systems Principles (SOSP), Saint-Malo, France, October 1997.

....is helpful for isolating the location of Alpha instructions that make unaligned references to memory. Over several years, Compaq has also invested in the development of several tools for the analysis and modification of programs. These tools include: ffl DCPI: Execution profiling and analysis [1], ffl PatchWrx: Executable tracing [4, 15] and ffl NT ATOM: Execution driven simulation [3] Next we will describe these program analysis tools available for the Compaq Alpha architecture. 3.1.1 DCPI The Compaq Continuous Profiling Infrastructure (DCPI) is a tool for Alpha platforms that ....

....program analysis tools available for the Compaq Alpha architecture. 3.1. 1 DCPI The Compaq Continuous Profiling Infrastructure (DCPI) is a tool for Alpha platforms that provides continuous profiling of entire systems, including user application code, libraries, drivers and operating system kernel [1]. DCPI samples the Alpha hardware performance counters and collects execution profile information in its own database. The performance event counters on Alpha 21164 processors are extensive, measuring CPU cycles, D cache misses, I cache misses, branch mispredicts, CPU stalls, etc. DCPI provides a ....

J.M. Anderson, et al., "Continuous Profiling: Where Have All the Cycles Gone?," ACM Transactions on Computer Systems, Vol. 15, No. 4, November 1997, pp. 357-390.

....gather virtually any information they wanted, and left the choice of online or off line analysis to them. Many recent approaches to improving microprocessor performance have attacked or blurred the traditional lines between run time (hardware) and compile time (software) techniques. Digital s CPI [1] and Harvard s Morph project [18] collect profile information using statistical sampling 2 techniques at very low overhead; these profiles can then be used later to optimize the program. Machine emulators such as Digital s FX 32 [REF] initially interpret al..l of a program s code, but then ....

J.M. Anderson, L. Berc, J. Dean, S. Ghemawat, M. Henzinger, S.-T. A. Leung, D. Sites, M. Vandevoorde, C. Waldspurger, and W. E. Weihl. "Continuous profiling: Where have all the cycles gone?" Technical Note 1997-016. Digital Equipment Corporation Systems Research Center, Palo Alto, Calif., July 1997.

....Heuristic techniques do not profile; instead they analyze properties of the program or branch instructions and attempt to infer profile like information. Several recent systems suggest that the expense critique of profiled branch prediction is becoming irrelevant. Systems such as Digital s DCPI [3] and Harvard s Morph project [67] continuously profile applications with virtually unmeasurable overhead. They Page 4 of 136 then use the profiling results to direct optimization of application programs. Similarly, the Java language [5] is designed for interpretation or just in time ....

J.M. Anderson, L. Berc, J. Dean, S. Ghemawat, M. Henzinger, S.-T. A. Leung, D. Sites, M. Vandevoorde, C. Waldspurger, and W. E. Weihl. "Continuous profiling: Where have all the cycles gone?" Technical Note 1997-016. Digital Equipment Corporation Systems Research Center, Palo Alto, Calif., July 1997.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC