A.D. Malony, D.A. Reed, H.A.G. Wijshoff, "Performance measurement intrusion and perturbation analysis," IEEE Transactions on Parallel and Distributed Systems, vol.3, no.4, 1992, pp. 433-50.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....testing, remove the probe in task A, scenario (2) may occur again and the erroneous calculation B(A(X) may be executed, leading to a failure. This non deterministic effect of intrusively observing a system is called the probe effect [31] 78] or the Heisenberg uncertainty in software [64][77]. Figure 4 1. The releases of tasks A and B Figure (a) Where task B has higher priority than task A. Task B enters the critical section before A, when A has its worst case execution time. Figure (b) depicts the resulting execution, where A is preempted by B. B A Critical section ....

Malony A. D., Reed D. A., and Wijshoff H. A. G. Performance measurement intrusion and perturbation analysis. IEEE Trans. on Parallel and Distributed Systems 3(4):433-450, July 1992.

....in this branch, however, are not related to the development process of parallel programs but do rather stand as contributions to the art of modeling for its own. First approaches to relate performance engineering to the development cycle were attempted by performance measurement and monitoring [Malo 92] of parallel systems, the visualization [Hari 93] of concurrent behaviors and the performance tuning [Ande 90] of codes. All these activities are launched at the very end of the development cycle, such that early performance mistakes, although detected, studied and fine tuned, cannot be corrected ....

A. D. Malony, D. A. Reed, and H. A. G. Wijshoff. "Performance Measurement Intrusion and Perturbation Analysis". IEEE Transactions on Parallel and Distributed Systems, Vol. 3, No. 4, pp. 433 -- 450, July 1992.

....high overhead anyway, but may be very problematic for the study of ne grain events related to communication, synchronization, and memory usage. One possible solution to this problem is to model the e ect of the instrumentation, thereby enabling it to be factored out of the measurement results [55]. This leads to results that re ect real system behavior (that is, una ected by the instrumentation) but leaves the problem of performance degradation while the measurements are being taken. An alternative is to selectively activate only those parts of the instrumentation that are needed at each ....

A. D. Malony, D. A. Reed, and H. A. G. Wijsho, \Performance measurement intrusion and perturbation analysis". IEEE Trans. Parallel & Distributed Syst. 3(4), pp. 433-450, Jul 1992.

....where parallel programming is applied, is a dominant issue [7] Attempts have been made to formulate methods following traditional life cycle models which include performance engineering in the final stages of the development. For example, the generation of information at run time for monitoring [13] and performance tuning [15] Performance measurements are obtained after the program has been developed and, if necessary, the program is modified (redesigned and reimplemented) in order to improve performance. The cost and effort necessary in modifying a design when an error is only found after ....

A. D. Malony, D. A. Reed, and H. A. G. Wijshoff. Performance measurement intrusion and perturbation analysis. IEEE Transactions on Parallel and Distributed Systems, 3(4):433--450, July 1992.

....to the return of the instrumentation point rid.ThisDExecution #rid# captures the time the instrumentation waits on mutex locks, full buffers or other costs. Small changes in application instruction sequences from this instrumentation can result in substantial aggregate performance perturbation[15]. We measure this aggregate perturbation instead of the execution time of each individual instrumentation point. Application runtime, # application = # application stop , # application start , is the runtime of the original application without any instrumentation points. Instrumented ....

A. Malony, D. Reed, and H. Wijshoff. Performance measurement intrusion and perturbation analysis. IEEE Trans. Parallel and Distributed Systems, 3(4):433--50, 1992.

....detection and or on the event processing phases, as suggested above. Also, one needs to look for a suitable balance depending on the allowable effective perturbation. Of course, this introduces the problem of precise measurement and evaluation of monitor intrusion, a very important subject (see [21]) After quantification of the perturbation, one tries to compensate it, by adjusting the event orderings and the event timings. Due to unacceptable monitor intrusion, sometimes software monitoring is ruled out, e.g. in real time systems with critical timing constraints. 3.2 Hardware Monitoring ....

A.D. Malony, D.A. Reed, H.A.G. Wijshoff. Performance measurement intrusion and perturbation analysis. IEEE Trans. on Parallel and Distributed Systems, vol 3, no 4, Jul 1992.

....improves system performance. 2 The students are assigned readings from the textbook The Art of Computer Systems Performance Analysis [2] and several relevant papers selected by the instructor from the current literature [3] 4] 5] 6] 7] 8] 9] 10] 11] 12] 13] 14] 15] [16], 17] The outline for this course is shown in Table I. It is assumed that, before taking the course, the students have developed an understanding of computer organization and design and basic computer architecture # have the ability to program in a general purpose computer language, such as C, ....

Allen D. Malony and Daniel A. Reed, "Performance measurementintrusion and perturbation analysis," in IEEE Tran. Parallel Distributed Systems, July 1992, vol. 3, pp. 433--450.

....executions upon the performance factors of other instruction executions. One application is to predict the performance of an original program from an instrumented execution with a model to compensate for overhead of direct and indirect perturbations (see section 4) We extend models like those of [22] to the Pentium and to metrics beyond execution time (see section 5) All but a few ( 22, 6] e.g. prior studies and tools do not account for overhead. A second application is to test assumptions about our hardware s organization and dynamic behavior. We seek full control and selectivity of ....

.... is to predict the performance of an original program from an instrumented execution with a model to compensate for overhead of direct and indirect perturbations (see section 4) We extend models like those of [22] to the Pentium and to metrics beyond execution time (see section 5) All but a few ([22, 6], e.g. prior studies and tools do not account for overhead. A second application is to test assumptions about our hardware s organization and dynamic behavior. We seek full control and selectivity of hardware counter measurements with quantified overheads. We then seek to minimize the overhead, ....

[Article contains additional citation context not shown here]

Allen D. Malony, Daniel A. Reed, and Harry A. G. Wijshoff. Performance measurement intrusion and perturbation analysis. IEEE Transactions on Parallel and Distributed Systems, 3(4):433--450, 1992.

....most parallel programs will yield different process or task instruction interleavings each time they are executed. This interleaving is likely to be dependent not only on characteristics of the execution architecture, but also on subtle timing perturbations introduced by the tracing software [12, 11], other users on the system, page faults, and so on. The effect on the overall execution of the program depends on its structure. Traces of statically scheduled parallel programs can be affected in subtle ways, such as the time to rendezvous at a barrier. Dynamically scheduled programs can be ....

Malony, A. D., Reed, D. A., and Wijshoff, H. Performance Measurement Intrusion and Perturbation Analysis. Tech. Rep. CSRD No. 923, Center for Supercomputing Research and Development, Oct. 1989.

.... data by multivariate analysis [9] and cluster analysis [28] the use of three dimensional visualizations [12, 31] the use of multimedia [5, 15] and virtual reality [29] characterization in the event and frequency domains [Abrams] and development of event based models of parallel system behavior [1, 24]. The consensus is that we need to call upon candidate resources to attack the problem of analyzing and presenting performance data. To this end, we have developed the multiple domain analysis approach [38] Consider this excerpt: The history of science shows us that a given problem can ....

Malony, A. D., D. A. Reed, and H. A. G. Wijshoff, "Performance Measurement Intrusion and Perturbation Analysis, " IEEE Trans. Par. and Distr. Sys., 3(4), July 1992.

....true. The basic technique of measuring is using instrumentation, which will introduce performance overheads and may also affect the system s dynamic behavior [Jelly94] This is called instrumentation perturbation. Different kinds of perturbations, direct or indirect, are described by Malony [Malony92] Besides, the accuracy of measurement also depends on several other things, such as time of the measurement and the workload used. Thus, measurement techniques may offer very accurate data or very inaccurate data, all depending on how well the above factors are treated. As for the cost, because ....

....system points (the connections are made with probes) where they detect signals characterizing the phenomena to be observed [Ferrari83] A hybrid monitor is a combination of software and hardware. Many examples of software monitors can be found in the literature (see e.g. Plattner81; Power83; Malony92] Examples of hardware monitors are 26 COMTEN and SPM [Ibbett78] Ries93] describes the Paragon performance monitoring environment that uses a hardware performance monitoring board. And examples of hybrid monitors are Diamond [Hughes80] and HMS [Hadsell83] Each class of monitors has its own ....

[Article contains additional citation context not shown here]

Malony, A. D., Reed, D. A. and Wijshoff, H. A. G., "Performance measurement intrusion and perturbation analysis," IEEE Transactions on Parallel and Distrubuted Systems, vol. 3, no. 4, pp. 443450, July 1992.

....often supports tools accessed by end users; the user typically sees the tool and not the IS. Consequently, tools are scrutinized, and the IS and its overheads receive little attention. Users may be unaware of the impact of the IS. Unfortunately, the IS can perturb the behavior of the application [17], degrading the performance of an instrumented application program from 10 to more than 50 according to various measurement based studies [8,19] Perturbation can result from contention for system resources among application and instrumentation processes. With increasing sophistication of system ....

....The IS can use the model to adapt its behavior in order to regulate overheads. Some initial work has already been done in this direction for Paradyn [12] Previous work related to IS modeling and overhead analysis has focused on analyzing the intrusion due to instrumenting parallel programs [17,32]. Several other researchers have given special attention to the monitoring overheads of their tools. Miller et al. present measurements of overheads of the IPS 2 tool and compare them with the 23 overheads of a functionally similar tool, gprof [19] Gu et al. use synthetic workloads to exercise ....

Malony, A. D., D. A. Reed, and H. A. G. Wijshoff, "Performance Measurement Intrusion and Perturbation Analysis," IEEE Transactions on Parallel and Distributed Systems, 3(4), July 1992.

....testing, remove the probe in task A, scenario (2) may occur again and the erroneous calculation B(A(X) may be executed, leading to a failure. This nondeterministic effect of intrusively observing a system is called the probe effect [18] 42] or the Heisenberg uncertainty in software [36][41]. 3 t Priority 4 10 Figure 1 2. The releases of tasks A and B figure (a) Where B has higher priority than task A. Due to shorter execution time task A starts and terminates before task B is released. Figure (b) depicts the resulting execution, where A precedes B. B A t Priority 4 ....

Malony A. D., Reed D. A., and Wijshoff H. A. G. Performance measurement intrusion and perturbation analysis. IEEE Trans. on Parallel and Distributed Systems 3(4):433-450, July 1992.

....different mechanisms (i.e. software, hardware or hybrid) to record events of interest related to performance. However, these profiling mechanisms may perturb the behaviour of the application being monitored, becoming a significant factor. This factor, often called invasiveness [27] perturbation [17] or intrusiveness [2] must be taken into account when measuring the performance of an application in order to provide more accurate information. The purpose of this paper is to determine the degree of invasiveness of PARMACS, a well known set of macros which provides a portable programming ....

....manifest themselves in several ways: execution slowdown, changes in memory reference patterns, event reordering, and even register interlock stalls. From a performance evaluation perspective, instrumentation perturbations must be balanced against the need for detailed performance data [17]. With the exception of passive hardware performance monitors, performance experiments rely on software instrumentation for performance data capture. Such instrumentation mandates a delicate balance between volume and accuracy. Excessive instrumentation perturbs the measured system; limited ....

[Article contains additional citation context not shown here]

Allen Malony, Daniel Reed, and H. Wijshoff. Performance-measurement intrusion and perturbation analysis. IEEE Transactions on Parallel and Distributed Systems, 3(4):433-- 450, 1992.

....interest. Likewise, we take the mean over 25 runs to assure a high degree of accuracy for the input parameter used. Additional benchmarking approaches are presented in Section 3.D.3. It is possible that adding timing statements to a code in order to benchmark sections of it may perturb the code [MRW92, HM96] If a large perturbation is indicated, the input parameter for this value should be tagged with meta data indicating the possible error in the value. 5. How are application profiles defined Profiles for applications running on a shared cluster of workstations may be dependent on the ....

A. Malony, D. Reed, and H. Wijshoff. Performance measurement intrusion and perturbation analysis. IEEE Transactions on Parallel and Distributed Systems, 3(4), July 1992.

....if any reordering of events does not affect the final result of the application. Greater invasiveness implies longer execution times and less accuracy. From a performance evaluation perspective, instrumentation perturbations must be balanced against the need for detailed performance data [34]. IV. PERFORMANCE ANALYSIS USING A DATA VISUALIZATION TOOL The amount of trace information generated by a program running on a large number of processors can be enormous. However, data visualization environments can represent this NA iM, HEY AND ZALUSKA: DO LOOP SURFACE. 9 Add instructions to ....

Allen Malony, Daniel Reed, and H. Wijshoff, "Performance-measurement intrusion and perturbation analysis", IEEE Transactions on Parallel and Distributed Systems, vol. 3, no. 4, pp. 433--450, 1992.

....mechanisms (i.e. software, hardware or hybrid) to record events of interest related to performance. However, these profiling mechanisms may perturb the behaviour of the application being monitored and can become a significant factor. This factor, often called invasiveness [137] perturbation [140] or intrusiveness [54] must be taken into account when measuring the performance of an application in order to provide more accurate information. The purpose of this Chapter is to describe our experience determining the degree of invasiveness of PARMACS, a well known set of macros which provides ....

....manifest themselves in several ways: execution slowdown, changes in memory reference patterns, 66 67 event reordering, and even register interlock stalls. From a performance evaluation perspective, instrumentation perturbations must be balanced against the need for detailed performance data [140]. With the exception of passive hardware performance monitors, performance experiments rely on software instrumentation for performance data capture. Such instrumentation mandates a delicate balance between volume and accuracy. Excessive instrumentation perturbs the measured system; limited ....

[Article contains additional citation context not shown here]

Allen Malony, Daniel Reed, and H. Wijshoff, "Performance-measurement intrusion and perturbation analysis", IEEE Transactions on Parallel and Distributed Systems, vol. 3, no. 4, pp. 433--450, 1992.

....in IS development. The IS represents enabling technology of growing importance for effectively using parallel and distributed systems. However, users typically see a tool and not the IS and therefore may be unaware of IS overheads. Unfortunately, the IS can perturb the behavior of the application [20], degrading the performance of an instrumented application program from 10 percent to more than 50 percent according to various measurement based studies [9] 22] Perturbation can result from contention for system resources among application and instrumentation processes. With increasing ....

....These limits can be used to adapt IS behavior in order to regulate overheads. Some initial work has already been done in this direction for Paradyn [13] Previous work related to IS modeling and overhead analysis has focused on analyzing the intrusion due to instrumenting parallel programs [20], 32] Several other researchers have given special attention to the monitoring overheads of their tools. Miller et al. present measurements of overheads of the IPS 2 tool and compare them with the overheads of a functionally similar tool, gprof [22] Gu et al. use synthetic workloads to ....

# A.D. Malony, D.A. Reed, and H.A.G. Wijshoff, "Performance Measurement Intrusion and Perturbation Analysis," IEEE Trans. Parallel and Distributed Systems, vol. 3, no. 4, pp. 433--450, July 1992.

....ease of use. Traditionally, application or OS level software instrumentation is easy to customize for particular measurement based experiments that provide specific information. The main drawbacks of this technique include the intrusion to the system under test and lack of hardwarelevel details [10]. Simulation based techniques are often used when detailed measurements are needed under carefully controlled operating conditions [4] However, these studies are limited by the accuracy of the simulator and the amount of time and computing resources required to execute a high fidelity ....

A.D. Malony, D.A. Reed, and H.A.G. Wijshoff, "Performance Measurement Intrusion and Perturbation Analysis," IEEE Transactions on Parallel and Distributed Systems, 3(6), Nov. 1992, pp. 657-671.

....programs in comparison with Gprof. Work has been done on compensating for the effects of program perturbation due to instrumentation [38] The goal of perturbation compensation is to reconstruct the actual program behavior from the perturbed behavior as it may be recorded by the IS. Malony et al. [20] describe a model for removing the effects of perturbation from the traces of parallel program executions. Presently, it is not standard practice to formally evaluate the performance and functionality of a tool early in its development. Usability and efficiency studies of prototypical tools are ....

Malony, A. D., D. A. Reed, and H. A. G. Wijshoff, "Performance Measurement Intrusion and Perturbation Analysis," IEEE Transactions on Parallel and Distributed Systems, 3(4), July 1992. 22

....new errors to appear. It will not only change the real time execution speed, but also change the probability of making particular non deterministic choices [7] Several approaches to postmortem analysis of the intrusion have been proposed. Such approaches include the models of Malony and Reed [15,16] which adjust the trace event times to remove perturbation through event trace and execution time analysis, and the trace analysis technique of Helmbold, McDowell and Wang [11] which determines possible event orders by analyzing the trace information collected at runtime. However, postmortem ....

A. Malony, D.A. Reed and H. Wijshoff, Performance measurement intrusion and perturbation analysis, IEEE Trans. Parallel and Distributed Systems, Vol.3, No.4, pp. 433-450, July 1992.

....problems, developers must use performance instrumentation systems that capture detailed information on a large number of these time varying performance metrics. Unfortunately, this instrumentation can influence the performance of the target system and can produce tremendous volumes of data [13]. To help combat these consequences of performance instrumentation, the instrumentation system should provide support for selecting metrics and measurement points so as to minimize the effects of its instrumentation. As a result, the challenge for developers of performance analysis systems is to ....

A. D. Malony, D. A. Reed, and H. A. G. Wijshoff, "Performance measurement intrusion and perturbation analysis," IEEE Trans. Parallel and Distributed Systems, 3(4):433-50, 1992.

.... to reliably generate computational results [22, 23, 20] Early work in the performance analysis of parallel systems mainly focused on the mechanics of empirical analysis: measurement and monitoring, automatic probe insertion, trace generation, post execution trace analysis, perturbation analysis [21] and trace visualization [16] It was soon recognized that starting with performance debugging activities after fully functional executable code has been developed, reduces the degrees of freedom for program modifications to those that do not require significant changes to the original program ....

A. D. Malony, D. A. Reed, and H. A. G. Wijshoff. Performance Measurement Intrusion and Perturbation Analysis. IEEE Transactions on Parallel and Distributed Systems, 3(4):433 -- 450, July 1992.

.... to reliably generate computational results [22, 23, 20] Early work in the performance analysis of parallel systems mainly focused on the mechanics of empirical analysis: measurement and monitoring, automatic probe insertion, trace generation, post execution trace analysis, perturbation analysis [21] and trace visualization [16] It was soon recognized that starting with performance debugging activities after fully functional executable code has been developed, reduces the degrees of freedom for program modifications to those that do not require significant changes to the original program ....

A. D. Malony, D. A. Reed, and H. A. G. Wijshoff. Performance Measurement Intrusion and Perturbation Analysis. IEEE Transactions on Parallel and Distributed Systems, 3(4):433 -- 450, July 1992.

No context found.

A.D. Malony, D.A. Reed, H.A.G. Wijshoff, "Performance measurement intrusion and perturbation analysis," IEEE Transactions on Parallel and Distributed Systems, vol.3, no.4, 1992, pp. 433-50.

First 50 documents  Next 50

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