J. Hollingsworth and B. Miller, `Parallel program performance metrics: a comparison and validation', Proceeding, Supercomputing '92, Minneapolis, MN, pp. 4--13.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....of performance problems. One class of monitoring tools is software Profilers such as Gprof [4] Prof [7] and Quartz [1] that collect information via sampling. Unfortunately, profiling tools record a limited number of metrics, which restricts the amount of insight a user can gain from such tools [6]. Performance visualization and instrumentation tools such as Pablo [13] Paradyn [11] and AIMS [10] allow programmers to instrument the application code to collect runtime information. Interactive steering tools such as Falcon [2, 5] and Progress [17] take monitoring a step further by allowing ....

J. K. Hollingsworth and B P. Miller. Parallel Program Performance Metrics: A Comparison and Validation. In Supercomputing '92 Proceedings, Minneapolis, MN, November 1992. ACM and IEEE.

....specific tools to help users instrument, analyze, predict, and tune the performance of their parallel and distributed applications using broad range of approaches. Most of the developed tools tend to be used at runtime, using performance data gathered by instrumenting the application program [2,3,4,5,8]. A survey conducted by Pancake and Cook [6] has revealed the fact that tool use is still appallingly low among the high performance computing community. Cherri et al. 7] have cited the following three critical causes for this situation; 1) current tools are difficult to understand by scientific ....

J.K. Hollingsworth and B.P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation," Proc. Supercomputing'92, Minneapolis, Nov 1992.

.... to understanding the details of Web transaction performance uses the method of critical path analysis [70] Critical path analysis is the natural tool for understanding distributed applications because it identifies the precise set of activities that determine an application s performance [51, 79, 131]. Prior to our work, CPA has not been applied to TCP transactions. In [105] Paxson discusses the notion of cause effect issues in TCP transactions, which is an idea that also underlies our approach. However, Paxson had difficulty fully developing this notion because he was interested in ....

J. Hollingsworth and B. Miller. Parallel program performance metrics: A comparison and validation. In Proceedings of Supercomputing '92, Nov 1992.

....latency. Our general approach is to study these transactions using the method of critical path analysis [25] Critical path analysis is the natural tool for understanding distributed applications because it identi es the precise set of activities that determine an application s performance [17, 27, 41]. The central observation of critical path analysis as applied to distributed systems is that only some of the component activities in a distributed application are responsible for the overall response time; many other activities may occur in parallel, so that their executions overlap each other, ....

Jerey K. Hollingsworth and Barton P. Miller. Parallel program performance metrics: A comparison 11 and validation. In Proceedings of Supercomputing '92, November 1992.

....performance. For the 8 processor case, the estimate is within 3 of the observation. 4 Related Work Numerous efforts have gone into developing tools to help users understand the performance of their parallel applications. Typically these tools tend to be used at run time or as post processors [4, 5, 6, 8, 9, 10]. In these approaches, performance data is gathered by monitoring the program execution and or by collecting execution trace. These tools serve a useful purpose as performance tuning aids after an application has been parallelized. As such, these tools are not directly useful in designing and ....

J. Hollingsworth and B. Miller, Parallel program performance metrics: A comparison and validation. In Supercomputing'92 Proceedings, 1992.

....Pablo environment may be invoked to explore critical and near critical path activities from the most appropriate perspectives. The goal of performance debugging metrics is to rank the importance of improving specific program activities. Six parallel program performance metrics were compared in [17], and although no single metric was universally superior, the Critical Path Metric (CPM) provided the best overall 14 guidance. CPM ranks activities according to the magnitude of their durations on the critical path. The Maximum Benefit Metric (MBM) is an extension of the Critical Path Metric ....

J. K. Hollingsworth and B. P. Miller, "Parallel program performance metrics: A comparison and validation," in Proc. Supercomputing'92, IEEE Comput. Soc., Nov. 16--20, 1992, pp. 4--13.

....data from single processor systems. The computational cost of data reductions must not 2 A count of basic blocks is the best known example. be excessive, and graphical displays of performance data, if used, must scale to thousands of processors. As an example, critical path calculations [9], which identify the path through an event trace that most limits the computation, do not scale well. The computation cost becomes prohibitive with large event traces and large numbers of processors. Similarly, many types of graphical performance data displays that show the network ....

Hollingsworth, J. K., and Miller, B. P. Parallel Program Performance Metrics: A Comparison and Validation. In Supercomputing '92 (Nov. 1992), pp. 4--13.

....that limits its performance. Based on our experience with commercial and scientific users, Critical Path Profiling is an effective metric for tuning parallel programs. It is especially useful during the early stages of tuning a parallel program when load imbalance is a significant bottleneck [11]. In this paper, we introduce a runtime, nontrace based algorithm to compute the critical path profile. Our algorithm also permits starting or stopping the critical path computation during program execution and reporting intermediate values. Previous algorithms to compute the critical path ....

....those procedures that consume a large percentage of the total CPU time of the application. Selecting high CPU time procedures works well since, although critical path profiling may assign a different ordering and importance to the top 10 procedures, the procedures generally remain the same [11]. If top procedures are not the same, this can be detected since their cumulative share of the critical path length will be small. In this case, the programmer selects a different set of procedures and computes the critical path share for them. We could also automate the identification of the top ....

[Article contains additional citation context not shown here]

# J.K. Hollingsworth and B.P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation," Proc. Supercomputing 1992, pp. 4-13, Minneapolis, Minn., Nov. 1992.

....three sections: an overview of Critical Path, a description of the Slack algorithm, and a short case study comparing the results of Critical Path and Slack for two applications. A complete case study comparing Critical Path, Slack, and several other parallel performance metrics can be found in [3]. 2 2. Critical Path Critical Path Profiling is based on identifying the path through the program s execution that consumed the most time. To calculate a program s Critical Path, we build a graph of the program s execution, called a Program Activity Graph (PAG) This graph consists ....

J. K. Hollingsworth and B. P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation", Supercomputing 1992, Minneapolis, MN, November 1992, pp. 4-13.

....is Critical Path Profiling[16] Based on our experience with commercial and scientific users, Critical Path Profiling is an effective metric for tuning parallel programs. It is especially useful during the early stages of tuning a parallel program when load imbalance is a significant bottleneck[7]. In this paper we introduce a runtime, non trace based algorithm to compute the critical path profile. Our algorithm also permits starting or stopping the critical path computation during program execution and reporting intermediate values. Previous algorithms to compute the critical path profile ....

....to identify those procedures that consume a large percentage of the total CPU time of the application. Selecting high CPU time procedures works well since although critical path profiling assigns a different ordering and importance to the top ten procedures, the procedures generally remain the same[7]. If top procedures are not the same, this fact can be detected since their cumulative share of the critical path length will be small. In this case, the programmer can select a different set of procedures and compute the critical path share for them. 1 struct node one record per PAG node ....

[Article contains additional citation context not shown here]

J. K. Hollingsworth and B. P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation", Supercomputing 1992. Nov. 1992, Minneapolis, MN, pp. 4-13. 2 We will show how to permit this computation do be done in a single execution of the program at the end of this Appendix.

....the network, are serviced by the memories, and then carry results over the result half of the network back to the issuing cpu. The simulator computes the ################## In a different study, we verified that procedure effect join is actually more important to the runtime of the application [9]. 13 state of each network device (processor, switch, or memory) in parallel for one clock cycle and then performs a barrier synchronization before beginning the next clock cycle. The psim program statically assigns processes to compute the states of network elements and achieves ....

....To continue our study of the system we could refine our view of specific threads and 18 processes with procedure level critical path and profile analyses. Procedure level analyses identify specific procedures to be tuned and have led to performance improvements during past studies [8, 9]. However, for the present study we were primarily interested in identifying which applications affected each other, and process level analysis was sufficient. 6. Summary The support of multiple applications in a parallel program performance tool is a simple idea that is easy to implement, yet is ....

J. K. Hollingsworth and B. P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation", Supercomputing '92, November 1992.

....of the performance problem by reducing large volumes of performance data into single values or tables of values. Many metrics have been proposed for parallel programs: Critical Path[21] NPT[1] MTOOL[8] Gprof[9] Each of these metrics can provide useful information; however in an earlier paper[12] we compared several of these metrics (and a few variations) and concluded that no single metric was optimal for all programs. However, we did discover several factors that can be used to help select appropriate metrics. For example, whether an application is well balanced (e.g. all processes do ....

J. K. Hollingsworth and B. P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation", Supercomputing 1992, Minneapolis, MN, November 1992, pp. 4-13.

....with insight into the performance of their program with a relatively small amount of data. Each of these techniques assumes the programmer knows what picture to draw or which metric is appropriate. Unfortunately, no single approach applies to all parallel programs. In one study we conducted [40], four performance metrics were compared head to head for a number of real parallel programs. In addition, we employed a technique to calibrate the accuracy of the guidance supplied by the metrics. That study showed that for different programs, different metrics provided the best guidance. Since ....

....for sophisticated users, and a convenient way for tool developers to explore new metrics. However, many users do not know what additional metrics might be useful and are unable to use this functionality. The major problem with having many metrics is knowing which one to use. In a previous paper[40], we compared several different metrics (including Critical Path, Logical Zeroing, and NPT) and concluded that there was no single best metric. However, we were able to characterize the types of applications where each metric would be useful. This information is valuable to the programmer and ....

J. K. Hollingsworth and B. P. Miller, "Parallel Program Performance Metrics: A Comparison and Validation", Supercomputing 1992, Minneapolis, MN, November 1992, pp. 4-13.

No context found.

J. Hollingsworth and B. Miller, `Parallel program performance metrics: a comparison and validation', Proceeding, Supercomputing '92, Minneapolis, MN, pp. 4--13.

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