J. K. Hollingsworth, R. B. Irvin and B. P. Miller, `The integration of application and system based metrics in a parallel program peformance tool', Proc. Third ACM SIGPLAN Symp. on Principles and Practice of Parallel Programming (PPOPP), 1991, pp. 189--200.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....programs. Metrics based on the distribution of elapsed time across the program can systematically direct the user towards time consuming regions of code where performance should be improved. Unfortunately, such simple measures are insufficient for parallel programs and may even be misleading [HIM91]. This problem arises primarily because aggregate values do not necessarily capture the dynamic interactions between various processes (and processing nodes) involved in the computation. In order to fully understand the performance of a particular program on a particular machine, a performance ....

J. Hollingsworth, R. Irvin, and B. Miller, The Integration of Application and System Based Metrics in a Parallel Program Performance Tool, Proc. Third ACM SIGPLAN Symp. on Principles and Practice of Parallel Programming (PPOPP), pp. 189-200, 1991.

....of papers. Workload characterizations also need a stable clock. It is common to present histograms of small intervals such as interarrival time of packets, packet service times, file open time, read write service times, time to acquire a lock, and time spent in idle loops or at barriers, etc. [BHKSO91, CDJM91, FE89, HIM91, Mog90]. If the clock used to measure these intervals is unstable, then their distribution (histogram) will be the convolution of the actual intervals and the clock error. In other words, the shape and noise characteristics of the distribution will be altered from that of the actual intervals. When the ....

Hollingsworth, J.K., Irvin, R.B., and B.P. Miller. "The Integration of Application and System Based Metrics." 3rd. ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming, April 1991, pp. 206-217.

....metrics, for differing visualisation modes) availability . memory latency analysis (identify memory bottlenecks) system performance (os, h w and network) instrumentation timing overhead minimisation . heterogeneous processing support 3.2 Available tools 3.2. 1 IPS 2 The IPS 2 [HOLL91] parallel program measurement tool is produced by the University of Wisconsin. It supports C and fortran, and both shared memory and distributed Tools Survey 20th June 1993 12 memory programming models. Results can be visualised from a choice of four different metrics . tabular summary . time ....

J.K. Hollingsworth, R.B. Irvin, B.P Miller, The Integration of Application and System Based Metrics in a Parallel Program Performance Tool, Computer Sciences Department, University of Wisconsin-Madison, to appear in Proceedings of the 1991 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming 1991.

....W. Nagel and M. Linn [76] ffl Toward a better parallel performance metric, X. Sun and J. Gustafson [77] ffl Toward a taxonomy of performance metrics, J. Worlton [78] ffl The integration of application and system based metrics in a parallel program performance tool, J. Hollingsworth et al. [79]. ffl Parallel Program Performance Metrics: A comparison and validation, J. Hollingsworth and B. Miller [80] ffl Measuring parallelism in computation intensive scientific engineering applications, M. Kumar [81] ffl Selective Monitoring Using Performance Metric Predicates, C. Fineman and P. ....

J. Hollingsworth, R. Irvin, and B. Miller, "The integration of application and system based metrics in a parallel program performance tool", in Proc. of the 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming, April 1991, pp. 189--200.

....Their system specifies a standard, extensible format for event logs and supports a number of tools for summarizing and presenting information, including auralization and 3 D views on head mounted displays. Miller s group has also created a flexible system for visualizing performance data. IPS 2 [37, 18] uses two kinds of data: a standard trace log and a time histogram that notes the value of some performance metric as it varies over time. Users can see the critical path through their code and usage profiles of several program resources, along with a number of other views. The visualization ....

Hollingsworth, J. K., Irvin, R. B., and Miller, B. P. The integration of application and system based metrics in a parallel program performance tool. In Proceedings of the Third ACM SIGLAN Symposium on Principles and Practice of Parallel Programming (April 1991), pp. 189--200.

....detection, analysis and resolution. One difficulty for the tuner is that the amount of trace data produced may be very large, requiring sophisticated visualization mechanisms to present it in a form which aids analysis. This aspect of tuning is well supported by numerous trace visualization tools [1, 5, 6, 12, 11, 14, 15]. In order to resolve the performance problem, however, the tuner must relate the low level account of program behaviour provided by the trace data to the high level representation of program behaviour contained within the source code. VISPAT (VISualization for Performance Analysis and Tuning) ....

J. Hollingsworth, R. Irvin and B. Miller, The integration of application and system based metrics in a parallel program performance tool, ACM Symposium on Principles and Practice of Parallel Processing, ACM Press, May, 1991.

....VISPAT (VISualization for Performance Analysis and Tuning) a tool for parallel program performance analysis and tuning. Numerous examples of such tools already exist, many of which employ the latest graphical techniques for user interface design and visualization of trace data [RAM 92, HE91, HIM91, Moh91, MHJ91, Sto88, BB92, FR92] In spite of this, however, their design often appears to rest upon largely unstated assumptions concerning the nature of the tuning task, and evidence for their effectiveness is often lacking. Graphical user interfaces are not a panacea for usability problems ....

J. K. Hollingsworth, R. B. Irvin, and B. P. Miller. The integration of application and system based metrics in a parallel program performance tool. In ACM Symposium on Principles and Practice of Parallel Processing, May 1991.

....over time, but its displays are limited primarily to context switching. Other system level activity and activity from the application and other levels of the system are not displayed simultaneously for correlation. The IPS 2 performance measurement system for parallel and distributed programs [5] [14] does integrate both application and system based metrics. However, system metrics are dealt with strictly in the form of external time histograms , each describing the value of a single performance metric over time, as opposed to more general event data. Thus, where non application data are ....

J.K. Hollingsworth, R.B. Irvin, and B.P. Miller "The Integration of Application and System Based Metrics in a Parallel Program Performance Tool", Proc. Third Symposium on Principles and Practice of Parallel Programming, SIGPLAN Notices 26(7), July 1991, pp. 189-200.

....charts, strip charts, contour plot,s dial, interval plot, kiviat diagrams, LEDs, X Y line scatter plots, matrix displays, pie chart, polar plots, 3 D scatter plot. 6.6 IPS 2 This tool was developed at the University of Wisconsin by Jeff Hollingsworth, Bruce Irvin and Barton P. Miller. IPS 2 [HIM91] is an interactive, trace driven performance measurement system for parallel and distributed programs . The IPS 2 can measure shared memory and message passing parallel programs running on a heterogeneous collection of machines. The events are only generated by software probes inserted in the ....

J. Hollingsworth, B. Irvin, and P. B. Miller. The integration of application and system based metrics in a parallel program performance tool. To appear in Proceedings of the 1991 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, 1991. accessible by anonymous FTP in grilled.cs.wisc.edu:technical reports/edcu.ps.Z.

....of papers. Workload characterizations also need a stable clock. It is common to present histograms of small intervals such as interarrival time of packets, packet service times, file open time, read write service times, time to acquire a lock, and time spent in idle loops or at barriers, etc. [BHKSO91, CDJM91, DE95, FE89, HIM91, Mog90]. If the clock used to measure these intervals is unstable, then their distribution (histogram) will be the convolution of the actual intervals and the clock error. In other words, the shape and noise characteristics of the distribution will be altered from that of the actual intervals. When the ....

Hollingsworth, J.K., Irvin, R.B., and B.P. Miller. The Integration of Application and System Based Metrics. 3rd. ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming, April 1991, pp. 206-217.

....incorporated into high performance microprocessors and systems. The IBM RP3[5] included performance monitoring hardware throughout the system. The Sequent Symmetry s also included hardware to measure bus utilization, bus read and write counts, and cache miss rates. Using this information, tools[12] were able to give the programmer insights that substantially simplified program tuning. Unfortunately, Sequent never widely disclosed the existence of these features due to fears that information obtained using the hardware instrumentation would be used by their competitors. Another early example ....

J. K. Hollingsworth, R. B. Irvin, and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool," 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming. April 21-24, 1991, Williamsburg, VA, pp. 189-200.

....forms a matrix (called a grid in Paradyn) containing the value of each metric for each focus. The Performance Consultant and visualizations receive performance data by specifying one or more metric focus grids. Paradyn stores performance data internally in a data structure called a time histogram [4]. A time histogram (also known as a time series or time array) is a fixed size array whose elements (buckets) store values of a metric for successive time intervals. Two parameters determine the granularity of the data stored in time histograms: initial bucket width (time interval) and number of ....

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", 1991 ACM SIGPLAN Notices Symposium on Principles and Practice of Parallel Programming, April 1991, pp. 189-200.

....have a working program that does not alter the performance of other procedures. We must also be alert for performance changes caused by interactions with architectural features such as caching and virtual memory. IPS 2 allows us to easily monitor the levels of cache and virtual memory activity [11], so we can check that the modified program does not significantly change these behaviors. Several approaches were used to compute the True Zeroing value for each procedure. 1) The easiest approach is to simply remove the procedure, which works well for printing routines and other parts of the ....

....had a large amount of system time from page faults. While this time is important, unless the programmer is aware that the problem with a particular procedure is due to system time, they will not know how to fix it. Currently IPS 2 collects system time information via an external sampling process [11]. This approach provides good coarse grained information, but does not isolate the cause of a system time bottleneck to a specific procedure. We also discovered that the slack metric generally failed to provide useful guidance. Because all of the programs considered in our case study were written ....

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", Proc. of the 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming, April 1991, pp. 189-200.

....the performance of individual applications in the presence of other applications. The user can also study the effect of the workload on various parts of the system. IPS 2 includes a simple, open interface for the incorporation of external data from hardware, network, or operating system monitors [8]. If the computing environment already includes such performance monitors then their output can be directed into IPS 2 through the external data interface. Data gathered by external monitors may be used in the same displays and analyses as data gathered with IPS 2 program tracing, and the user can ....

....at which program events occur. Figure 6. Figure 7. Critical Path with 10 msec clock. Critical Path with 1 usec clock. 5.2. Scheduling, Synchronization Policies, and Workload Performance This case study examines a widely studied application [14] that was previously tuned in isolation using IPS 2 [8]. The application, called psim, simulates an indirect k ary, n cube processor memory interconnection network. Over the course of a simulation several memory request packets are issued from each simulated cpu. The packets travel over the request half of the network, are serviced by the memories, ....

[Article contains additional citation context not shown here]

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming, April 1991, pp. 189-200.

....is used only to periodically report the collected data. Varying the sampling rate affects only our rate of decision making and granularity of phase boundaries; it does not affect the accuracy of the underlying performance data. Collected data is stored in a data structure called a time histogram[5]. A time histogram is a fixed size array whose elements store values of a performance metric for successive time intervals. Two parameters determine the granularity of the data stored in time histograms: initial bucket width (timer interval) and number of buckets. Both parameters are supplied by ....

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming, April 1991, pp. 189-200.

....event traces during a program s execution. Each event (e.g. procedure call or synchronization operation) contains both wall clock and process time stamps in addition to some event specific data. In addition to normal IPS 2 instrumentation, we ran the programs with two External Data Collectors[11]. External Data Collectors are dedicated sampling processes that collect additional information not available via tracing. One collector gathered information about the behavior of the operating system (e.g. page faults, context switch rate) The other collected data about the hardware (e.g. ....

....one axis ( where in this case) we are able to use another axis ( when ) to help isolate the problem. This flexible approach to finding bottlenecks is an important feature of our work. To validate this result, we again used the IPS 2 performance tools. Since we had previously studied this program[11], we recognized the page fault problem as one of the problems in this program. The problem was due to the creation of new user data in the program. A few small changes to the program reduced this page fault behavior and improved the execution time by 10 . Figures 11 and 12 show the volume of ....

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", Proc. of the 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming , Williamsburg, VA, April 21-24 1991, pp. 189-200. appears as SIGPLAN Notices, July 1991.

No context found.

J.K. Hollingsworth, R.B. Irvin, and B.P. Miller. The Integration of Application and System Based Metrics in a Parallel Program Performance Tool. 3rd ACM Symp. on Principles and Practice of Parallel Programming, Williamsburg, VA (April 1991).

....forms a matrix (called a grid in Paradyn) containing the value of each metric for each focus. The Performance Consultant and visualizations receive performance data by specifying one or more metric focus grids. Paradyn stores performance data internally in a data structure called a time histogram [3]. A time histogram is a fixed size array whose elements (buckets) store values of a metric for successive time intervals. Two parameters determine the granularity of the data stored in time histograms: initial bucket width (time interval) and number of buckets. Both parameters are supplied by ....

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", 1991 ACM SIGPLAN Notices Symposium on Principles and Practice of Parallel Programming, April 1991, pp. 189-200.

....force them to select the appropriate resources to profile their application. Building a tool that includes all potentially useful metrics is difficult. An alternative is provide a library of metrics, and make the tool extensible to permit users to create their own metrics. One such tool is IPS 2[39] which permits users to create new metrics as algebraic expressions of previously defined metrics. For example, IPS 2 does not have a built in metric to indicate what fraction of the actual CPU time used went to each process. However, it is possible to define this metric in terms of a process ....

....environment. The Mach Kernel Monitor[56] instruments context switches to trace the state of processes through time. If modifying the operating system is not feasible or desirable, monitoring can be accomplished via dedicated data collection processes. The External Data Collection facility of IPS 2[39] provides this capability. Information from the operating system is gathered by the collector processes and reported to the user. Most operating systems collect many statistics and make this information available via system calls. By using these system calls, data that could only be collected by ....

[Article contains additional citation context not shown here]

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, "The Integration of Application and System Based Metrics in A Parallel Program Performance Tool", 1991 ACM SIGPLAN Symposium on Principals and Practice of Parallel Programming , Williamsburg, VA, April 21-24 1991, pp. 189-200. appears as SIGPLAN Notices, July 1991.

No context found.

J. K. Hollingsworth, R. B. Irvin and B. P. Miller, `The integration of application and system based metrics in a parallel program peformance tool', Proc. Third ACM SIGPLAN Symp. on Principles and Practice of Parallel Programming (PPOPP), 1991, pp. 189--200.

No context found.

J.K. Hollingsworth, R.B. Irvin, and B.P. Miller. The Integration of Application and System Based Metrics in a Parallel Program Performance Tool. 3rd ACM Symposium on Principles and Practice of Parallel Programming, Williamsburg, VA (April 1991).

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