Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Proceeding of Scalable High Performance Computing Conference, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....of software systems using hardware, software and hybrid systems[12] has been studied for years. Many works dealt with the techniques used to instrument a system at compile or runtime in various use cases. Dynamic instrumentation systems such as Paradyn[3] KernInst[10] or the system presented in [4] insert the instrumentation code into running binaries. Other binary rewriters like not opti stripped fully opmized down timized Full sensor 104 104 53 Minimal sensor 104 32 7 Example (sensors) 2496 426 344 Example (per event) 104 35.5 28.7 Table 5. Summary: possible sensor sizes (byte) ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Proceeding of Scalable High Performance Computing Conference, 1994.

....below) communicate using CORBA. Tool Tool OCB AM Agg. PM Dyninst Appl. Data flow Control flow PM Dyninst Appl. Figure 1: The FIRST Basic Architecture. The Process Manager (PM) controls and instruments processes using the dynamic instrumentation mechanisms provided by the Dyninst [3] library. In Dyninst terminology, the part of the PM that performs the instrumentation is known as the mutator and the process it instruments, the mutatee. With Dyninst, one mutator can instrument multiple mutatees on a single processor. A PM must reside on the same processor as the application ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic Program Instrumentation for Scalable Performance Tools. In Proceedings of the 1994 Scalable High Performance Computing Conference, Knoxville, TN, May 1994.

....RS6000 AIX exists) EEL s RTL is also very RISC like, and thus it remains to be seen whether EEL could successfully support other processor types such as x86. DYNINST is a system for run time instrumentation that has been successfully deployed 76 on the x86, Alpha, PowerPC, and SPARC platforms [13] [14] It adds instrumentation at run time by replacing instructions with fixed jumps to instrumentation routines, in a way similar to the application of SFI checking in Naccio Win32. DYNINST uses a platformindependent language called MDL that is tailored to program instrumentation and allows ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic Program Instrumentation for Scalable Performance Tools. In Scalable High Performance Computing Conference, May 1994.

....bart cs.wisc.edu Informix Software, Inc. 921 SW Washington St. Portland, OR 97205 Computer Sciences Department University of Wisconsin 1210 W. Dayton Street Madison, WI 53706 1685 Page 2 November 27, 1995 structures. We show how these techniques can be combined with dynamic instrumentation [13,19] (a technique for on the fly instrumentation of executable programs) from the Paradyn Parallel Performance Tools. We present two of our case studies of the use of this tool to profile and improve the performance of real world parallel applications. Our approach to tools for high level parallel ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Scalable High Performance Computing Conference, May 1994.

....emphasizes low latency, on line monitoring, performed by capturing only those program attributes required for specific performance analyses or for specific program steering. This distinguishes our work from related research on performance monitoring and tuning, including that of Reed[42] and Miller[21], both of which generally address the issue of performance debugging using program traces stored in intermediate files. These projects primary concern is not the latency with which program events are transferred from the program to the end user (i.e. to an interactive user interface or to an ....

....the selection of suitable monitoring performance for specific monitoring and steering tasks, and they may be used to adapt the monitoring system to dynamic changes in workload imposed by the target application. For example, when heavy monitoring is 4 Related work by Hollingsworth and Miller [21] removes instrumentation points in order to completelyeliminate the overheads of such turned off instrumentation points. detected by a simple monitor monitor mechanism, new local monitoring threads may be forked. Similarly, when bursty monitoring traffic is expected with moderate requirements ....

[Article contains additional citation context not shown here]

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Proceedings of SHPCC'94, pages 841--850, Knoxville, TN, May 1994. IEEE Computer Society Press.

....tools, lan 15 guage specific tools, and event mapping tools. General performance tools offer a variety of data collection techniques, from specialized hardware [9,64,67] to library instrumentation [37,40,44] to source code instrumentation [54] to dynamic rewriting of code in execution [24,43]. The NV model does not require any particular method of collection and the tools described in this dissertation employ a variety of data collection techniques pioneered by other tools. However, various techniques for dynamic mapping of performance data are more practically implemented with ....

....performance tool can use the dynamic mapping information during or after run time to relate performance measurements to abstract program constructs and activities. In this section we discuss two important techniques for collecting dynamic mapping information. The first uses dynamic instrumentation [24] to reduce the perturbation effects of collecting dynamic mapping information, and the second uses a data structure called the Set of Active Sentences to discover verb mappings that are otherwise difficult to detect. 4.4.1 The Use of Dynamic Instrumentation for Dynamic Mapping A mapping point is ....

[Article contains additional citation context not shown here]

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Scalable High Performance Computing Conference, May 1994.

....parameters. The objective of SIMPARC is to make the program independent of the destination machine s parameters, whereas the losses in performance that MTOOL reveals in the program, are completely dependent on the specific architecture on which the program is running. Other instrumentation systems [21, 8, 11, 15, 16] include: advance display features (e.g. program view and processors states) memory utilization, network load and some hardware support for monitoring. In contrast to Simparc all the above systems are intrusive thus changing the scheduling during execution. Another advantage of the Simparc ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Scalable High-Performance Comput. Conf., pages 841--850, May 1994.

....emphasizes low latency, on line monitoring, achieved by capturing only those program attributes required for specific performance analyses or for specific program steering. This distinguishes our work from related research on performance monitoring and tuning, including that of Reed[43] and Miller[23], both of whom generally address the issue of performance debugging using program traces stored in intermediate files. These projects primary concern is not the latency with which program events are transferred from the program to the end user (i.e. to an interactive user interface or to an ....

....sensors. At the local monitor level, trace data is further analyzed to produce high level information. The partially processed monitoring information can be fed to Falcon s steering mechanism to effect on line changes to the program and its execution 2 Related work by Hollingsworth and Miller [23] removes instrumentation points in order to completelyeliminate the overheads of such turned off instrumentation points. environment. It can be sent to Falcon s central monitor for further analysis and for display to end users, and it can also be stored in trace data files for postmortem ....

[Article contains additional citation context not shown here]

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Proceedings of SHPCC'94, pages 841--850, Knoxville, TN, May 1994. IEEE Computer Society Press.

.... work in program monitoring has focused on helping programmers understand the performance of their parallel codes, minimizing or correcting program perturbation due to monitoring, reducing the amounts of monitoring or trace information captured for parallel or distributed program debugging [OSS93, HMC94] and the effective replay [LMC87] or long term storage of monitoring information. In comparison, interactivity, in the form of on line program steering, specifically requires its on line monitoring system to be able to: 1) capture application specific information, 2) impose controlled ....

....ineffective distance checks are detected. Using Falcon s monitoring specification language [Sno87] programmers may define application specific sensors for capturing both the program and performance behavior to be monitored and the program attributes 2 Related work by Hollingsworth and Miller [HMC94] removes instrumentation points completely to reduce the overheads of these turned off instrumentation points to zero. based on which steering may be performed. The specification of a sample tracing sensor is shown below: sensor workload attributes int domainnum; double workload; The ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic program instrumentation for scalable performance tools. In Proceedings of SHPCC'94, pages 841--850, Knoxville, TN, May 1994.

....problems associated with profile and trace based tools. The performance data collection part of the tool is designed to scale to long running, massively parallel applications by using two mechanisms: fixed length data structures to store time varying performance data and dynamic instrumentation [27] (instrumentation code that can be inserted or removed from a running program at any point in its execution) Digital s continuous profiling infrastructure [51] uses a slightly different technique to provide low overhead performance data. It periodically samples the Alpha performance counter ....

....Frame Local Vars Operand Stack args local vars object Method Vars Frame 34 made for the class methods. Measurement of the loading time is initiated before the class resources are created. 4.1. 3 Dynamic Instrumentation for VM Code Paradyn J uses Paradyn s dynamic instrumentation [27] to insert and delete instrumentation code into Java virtual machine code at any point in the interpreted execution. Paradyn s method for instrumenting functions is to allocate heap space in the application process, generate instrumentation code in the heap, insert a branch instruction from the ....

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille. Dynamic Program Instrumentation for Scalable Performance Tools. In Proceedings of the Scalable High-performance Computing Conference (SHPCC), Knoxville, Tennessee, May 1994.

....steering configuration that uses the mechanisms of Paradyn and a symbolic debugger for performance steering. 2 Paradyn as a Performance Tool Paradyn[5] is a performance tool based on a novel mechanism for automating the search for performance bottlenecks and low overhead instrumentation techniques[3]. Paradyn is also extensible any tool external to Paradyn may access the performance information that Paradyn collects via a well defined 6 visualization interface. The combination of the search mechanism, dynamic instrumentation, and the visualization interface allows us to use Paradyn as the ....

Jeffrey K. Hollingsworth, Jon Cargille, and Barton P. Miller. Dynamic Program Instrumentation for Scalable Performance Tools. In Proceedings of the 1994 Scaleable Highe Performance Computing Conference, pages 841--850, Knoxville, TN, May 1994.

....as necessarily representing the official policies or endorsements, either expressed or implied, of the Wright Laboratory Avionics Directorate or the U.S. Government 2 The best way to handle instrumentation overhead is to avoid introducing it in the first place. In a previous paper [3], we described a new approach to performance monitoring called Dynamic Instrumentation. Dynamic Instrumentation delays instrumenting an application program until it is in execution, permitting dynamic insertion and alteration of the instrumentation during program execution. This strategy of ....

....the predicted perturbation. The sum of this information for all points is the predicted cost for an instrumentation request. Metric definitions are used to enumerate what instrumentation primitives and predicates need to be inserted and where. Based on measurements of Dynamic Instrumentation [3], we know the precise cost of each instrumentation primitive and trampoline request. The difficult part is estimating the frequency of execution of each point. Application Processes Primitive Overheads Predicted Cost Observed Cost Static Event Frequency Observed Cycle Counts Super Scalar Parameter ....

[Article contains additional citation context not shown here]

Jeffrey K. Hollingsworth, Barton P. Miller, and Jon Cargille, "Dynamic Program Instrumentation for Scalable Performance Tools," 1994 Scalable High-Performance Computing Conf., Knoxville, Tenn., 1994.

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