D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. Scalable performance environments for parallel systems. In Proceedings of the Sixth Distributed Memory Computing Conference, pages 562--569, Los Alamitos, CA, Apr. 1991. IEEE Computer Society Press.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....in terms of network and host animations except now the hosts become the individual nodes of the MPP and the network is replaced by an animation of the given MPP s communication network. Examples include grid, bus, crossbar, and tree networks. Visualization tools such as ParaGraph [4] and Pablo [5] have more than 25 different views selectable by the user. These tools were developed to help parallel programmers do post mortem analysis on codes run on massively parallel processors (MPP) Many of these views are not meaningful and in 5 fact can be misleading in the heterogeneous distributed ....

D. Reed, R. Olson, R. Aydt, T. Madhyastha, T. Birkett, D. Jensen, B. Nazief, B. Totty, "Scalable Performance Environments for Parallel Systems," Proceedings of the Sixth Distributed Memory Computing Conference, IEEE Computer Society Press, April 1991.

....to performance bottlenecks and unexpected behavior. Usually these types of errors can not be found with conventional debugging systems. But the most complex 1 Funded by the German Science Foundation, Contract: SFB 0342 problem to solve when writing parallel programs is the scalability problem [22]. Programs running on massively parallel architectures consist of hundreds and thousands of interacting processes. This massive amount of processes or information in general has to be handled by the programmer. As the state of the art in parallel programming is still based on the approach of ....

Daniel A. Reed, Robert D. Olson, Ruth A. Aydt, Tara M. Madhyastha, Thomas Birkett, David W. Jensen, Bobby A. A. Nazief, and Brian K. Totty. Scalable Performance Environments for Parallel Systems. In Proc. of the Sixth Distributed Memory Computing Conference, pages 562--569, Portland,Or, April 28 - May 1 1991. IEEE.

....this by controlling multiple, independent applications (e.g. MsgView, PVMSR, ParaGraph, etc. in the figure) each of which is nominally a program that reacts to events from the trace (and presumably provides one or more related views) thus an application is analogous to a Pablo analysis module [5]. The perspective manager also reads events from the trace, then distributes them to applications. 2.2 Trace File Management PICL trace files traditionally cause di#culty for analysis routines, primarily due to the nature of the instrumentation that is used to create events. The first problem ....

Daniel A. Reed and The Picasso Group. Scalable performance environments for parallel systems. In Proceedings of the Sixth Distributed Memory Computing Conference, pages 562--569, 1991.

....In the past five years, the application of visualization technologies to manage the vast amount of performance data collected in multiprocessors has also been proposed. Many tools supporting performance visualization are available today, either in the public domain (e.g. AIMS [YSM95] Pablo [RO 91], Paradyn [MC 95] and ParaGraph [HE91] or as part of a multiprocessors system software (e.g. CXTRACE on Convexs SPP 1, ParAid on Intels Paragon, MPP Apprentice on CRIs T3D, PRISM on TMCs CM5, and PV on IBMs SP2) Typically, performance data are displayed post mortem. These displays provide ....

....Our studies indicated that trace record merging resulted in a reduction of 27 of the number of trace records, which translated to an average reduction of 38 in actual trace file length in ASCII. Binary trace encoding is the second technique to reduce trace file size. Researchers working on Pablo [RO 91] reported 40 savings with the use of binary (vs. ASCII) representation. Our own calculations also indicated similar savings (from 40 to 50 ) In summary, the trace file would only shrink by a factor of 4 even if both techniques were applied simultaneously. This was insufficient because ....

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. Scalable Performance Environments for Parallel Systems. In Proceedings of the 6th Distributed Memory Computing Conference. April 1991.

....fairly rapid evaluation of a code s scalability and the possibility of monitoring network behavior albeit simulated behavior in a way that is not normally possible on actual codes. An interesting prospect is the integration of such a tool within a performance monitoring system such as Pablo [16]. Several other projects use direct execution of application processes to drive simulations of multiprocessor systems. Among these we find two pertinent characteristics, i) the type of network being simulated, and (ii) whether the simulation is itself parallelized. Table 1 uses these attributes ....

D.A. Reed, R.D. Olson, R.A. Aydt, T.M. Madhyastha, T. Birkett, D.W. Jensen, B.A.A. Nazief, and B.K. Totty. Scalable performance environments for parallel systems. In Proceedings of the Sixth Distibuted Memory Computing Conference, pages 562--569, 1991.

.... features together withthe2428 5 message passing toolkit function wrapper library (described in the lastsection) 489 for analysis of a CRAY T3E message passing program in a much greater detail than it would bepossible760 other currently available performance tools (e.g. AIMS [2] or Pablo [3]) By using the default wrappers for MPI or PVM functions, but the special wrappers for SHMEM routines, wecan994 inside the former and can determine how they are implemented the necessarycommunication 43610 in terms of the latter. Two small examples will demonstrate this. As a first example, we ....

....make PAT a powerful and unique event trace analysis tool: 1. It s object instrumentation allows for analysis of programs written inany30 577 currently available on the T3E (Fortran, C, and C ) and is extensible toother 56440 Other currently available performance tools (e.g. AIMS [2] or Pablo [3]) use source code instrumentation and only support Fortran77 and or C. Paradyn [6] also can instrument object code but is not available for the CRAY T3E. In addition, PAT caneasily 52450 mixed language applications and can instrument system or 3rd party library code,41261 which are not ....

D.A. Reed, R.D. Olson, R.A. Aydt, T.M. Madhyasta, T. Birkett, D.W. Jensen, A.A. Nazief, B.K. Totty, Scalable Performance Environments for Parallel Systems, in: Proc. 6th Distributed Memory Computing Conference, pages 562-569, IEEE Computer Society Press, 1991

....computation events in an application, without having to go through the above contortions, would be helpful. In addition, it would be helpful if the analysis of communication overhead (communication not overlapped with any computation) could be done automatically using the trace information. Pablo [26, 27] is a portable, scalable, and extensible performance environment being developed at the University of Illinois, Urbana. The goals of the project are to create a performance data analysis environment for both SIMD and MIMD distributed memory parallel systems. Pablo consists of two components: ....

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty, Scalable Performance Environments for Parallel Systems (1991).

....non negligible for software instrumentation, we argue that we can characterize and compensate for its effects and infer the original program behavior from the tracefile. There are two possible mechanisms for software instrumentation: i. instrumented system software [2] or ii. event recorders [7, 5]. The use of instru mented versions of communication libraries and operating systems is most convenient since performance data can be obtained without modifying the application code. However, this requires vendor participation and does not provide an easy mechanism for the user to turn monitoring ....

....of compiler effectiveness and the utilization of hardware monitors . Flexibility: Application specialists want different questions answered for the same data set. It is impossible to build a visualization analysis toolkit that will make everybody happy. AIMS tracefiles are readable by Pablo [7] (which provides users a mechanism to customize the analysis toolkit) and ParaGraph [1] which provides over 30 animated views) We hope that performance data gathered using AIMS will eventually help us build more accurate models of parallel programs machines and predict their scalability ....

D.A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. "Scalable Performance Environments for Parallel Systems." In Proceedings of the 6 th Distributed Memory Computing Conference . Apil 1991

....execution characteristics. In the past five years, visualization technologies has been applied to manage the vast amount of performance data collected in multiprocessors. Many tools supporting performance visualization are available today, either in the public domain (e.g. AIMS [YSM95] Pablo [RO 91], Paradyn [MC 95] and ParaGraph [Hea93] or as part of a multiprocessor s system software (e.g. CXTRACE on Convex s SPP 1, ParAid on Intel s Paragon, MPP Apprentice on CRI s T3D, PRISM on TMC s CM5, and PV on IBM s SP2) Typically, performance data are displayed postmortem. These displays ....

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. "Scalable Performance Environments for Parallel Systems." In Proceedings of the 6th Distributed Memory Computing Conference. April 1991.

....increase in complexity of inter processor interactions and advancements in load balancing strategies for such codes, performance tools that expose interactions within each partition and across different partitions are essential for developing efficient code. Several performance tools (e.g. Pablo [3], ParaGraph PICL[4,5] IPS 2 [6] PRISM on the CM 5 [7] and ParAide on the Paragon) have been available for monitoring SPMD programs on MIMD Multiprocessors for some time. However, none of them meets all of the requirements for a performance tool applicable in this context which include: ....

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. "Scalable Performance Environments for Parallel Systems." In Proceedings of the 6 th Distributed Memory Computing Conference. April 1991

....monitoring visualization tools. These include ParAide for Intel s Paragon, MPP Apprentice for CRI s T 3D, PRISM for TMC s CM 5, PV for IBM s SP2 and CXTrace for Convex s SPP1. Many public domain tools have also emerged from various national laboratories and universities including UIUC s Pablo [RO 91], ORNL s XPVM [Kohl95] and NASA s AIMS [Yan94, YSM94] A few observations can be made: Tools provided by the vendors only run on their own machines (as expected) Compounded with unreliable system software, a lot of these sophisticated performance tools are glanced over by many users who prefer ....

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. "Scalable Performance Environments for Parallel Systems." In Proceedings of the 6 th Distributed Memory Computing Conference. April 1991.

....input output; however, in depth study of application level input output is also necessary to understand programmer input output requirements. Many researchers have addressed this problem, studying the application behavior of a variety of scientific parallel applications. Using the the Pablo [73, 74] performance analysis software, researchers have characterized the application level behavior of a wide variety of parallel applications on the Intel Paragon 10 XP S. Application areas include modeling of electron molecule collisions, a 3 D numerical simulation of the Navier Stokes equations, an ....

Reed, D. A. Scalable Performance Environments for Parallel Systems. In Proceedings of the Sixth Distributed Memory Computing Conference, to appear (Apr. 1991).

....PARADISE [10] In that tool, users wrote code to define objects that modeled various program elements and compiled them into a display program. The program lets users graphically define the interactions among these elements and then feed in data from a program log to animate the display. Pablo [11] is an environment for performance monitoring. It defines a standard, flexible format in which programs can record data about their execution, and tools built in this environment can summarize and present this data in a variety of forms. Similarly, IPS 2[12] records several kinds of performance ....

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty, "Scalable performance environments for parallel systems," technical report, University of Illinois, March 1991.

....interact with the running program through Voyeur or query its state. IMPROV [5] lets users define debugging views by providing a graphical editor and a language for defining the connection between events in the execution of a program and the desired transformations in a graphical display. Pablo [11] offers a framework for integrating a number of parallel performance analysis tools, including novel tools for presenting data, such as sonification and headmounted displays. It also provides a flexible format for program trace data. All of these tools offer the user important benefits, and there ....

REED, D. A., OLSON, R. D., AYDT, R. A., MAD- HYASTHA, T. M., BIRKETT, T., JENSEN, D. W., NAZIEF, B. A. A., AND TOTTY, B. K. Scalable performance environments for parallel systems. Technical report, University of Illinois, March 1991.

....assigned to active and idle states on various processors combine to create characteristic audio patterns that users can recognize as they analyze the behavior of their programs. Reed and his colleagues have experimented with a variety of media for presenting performance data using their Pablo [47] environment. 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 ....

Reed, D. A., Olson, R. D., Aydt, R. A., Madhyastha, T. M., Birkett, T., Jensen, D. W., Nazief, B. A. A., and Totty, B. K. Scalable performance environments for parallel systems. Technical report, University of Illinois, March 1991. 130

....tool is the incorporation of a 3 D spreadsheet to analyse data between program runs. For example, it is possible use the 3 D spreadsheet package to compare the run times for different program runs and display them in a bar chart. Halstead s Vista system [8] based on an earlier system called Pablo [21]) provides a generic browser that can be used to analyse different types of performance data. In particular, Halstead has used Vista to analyse parallel symbolic programs [9] Vista presents the programmer with different views of the parallel execution which resemble the activity and per thread ....

D. Reed, R. Aydt, T. Madhyastha, R. Noe, K. Shields, and B. Schwartz. Scalable Performance Environments for Parallel Systems. In Sixth Distributed Memory Computing Conference (IEEE), 1991.

....of the format itself, an explanation of the C interface library, and sample files and code demonstrating the use of SDDF. 1. 1 Motivation for the Development of SDDF The Pablo group at the University of Illinois is developing an integrated performance data analysis environment called Pablo [2]. The goal of the Pablo system is to provide tools for the collection and analysis of performance data across both SIMD and MIMD distributed memory parallel systems. The data capture component of the environment produces files containing performance trace event records for the system being ....

....method setCellString must be used. For Values containing a single character string (CHARACTER type Arrays with dimension one) either the Value class string manipulation methods discussed in the previous section or the Array setCellString method may be used. 1 int strLength = 30; 2 int dimSizes[2] = 3, 20 ; 3 4 Array oneString = new Array( CHARACTER, 1 ) 5 oneString setDimSizes( strLength ) 6 oneString setCellString( Here is my character string. 7 8 Array moreStrings = new Array( CHARACTER, 2 ) 9 moreStrings setDimSizes( dimSizes ) 10 moreStrings setCellString( ....

[Article contains additional citation context not shown here]

Reed, D. A., Olson, R. D., Aydt, R. A., Madhyastha, T. M., Birkett, T., Jensen, D. W., Nazief, B. A. A., and Totty, B. K. Scalable Performance Environments for Parallel Systems. In Proceedings of the Sixth Distributed Memory Computing Conference (1991), IEEE Computer Society Press.

.... More information about the Pablo Self Defining Data Format can be found in [1] Information about the instrumentation system can be found in the users guides [11, 17] Information about sonification can be found in [4, 5, 6, 7] Other good documentation about the Pablo environment can be found in [3, 8, 9, 10, 12, 13, 15, 16]. In addition to the Pablo documents referenced, the OSF Motif, and X Window System manuals may be useful. 2 add 2 modules load configuration delete graph add module(s) delete module(s) arrange module(s) Unconnected Modules connect modules runPablo Idle load layout Quit add ....

Reed, D. A., Olson, R. D., Aydt, R. A., Madhyastha, T. M., Birkett, T., Jensen, D. W., Nazief, B. A. A., and Totty, B. K. Scalable Performance Environments for Parallel Systems. In Proceedings of the Sixth Distributed Memory Computing Conference (1991), IEEE Computer Society Press.

....of the format itself, an explanation of the C interface library, and sample les and code demonstrating the use of SDDF. 1. 1 Motivation for the Development of SDDF The Pablo group at the University of Illinois is developing an integrated performance data analysis environment called Pablo [2]. The goal of the Pablo system is to provide tools for the collection and analysis of performance data across both SIMD and MIMD distributed memory parallel systems. The data capture component of the environment produces les containing performance trace event records for the system being studied. ....

....method setCellString must be used. For Values containing a single character string (CHARACTER type Arrays with dimension one) either the Value class string manipulation methods discussed in the previous section or the Array setCellString method may be used. 1 int strLength = 30; 2 int dimSizes[2] = 3, 20 ; 3 4 Array oneString = new Array( CHARACTER, 1 ) 5 oneString setDimSizes( strLength ) 6 oneString setCellString( Here is my character string. 7 8 Array moreStrings = new Array( CHARACTER, 2 ) 9 moreStrings setDimSizes( dimSizes ) 10 moreStrings setCellString( ....

[Article contains additional citation context not shown here]

Reed, D. A., Olson, R. D., Aydt, R. A., Madhyastha, T. M., Birkett, T., Jensen, D. W., Nazief, B. A. A., and Totty, B. K. Scalable Performance Environments for Parallel Systems. In Proceedings of the Sixth Distributed Memory Computing Conference (1991), IEEE Computer Society Press.

No context found.

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief, and B. K. Totty. Scalable performance environments for parallel systems. In Proceedings of the Sixth Distributed Memory Computing Conference, pages 562--569, Los Alamitos, CA, Apr. 1991. IEEE Computer Society Press.

No context found.

Reed, D. A.; R. D. Olson; R. A. Aydt; T. M. Madhyastha; T. Birkett; D. W. Jensen; B. A. A. Nazief; and Brian K. Totty. 1991. "Scalable Performance Environments for Parallel Systems." In Proceedings of the Sixth Distributed Memory Computing Conference (Apr.) IEEE Computer Society Press, Los Alamitos, CA, 562-569.

No context found.

Reed, D. A.; R. D. Olson; R. A. Aydt; T. M. Madhyastha; T. Birkett; D. W. Jensen; B. A. A. Nazief; and Brian K. Totty. 1991. "Scalable Performance Environments for Parallel Systems." In Proceedings of the Sixth Distributed Memory Computing Conference (Apr.) IEEE Computer Society Press, Los Alamitos, CA, 562-569.

No context found.

Daniel A. Reed, James Arendt, Ruth Aydt, Thomas Birkett, David Jensen, Tara Madhyastha, Bobby Nazief, Ted Nelson, Robert Olson, and Brian Totty. Scalable performance environments for parallel systems. In Sixth Distributed Memory Computing Conference, pages 562--569, Portland OR, April-May 1991.

No context found.

Daniel A. Reed, James Arendt, Ruth Aydt, Thomas Birkett, David Jensen, Tara Madhyastha, Bobby Nazief, Ted Nelson, Robert Olson, and Brian Totty. Scalable performance environments for parallel systems. In Sixth Distributed Memory Computing Conference, pages 562--569, Portland OR, April-May 1991.

No context found.

D. A. Reed, R. D. Olson, R. A. Aydt, T. M. Madhyastha, T. Birkett, D. W. Jensen, B. A. A. Nazief and B. K. Totty, `Scalable performance environments for parallel systems', Proc. 6th Distributed Memory Computing Conference, April 1991.

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