Fredrik Vraalsen, Ruth A. Aydt, Celso L. Mendes, and Daniel A. Reed. Performance contracts: Predicting and monitoring grid application behavior. In Proceedings of GRID 2001.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....performance expectations are. For example, they might specify a flexible desired turnaround time or a hard deadline for application completion. These specifications are used to create a performance contract, which formally specifies program performance expectations. More details are available in [30, 51]. vices of the Metacomputing Directory Service (MDS) 11, 13] and the Network Weather Service [57, 58] The NWS and MDS are centralized information servers that provide dynamic and static information about the current state of Grid resources. Since these interfaces are well known standards for ....

....II.C Current status The GrADSoft system architecture described in Section II.A is an ongoing design e#ort that is continually evolving as the GrADS project evolves. E#orts are underway to develop each of the new component technologies that will be required to realize the GrADSoft design [7, 23, 51, 3, 29, 56, 37]. While the individual GrADSoft components such as the compiler and the performance monitor are fundamental to the success of the GrADSoft system, these components must be able to interact and coordinate information flow and decision procedures. To this end, there is also an e#ort to prototype a ....

Frederik Vraalsen, Ruth A. Aydt, Celso L. Mendes, and Daniel A. Reed. Performance Contracts: Predicting and monitoring grid application behavior. In Proceedings of the 2nd International Workshop on Grid Computing, Nov 2001. 7, 11

....of the intermediate representation code. Next, the executable is launched on the selected grid resources and a real time performance monitor is used to track program performance and detect violation of performance guarantees. Performance guarantees are formalized in a performance contract [33]. In the case of a performance contract viola tion, the rescheduler is invoked to evaluate alternative schedules. 2.1 Grid technology The previous section outlined the GRADS vision; as components of that vision are prototyped and validated, they are incorporated in a shared prototype ....

....a common metric is iteration number. Alternatively, the sensors can use PAPI [22] to access measurements from the processor s hardware performance counters (i.e. flop rate) and the MPICH profiling library for communication metrics (i.e. bytes sent per second) The performance contracts [33] specify predicted metric values and tolerance limits on the difference between actual and predicted metric values. For example, contracts can specify predicted execution times and tolerance limits on the ratio of actual execution times to predicted execution times. The contract monitor receives ....

Frederik Vraalsen, Ruth A. Aydt, Celso L. Mendes, and Daniel A. Reed. Performance Contracts: Predicting and monitoring grid application behavior. In Proceedings of the 2rid International Workshop on Grid Computing, Nov 2001.

.... steering parameters (e.g. 2] 2) use measured or user provided estimates of application execution time as a function of system configuration, to heuristically adapt system resource management policies or application configuration parameters, automatically at runtime [17] 6] 14] 16] 20][27], or (3) use very simple models that compute execution time for alternative configurations [3] 8] 21] 23] 24] 25] from userprovided or measured deterministic processing and communication requirements per task, possibly adjusted for runtime measured node processing capacities and available network ....

.... adaptive control algorithms, Ribler et al. 20] describe the Autopilot system that filters data from instrumented client tasks (e.g. to characterize the dominant file access pattern) and adapts the system resource management policies (e.g. file prefetch policy) The related GRADS Project work [27] measures the application signature (i.e. processing, I O, and communication cycles as a function of time) and uses fuzzy logic to determine whether the rates for each task are within ranges defined by user specified performance contracts . Algorithms for changing the configuration when the ....

Vraalsen, F., R. A. Aydt, C. L. Mendes, and D. A. Read, "Performance Contracts: Predicting and Monitoring Grid Application Behavior", Proceedings of the 2 "a lnt'l. Workshop on Grid Computing, November 2001.

....on a Grid application development environment is the most relevant here. They propose an environment in which an application is continuously adapted at runtime to changes in Grid resources. They have described the use of perfor mance contracts for guiding dynamic reconfiguration and optimization [28]. They 25 Comparison of synchronous and asynchronous adaptation with MPIPOV Synchrorous Asynchronous 25O . 200 150 lOO 5o I I I 110 112 I Fig. 7. Comparison of synchronous and asynchronous adaptation for MPIPOV plan to incorporate some form of compiler support for suitable ....

F. Vraalscn, R.Aydt, C.Mcndcs, D. Reed, Performance Contracts: Predicting and Monitoring Grid Application Behavior, Tech. rcp., Computer Science Department, Univ. of Illinois at Urbana-Champaign (2001).

....on runtime techniques to judge if an assertion has met its expectation. Perl s work checked for these properties in event logs, 11 not in the application at runtime. The GRADS project (http: nhse2.cs. rice.edu grads ) is addressing issues of application performance and performance contracts [17] on computational grids. In a different effort, work by the APART consortium has culminated in a performance property specification language: ASL. ASL allows developers to write complex properties describing patterns in performance data, but current implementations do not allow users to plant ....

F. Vraalsen, R.A. Aydt, C.L. Mendes, and D.A. Reed, "Performance Contracts: Predicting and Monitoring Grid Application Behavior," Proc. GRID 2001.

.... (GRADS) project is broad in scope and is investigating many aspects of software support for grid applications [5] They have proposed the use of performance contracts for guiding dynamic reconfiguration and optimization strategies by comparing predicted and observed application performance [27]. They have also developed Grid enabled prototypes of part of the ScaLAPACK library [23] and the Cactus numerical simulation framework [3] both of which use performance models to select the best resources for a particular run (and they aim to support dynamic optimization guided by performance ....

F. Vraalsen, R.Aydt, C.Mendes, and D. Reed. Performance Contracts: Predicting and Monitoring Grid Application Behavior. Technical report, Computer Science Department, Univ. of Illinois at Urbana-Champaign, 2001.

No context found.

F. Vraalsen, R. A. Aydt, C. L. Mendes, and D. A. Reed. Performance contracts: Predicting and monitoring grid application behavior. In Lecture Notes in Computer Science, volume 2242, pages 154--165. Springer Verlag, November 2001.

....changing and heterogeneous nature of resource availability on the Grid makes it necessary to develop performance monitoring and adaptation mechanisms to enable better use of resources. To achieve this goal, a core component in runtime support for the Grid is the performance contract framework [2] [17]. In this framework, each application has a performance contract that may specify performance expectations, cost benefit limits, scheduling constraints, and other requirements. When the runtime system loads the application, the corresponding contract is submitted to resource brokers to enable ....

....disadvantage is that a signature created by historical performance data gives little insight about how changes in input parameters will influence the performance. Therefore, we have to create di#erent signatures for di#erent input parameters. 1.1. 2 Performance Contracts A performance contract ([17]) contains two essential notions: Specification. Each application is associated with a performance contract which specifies that, given a set of resources with certain characteristics and for particular problem parameters, an application will achieve the specified performance during its ....

Vraalsen, F., Aydt, R., Mendes, C., and Reed, D. A. Performance Contracts: Predicting and Monitoring Grid Application Behavior. In Proc. 2nd International Workshop on Grid Computing, 2001

....measurements from one execution may not be representative of future executions, nor may they highlight critical bottlenecks. This is even more true for emerging applications that execute atop computational Grids [1] where resource availability and performance can change in unexpected ways [2, 11]. Event tracing is the standard approach for obtain # This work was supported in part by the National Science Foundation under grants NSF EIA 99 72884 and ASC 97 20202; by the Department of Energy under contracts DOE W 7405 ENG 36, LLNL B341494, DOE SciDAC DEFC0201ER41205, and LLNL B505214; and ....

....the notion of performance contracts. 2.2 Performance Contracts The motivation for performance signatures is quantitative comparison and validation of application execution dynamics across platforms, input data sets, and configurations. In this model, each application has a performance contract [11] that may specify performance expectations, scheduling constraints, and other requirements. We envision application signatures being used as shown in Figure 2. The application signature module accepts real time performance data and a performance signature from previous executions, compares ....

Vraalsen, F., Aydt, R., Mendes, C., and Reed, D. A. Performance Contracts: Predicting and Monitoring Grid Application Behavior. Proc. 2nd International Workshop on Grid Computing, 2001

No context found.

Fredrik Vraalsen, Ruth A. Aydt, Celso L. Mendes, and Daniel A. Reed. Performance contracts: Predicting and monitoring grid application behavior. In Proceedings of GRID 2001.

No context found.

F. Vraalsen, R. Aydt, C. Mendes, and D. Reed, "Performance contracts: Predicting and monitoring grid application behavior," in Proceedings of the 2nd International Workshop on Grid Computing(GRID 2001.

No context found.

F.Vraalsen,R.Aydt,C.Mendes,D.Reed.Performance contracts: predicting and monitoring grid application behaviour.TR,Comp.Sc.Dept,Univ.Illinois at Urbana-Champaign, 2001.

No context found.

Vraalsen, F., et al. Performance Contracts: Predicting and Monitoring Grid Application Behavior.inSubmitted to the 2nd International Workshop on Grid Computing. 2001. Denver CO.

No context found.

Frederik Vraalsen, Ruth A. Aydt, Celso L. Mendes, and Daniel A. Reed. Performance Contracts: Predicting and monitoring grid application behavior. In Proceedings of the 2nd International Workshop on Grid Computing, Nov 2001.

No context found.

F. Vraalsen, R. Aydt, C. Mendes, and D. Reed. Performance contracts: Predicting and monitoring grid application behavior. In Proceedings of the 2nd International Workshop on Grid Computing/LNCS (GRID 2001.

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