J. M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....To achieve performance in multi user grid environments, adaptive scheduling of these components is critical. As part of the AppLeS project, we have found that we can achieve good schedules by using dynamic system resource information in application models which predict application performance [8, 26, 27, 28, 29]. Even for the simple client server model in Figure 1, effective adaptive scheduling may be non trivial. For example, for the model shown, the scheduler must determine how these operations can be split between client and server. To illustrate, suppose the user of an SRB geographical image server ....

J. M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998. Also available as UCSD CS Dept. Technical Report, Number CS98-607, http://www.cs.nwu.edu/jms/Thesis/ thesis.html.

....of potential performance. One insight that has emerged from the AppLeS experience is that performance predictions and parameters can be more accurately represented as distributions, and that scheduling policies can be developed which utilize the extra information provided by distributions [Sch98] Figure 3 shows the range of execution time values exhibited by a distributed red black Successive Over relaxation (SOR) application during a set of experi5 Figure 2: Plot of actual measurements for back to back Jacobi2D application executions scheduled using a static block decomposition and a ....

....not use resource variance information to make decisions. The figure shows that for the given set of experiments, the conservative scheduling policy avoids the dramatically poor performance sometimes exhibited by the non conservative policy, but in addition, can sometimes be too conservative [Sch98] Meta information was also used in [SW98] in the form of an estimation of prediction accuracy (automatically provided by the NWS) to gain a factor of 2 in execution performance on a wide area Computational Grid. These preliminary experiments indicate that meta information can provide important ....

J. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998.

No context found.

J. Schopf, Performance Prediction and Scheduling for Parallel Applications on Multi-user Clusters, Ph.D. thesis, University of California, San Diego, December 1998.

No context found.

J. Schopf, "Performance Prediction and Scheduling for Parallel Applications on Multiuser Clusters," Ph.D. thesis, Univ. of California, San Diego, Dec. 1998

No context found.

Schopf, J.M., Performance Prediction and Scheduling for Parallel Applications on Multi-User Cluster. Department of Computer Science and Engineering, University of California San Diego, San Diego, 1998, pp. 247.

....o line mappings of the graph onto the resources. SEA [62] uses its data ow style program graph to 27 determine which tasks are ready for execution. By contrast, AppLeS assumes that the performance model is provided by the user. Current AppLeS applications rely on structural performance models [67], which compose performance activities into a prediction of application performance. Almost all the aforementioned projects do not provide much latitude for user provided scheduling policies or performance criteria: minimization of execution time is the only goal. Scheduling in MARS [60] and Dome ....

J. Schopf, Performance Prediction and Scheduling for Parallel Applications on Multi-user Clusters, Ph.D. thesis, University of California, San Diego, December 1998.

....at any given time their performance may di er radically. The amount of data to be assigned to each machine in this case is no longer a simple decision, but will depend on the goals of the user or application developer and the optimization function of the scheduling policy, such as described in [Sch98] 1.2 Outline In this paper we de ne an extension to structural modeling that allows for model parameters that are stochastic values represented by an upper bound and a lower bound. Section 2 gives a brief overview of the structural modeling approach. Section 3 describes how intervals are de ....

....describe the application behavior, as shown experimentally in Section 5. Section 5.6 compares this approach to previous work that used normal distributions to represent stochastic values. We conclude and give future work in Section 7. 2 Overview of Structural Modeling In previous work [Sch97, Sch98] we developed a technique called structural modeling as an approach to accurate performance modeling. Structural modeling uses the functional structure of the application to de ne a set of equations that re ect the time dependent, dynamic mix of application tasks occurring during execution in a ....

[Article contains additional citation context not shown here]

Jennifer M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998. Also available as UCSD CS Dept. Technical Report, Number CS98-607, http://www.cs.nwu.edu/~jms/Thesis/thesis.html. 11

....results, and summary. 2 Overview of Structural Modeling In order to predict an application s behavior on shared resources, we need both good models of the implementation and accurate information about the application and system with which to parameterize the models. In previous work [Sch98] we developed a technique called structural modeling that uses the functional structure of the application to define a set of equations that reflect the time dependent, dynamic mix of application tasks occurring during execution in a distributed parallel environment. A structural model consists ....

....The primary source of varying behavior for this data was the available CPU values for the contended system. This data was supplied by the Network Weather Service (NWS) Wol97] an online tool that provides a time series of data for CPU availability, bandwidth and latency. In previous experiments [Sch98] we had determined the amount of data from the time series to use in our predictions, namely a 5 minute window. This had been shown to be no worse than using larger or smaller windows of data. We experimentally examined the accuracy of the histogram for a range of interval sizes and measured ....

[Article contains additional citation context not shown here]

Jennifer M. Schopf. Performance Prediction and Scheduling for Parallel Applications on MultiUser Clusters. PhD thesis, University of California, San Diego, 1998.

No context found.

Jennifer M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998.

....application behavior, as shown experimentally in Section 4. Section 4.5 compares this approach to previous work that used normal distributions to represent stochastic values, and other related work. We conclude and give future work in Section 6. 2 Overview of Structural Modeling In previous work [13, 14] we developed a technique called structural modeling as an approach to accurate performance modeling. Structural modeling uses the functional structure of the application to define a set of equations that reflect the time dependent, dynamic mix of application tasks occurring during execution in a ....

....network of Sparc workstations located in the UCSD Parallel Computation Lab. The machines are connected with a mix of 10 Mbit (slow) and 100 Mbit (fast) ethernet over an Intel switch, and all run Solaris. Both the CPU and the network were shared with other users. Additional experiments are given in [14]. 4.1 Performance Metrics To evaluate the stochastic predictions, we use two different performance metrics. The first we call capture. This measures the percentage of actual values falling within the predicted range for a set of application runs. This statistic is important because the primary ....

[Article contains additional citation context not shown here]

J. M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998. Also available as UCSD CS Dept. Technical Report, Number CS98-607, www.cs.nwu.edu/jms/Thesis/thesis.html.

....approximately 70 of the values and plus or minus two standard deviations captures approximately 95 of the values. 5 between 30 and 90 seconds. For each set of experiments we show approximately 25 runs, but the experimental data is consistent with larger runs on similarly loaded platforms [Sch98] The experiments were run on UCSD s Parallel Computation Laboratory (PCL) and Linux clusters. The PCL cluster consists of 4 heterogeneous Sparc workstations connected by a mixture of slow and fast ethernet; the Linux Cluster consists of 4 PCs running Linux connected by fast ethernet. Both ....

....variation when the CPU values varied over half the range. A sample of the actual CPU values during the scheduling execution times are given for the runs shown in Figure 4 in graphs 5 through 8. For the rest of the experiments only summaries are given, although the full data sets may be found in [Sch98] 3.1 Performance Metrics Each of the following graphs show the actual execution times using the three scheduling approaches, run one after the other so that each run will experience approximately equivalent system conditions. We consider a scheduling policy to be better than others if it ....

Jennifer M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998. Also available as UCSD CS Dept. Technical Report, Number CS98-607, http://www.cs.nwu.edu/jms/Thesis/thesis.html.

....In Section 5, we compare these two representations with respect to the accuracy of the resulting performance predictions using several applications in various multi user environments. Related work is presented in 6. We conclude in Section 7. 2. Overview of Structural Modeling In previous work [29, 30] we developed a technique called structural modeling that uses the functional structure of the application to de ne a set of equations to re ect the dynamic mix of application tasks occurring during execution in a distributed parallel environment. A structural model consists of a top level model, ....

....in bytes per second between processor x and processor y, available from a priori benchmarks or a dynamic measurement system. Unpack(x) Time to unpack a byte for processor x. Structural models provide an e ective framework in which to model the performance of distributed parallel applications [29, 30]. In the following we assume that accurate structural models are available to the user. We discuss the possible repercussions on performance should the structural model itself not provide accurate predictions as part of Section 3.3 and with the experimental results in Section 5. In the next ....

[Article contains additional citation context not shown here]

Schopf, J. M. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, San Diego, 1998. Also available as UCSD CS Dept. Technical Report, Number CS98-607, http://www.cs.nwu.edu/~jms/Thesis/thesis.html.

No context found.

J. M. Schopf. Performance Prediction and Scheduling for Parallel Applications on Multi-User Clusters. PhD thesis, University of California, 1998.

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