Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Univ. of Illinois at Urbana-Champaign, Center for Supercomputing Res. & Dev., January 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....perform matrix inversion and convolution. Simulation Analysis Figure 13 shows the results of the simulated execution of the second application phase on an ideal parallel machine. The simulator identifies the maximum degree of parallelism that is inherent an execution of the application [17, 20]. The figure shows the repetitive patterns in Seismic that follows the compute and data exchange phases. The maximum number of parallel operations found in this code is up to several million. Although this analysis does not show how realistic it is to exploit this degree of parallelism, it makes ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Univ. of Illinois at Urbana-Champaign, Center for Supercomputing Res. & Dev., January 1993.

....presented in [7, 21] It is shown in [39] how the HTG can be used for instruction scheduling in superscalar processors, in order to exploit instruction level parallelism. The motivation to exploit functional parallelism comes from measurements of the dynamic behavior of programs, such as those in [8, 29], in which large amounts of parallelism have been found beyond that available from totally parallel loops alone. In [7] however, it is shown that the amount of functional parallelism exploitable through the HTG is modest for scientific programs (in the order of 2 to 4) Another effort to ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Department of Computer Science, University of Illinois at UrbanaChampaign, January 1993.

.... for granularity control, were defined in [21] Efficient implementation schemes for the scheduling operations and algorithms for optimizing the HTG are presented in [6] The motivation to exploit functional parallelism comes from measurements of the dynamic behavior of programs, such as those in [7, 19], in which large amounts of parallelism have been found beyond that available from totally parallel loops alone. Another effort to exploit functional parallelism is described in [23] where scheduling and allocation algorithms for the exploitation of functional parallelism on distributed memory ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Department of Computer Science, University of Illinois at Urbana-Champaign, January 1993.

....collects and combines information from various sources. Timing information is gathered from instrumented program runs. The tool performing this instrumentation is a Polaris based utility, not discussed further in this paper [Eig93] Maximum parallelism estimates are supplied by the Max P tool [Pet93, KE97] Information about which loops are serial or parallel is provided by the actual Polaris compiler. The Ursa Minor tool includes a subroutine and loop nest structure analyzer, also implemented using the Polaris infrastructure. In the current implementation, these information sources are ....

....loop table view, each line displaying information for an individual loop. Currently, the table displays information such as timing results from various program runs, the number of invocations of each loop, the parent in the nest structure, and the maximum degree of parallelism provided by Max P [Pet93, KE97] It also indicates whether a loop is serial or parallel as detected by Polaris. If it is serial, the reason given by the compiler can be displayed on mouse clicking. In Figure 2, Figure 2: Loop Table View of the Ursa Minor tool. the user has clicked on loop RESTAR do560 to see the reason ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Univ. of Illinois at Urbana-Champaign, Center for Supercomputing Res. & Dev., January 1993.

....and the predictions of behaviors that may be undetected by static analysis methods. Of great interest is understanding the dynamic behavior of parallelism, one of the most dominant factors of performance. Max P is a Polaris based tool, originally developed at the University of Illinois [KE97, Pet93] It evaluates the inherent parallelism of a program at runtime. The inherent parallelism is defined as the ratio of the total number of operations in a program, or program section, to the number of operations along the critical path. The critical path is the longest path in the program s ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Univ. of Illinois at Urbana-Champaign, Center for Supercomputing Res. & Dev., January 1993.

....easy comparison. Figure 2 shows the table, each line being the name of an individual loop. Currently, the table displays timing results from various program runs, the number of invocations of each loop, the parent in the calling structure, and the maximum degree of parallelism provided by Max P [Pet93, KE97] It also indicates whether a loop is serial or parallel as detected by the parallelizing compiler. If it is serial, the reason given by the compiler can be displayed on mouse clicking. In Figure 2, the user is viewing the database of program ARC2D and its performance on a SGI Power ....

....Minor tool collects and combines information from various sources. Timing information is gathered from instrumented program runs. The tool performing this instrumentation is a Polaris based utility, not discussed further in this paper. Maximum parallelism estimates are supplied by the Max P tool [Pet93, KE97] Information on which loops are serial or parallel is provided by the actual Polaris compiler. The Ursa Minor tool includes a subroutine and loop calling structure analyzer, also implemented using the Polaris infrastructure. The tool enables the collection of the described information for ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Univ. of Illinois at Urbana-Champaign, Center for Supercomputing Res. & Dev., January 1993.

....Minor tool collects and combines information from various sources. Timing information is gathered from instrumented program runs. The tool performing this instrumentation is a Polaris based utility, not discussed further in this paper. Maximum parallelism estimates are supplied by the Max P tool [Pet93, KE97] Information on which loops are serial or parallel is provided by the actual Polaris compiler. The Ursa Minor tool includes a subroutine and loop calling structure analyzer, also implemented using the Polaris infrastructure. In the current implementation, these information sources are ....

....table view, each line displaying information for an individual loop. Currently, the table displays information such as timing results from various program runs, the number of invocations of each loop, the parent in the calling structure, and the maximum degree of parallelism provided by Max P [Pet93, KE97] It also indicates whether a loop is serial or parallel as detected by Polaris. If it is serial, the reason given by the compiler can be displayed on mouse clicking. In Figure 2, the user has clicked on loop RESTAR do560 to see the reason inhibiting parallelization. Whenever additional ....

Paul Marx Petersen. Evaluation of Programs and Parallelizing Compilers Using Dynamic Analysis Techniques. PhD thesis, Univ. of Illinois at Urbana-Champaign, Center for Supercomputing Res. & Dev., January 1993.

.... form of parallelism in programs is nonloop or functional parallelism [42, 85, 86] While data parallelism can provide large amounts of parallelism that often scales with problem size, recent studies have suggested that a significant amount of nonloop parallelism also exists in numerical codes [22, 82]. Theoretical foundations and methods for detecting functional parallelism automatically in a compiler were laid down in previous studies [42] The focus of this thesis, by contrast, is the exploitation of functional parallelism at run time once it has been detected by the compiler. This, too, ....

....loop carried dependences, but has the added advantage of synchronizing functional parallelism within each loop body as well. Measurements of nonloop parallelism have been made by several researchers by using a compiler to insert source code which calculates a program s critical path at run time [22, 82]. In some cases, large amounts of parallelism have been found beyond that available from doall loops alone. A drawback of this approach, however, is that it is not based on a structured analysis of the program, and hence the measurements give little indication of the source of this nonloop ....

Paul Marx Petersen. Evaluation of programs and parallelizing compilers using dynamic analysis techniques. PhD thesis, University of Illinois at Urbana-Champaign, Center for Supercomputing Research and Development, January 1993.

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