초록 ▼

A Veil program analyzes the source code and/or data of an existing sequential target program and determines how best to distribute the target program and data among the processing elements of a multi-processing element computing system. The Veil program analyzes source code loops, data sizes and typ

A Veil program analyzes the source code and/or data of an existing sequential target program and determines how best to distribute the target program and data among the processing elements of a multi-processing element computing system. The Veil program analyzes source code loops, data sizes and types to prepare a set of distribution attempts, whereby each distribution is run under a run-time evaluation wrapper and evaluated to determine the optimal distribution across the available processing elements.

대표청구항 ▼

1. A system for optimizing the operation of a program in a multiple-processing element computing system, the system comprising: a computer system comprising at least one processor and a plurality of processing elements;a target program written to run on the computer system;a target program written t

1. A system for optimizing the operation of a program in a multiple-processing element computing system, the system comprising: a computer system comprising at least one processor and a plurality of processing elements;a target program written to run on the computer system;a target program written to run on the computer system;a program having coded therein a means for: (a) wrapping the target program with a run-time analyzer;(b) selecting one parallelization strategy from a plurality of parallelization strategies;(c) parallelizing the target program according to the selected one parallelization strategy;(d) executing the target program on the processing elements;(e) recording execution results of the target program; and(f) if there are further parallelization strategies in the plurality of parallelization strategies, changing the selected one parallelization strategy to a next parallelization strategy from the plurality of parallelization strategies and repeat steps a-f;(g) comparing the execution results and selecting a best parallelization strategy having a best execution result;(h) parallelizing the target application using the best parallelization strategy with the best execution result;(i) analyzing a target data on which the target program is designed to operate;(j) if the target data is a contiguous dataset then: determining a number of available processing elements;copying at least part of the target program onto the available processing elements;dividing the target data into a number of data segments;executing the at least part of the target program on the available processing elements, each operating upon one of the data segments;combining the results from the available processing elements;(k) if the target data is burst data then: measuring an interval between bursts;measuring a computation time of the target program;if the interval between bursts is less than the computation time then determining the number of the available processing elements and distributing at least part of the target program over the available processing elements;if the interval between bursts is not less than the computation time then executing the program on one of the available processing elements;(l) if the target data is a continuous data stream then: determining the number of the available processing elements;duplicating at least part of the target program onto the available processing elements; anddemultiplex the target data into the number of sub-streams, where the number of sub-streams matches the number of the available processing elements. 2. The system of claim 1, wherein the plurality of parallelization strategies includes detecting a program loop. 3. The system of claim 1, wherein the run-time analyzer records types of input-output operations requested by the target program. 4. The system of claim 1, wherein the run-time analyzer is records a number of files accessed by the target program. 5. The system of claim 1, wherein the run-time analyzer is records input/output read/write patterns requested by the target program. 6. The system of claim 1, wherein the number of available processing elements is read from a configuration profile. 7. A computer implemented method for parallelizing of a target program onto a plurality of processing elements, the method comprising: (a) wrapping the target program with a run-time analyzer;(b) selecting one parallelization strategy from a plurality of parallelization strategies;(c) parallelizing the target program according to the one parallelization strategy;(d) executing the target program on the plurality of processing elements;(e) recording execution results of the target program;(f) if there are further parallelization strategies in the plurality of parallelization strategies, changing the one parallelization strategy to a next parallelization strategy from the plurality of parallelization strategies and repeating steps a-f; and(i) analyzing a target data on which the target program is designed to operate;(j) if the target data is a contiguous dataset then: determining a number of available processing elements;copying at least part of the target program onto the available processing elements;dividing the target data into a number of data segments;running the at least part of the target program on the available processing elements, each of the available processing elements operating upon one of the data segments;combining the results from the available processing elements;(k) if the target data is burst data then: measuring an interval between bursts;measuring a computational time of the target program;if the interval between bursts is less than the computational time then determining the number of the available processing elements and distributing the at least part of the target program over the number of available processing elements;if the interval between bursts is not less than the computational time then executing the program on one available processing elements;(l) if the target data is a continuous data stream then: determining the number of available processing elements;duplicating the at least part of the target program onto the available processing elements; anddemultiplexing the target data into a number of sub-streams, where the number of sub-streams matches the number of available processing elements. 8. The computer implemented method of claim 7, wherein the plurality of parallelization strategies includes detecting a program loop. 9. The computer implemented method of claim 8, wherein the execution results include a number of times the program loop iterates. 10. The computer implemented method of claim 7, wherein the run-time analyzer is adapted to measure types of input-output operations requested by the target program. 11. The computer implemented method of claim 7, wherein the run-time analyzer is adapted to measure a number of files accessed by the target program. 12. The computer implemented method of claim 7, wherein the run-time analyzer is adapted to measure input/output read/write patterns requested by the target program. 13. The system of claim 7, wherein the number of available processing elements is read from a configuration profile. 14. The system of claim 7, further comprising: if there still remains unassigned processing elements, then allocating the unassigned processing elements to assist the assigned processing elements as a sub-group. 15. A system for optimizing the operation of a program in a multiple-processing element computing system, the system comprising: (a) a computer system comprising at least one processor and a plurality of processing elements;(b) a target program written to run on the computer system;(c) a means for wrapping the target program with a means for analyzing run-time performance;(d) a means for selecting one parallelization strategy from a plurality of parallelization strategies;(e) a means for parallelizing the target program according to the one parallelization strategy;(f) a means for executing the target program one the processing elements;(g) a means for recording execution results of the target program; and(h) if there are further parallelization strategies in the plurality of parallelization strategies, a means for changing the one parallelization strategy to a next parallelization strategy from the plurality of parallelization strategies and repeat steps c-h;(i) a means for comparing the execution results and for selecting a best parallelization strategy having a best execution time;(j) a means for parallelizing the target application using the best parallelization strategy;(k) a means for analyzing a target data on which the target program is designed to operate;(l) if the target data is a contiguous dataset then: a means for determining a number of available processing elements;a means for copying at least part of the target program onto the available processing elements;a means for dividing the target data into a number of data segments;a means for running the target program on the available processing elements, each processing elements operating upon one of the data segments;a means for combining the results from the available processing elements;(m) if the target data is burst data then: a means for measuring an interval between bursts;a means for measuring a computational time of the target program;if the interval between bursts is less than the computational time then a means for determining the number of the available processing elements and a means for distributing the target program over the available processing elements;if the interval between bursts is not less than the computational time then a means for executing the program on one of the processing elements;(n) if the target data is a continuous data stream then: a means for determining the number of available processing elements;a means for duplicating the at least part of the target program onto the available processing elements; anda means for demultiplexing the target data into a number of sub-streams, where the number of sub-streams matches the number of available processing elements. 16. The system of claim 15, wherein the plurality of parallelization strategies includes a means for detecting a program loop. 17. The system of claim 15, wherein the means for analyzing run-time performance measures a number of files accessed by the target program. 18. The system of claim 15, wherein the means for analyzing run-time performance measures a type of input/output operation requested by the target program. 19. The system of claim 15, wherein the number of available processing elements is read from a configuration profile. 20. The system of claim 15, further comprising: if there still remains unassigned processing element, then allocating the unassigned processing elements to assist the assigned processing elements as a sub-group.