Partitioning and Scheduling Dsp Algorithms for Parallel Execution Using the Order Graph Method. (under the Direction of Dr. Partitioning and Scheduling Dsp Algorithms for Parallel Execution Using the Order Graph Method

Cato Gavin Richard Partitioning and Scheduling DSP Algorithms for Parallel Execution Using the Order Graph Method Under the direction of Dr Douglas S Reeves and Dr Winser E Alexander Recent e orts to speed up signal and image processing algorithms have focused on executing them in parallel on multiprocessor systems An important issue in achieving the maximum parallelism is how to automatically partition and schedule algorithms onto the individual processors In this research we present the Order Graph Method OGM as an e ective technique for parallel task scheduling and implement OGM in a scheduling tool to highlight the method s e ectiveness OGM is developed as a methodology to automatically partition and schedule hier archical fully speci ed ow graphs for parallel computing while considering interpro cessor communication and network topology Using OGM Digital Signal Processing DSP algorithms may be scheduled to optimize the processing rate for a speci ed number of processors to optimize the number of processors for a speci ed processing rate or to optimize both the processing rate and the number of processors The use of hierarchical ow graphs allows OGM to exploit ne to coarse grained paral lelism while minimizing the scheduling time by partitioning algorithms at the highest possible level in the hierarchy The reduction of inter processor communication by preprocessing the ow graphs minimizes the need for bi directional communication In turn the lower level of bi directional communication reduces the probability of con ii tention on the network topology Validating schedules against hardware limitations within individual processors and on the network topology guarantees the e ective ness of the schedules during physical implementation In addition OGM guarantees a parallel computing schedule by supporting two potential schedules parallel execu tion and software pipelining For either schedule OGM balances the workload of the individual processors We have implemented OGM in a scheduling program named Calypso Using Calypso a variety of signal processing and matrix operations algorithms are scheduled onto a multi processor system called the Block Data Parallel Architecture BDPA Results for scheduling a nd order IIR lter a th order FIR lter a th order lattice lter a nd order D IIR lter LU Decomposition and QR Factorization are presented For each algorithm a variety of hierarchical speci cations demonstrates OGM s ability to e ectively handle hierarchy during the scheduling process Insights into the e ects of hierarchy on scheduling image processing algorithms are identi ed and incorporated into OGM Runtime results and potential schedules for all of the algorithms are presented In addition some standard benchmarks are proposed for comparing methods of parallel task scheduling Partitioning and Scheduling DSP Algorithms for Parallel Execution Using the Order Graph Method by Gavin Richard Cato A dissertation submitted to the Graduate Faculty of North Carolina State University in partial ful llment of the requirements for the Degree of Doctor of Philosophy Department of Electrical and Computer Engineering Raleigh NC May Computer Engineering