Exploitation of control parallelism in data parallel algorithms

This paper considers the matrix decomposition A=LDL^T, as a vehicle to explore the improvement in performance obtainable through the execution of multiple streams of control on SIMD architectures. Several methods for partitioning the SIMD array are considered. Architectural support for and feasibility of using control parallelism in SIMD algorithms is briefly considered. Techniques for converting the extracted control parallelism into increased performance are illustrated via their application to the example algorithm. Analytical expressions for execution times are expressed in terms of execution times of the constituent operations. Experimental results for the various partitioning schemes based on execution traces are also presented. Timings based on MasPar MP-2 operations and extrapolated from experimental data are used to compare the various control parallel versions of the algorithm and the traditional SIMD counterpart