Clustered Data Parallelism

Many data layout optimizations cluster data accesses and memory into high-locality groups in order to optimize for the memory hierarchy. In this paper, we demonstrate that similar clustering program transformations enable efficient vectorization. We call this approach clustered data parallelism (CDP). CDP enables fast and power-efficient parallelism by partitioning a data structure into clusters such that SIMD evaluation is efficient within a cluster. We describe the CDP latent in three common computational patterns: map, reduce, and graph traversals. Demonstrating the benefits of CDP, we present case studies of instantiating the CDP patterns in order to design fast and power-efficient binary search and webpage layout algorithms. First, we increase binary search SIMD scalability by using CDP to expose speculative parallelism. Second, we achieve the first SIMD webpage layout algorithm by using CDP to eliminate heavy branching. We report strong performance improvements. Targeting AVX, we see a 5.5X speedup and 6.9X performance/Watt increase over FAST, the previously fastest SIMD binary search algorithm. Running webpage layout with SSE4.2 instructions, we observe a 3.5X speedup and 3.6X performance/Watt increase over an already optimized baseline. [Copyright notice will appear here once ’preprint’ option is removed.] Clustered data parallelism 1 2012/2/1