Locality Transformations and Prediction Techniques for Optimizing Multicore Memory Performance

Abstract

Chip Multiprocessors (CMPs) are here to stay for the foreseeable future. In terms of programmability of these processors what is different from legacy multiprocessors is that sharing among the different cores (processors) is less expensive than it was in the past. Previous research suggested that sharing is a desirable feature to be incorporated in new codes. For some programs, more cache leads to more beneficial sharing since the sharing starts to kick in for large on chip caches. This work tries to answer the question of whether or not we can (should) write code differently when the underlying chip microarchitecture is powered by a Chip Multiprocessor. We use a set three graph benchmarks each with three different input problems varying in size and connectivity to characterize the importance of how we partition the problem space among cores and how that partitioning can happen at multiple levels of the cache leading to better performance because of good utilization of the caches at the lowest level and because of the increased sharing of data items that can be boosted at the shared cache level (L2 in our case) which can effectively be a prefetching effect among different compute cores. The thesis has two thrusts. The first is exploring the design space represented by different parallelization schemes (we devise some tweaks on top of existing techniques) and different graph partitionings (a locality optimization techniques suited for graph problems). The combination of the parallelization strategy and graph partitioning provides a large and complex space that we characterize using detailed simulation results to see how much gain we can obtain over a baseline legacy parallelization technique with a partition sized to fit in the L1 cache. We show that the legacy parallelization is not the best alternative in most of the cases and other parallelization techniques perform better. Also, we show that there is a search problem to determine the partitioning size and in most of the cases the best partitioning size is smaller than the baseline partition size. The second thrust of the thesis is exploring how we can predict the best combination of parallelization and partitioning that performs the best for any given benchmark under any given input data set. We use a PIN based reuse distance profile computation tool to build an execution time prediction model that can rank order the different combinations of parallelization strategies and partitioning sizes. We report the amount of gain that we can capture using the PIN prediction relative to what detailed simulation results deem the best under a given benchmark and input size. In some cases the prediction is 100% accurate and in some other cases the prediction projects worse performance than the baseline case. We report the difference between the simulation best performing combination and the PIN predicted ones as well as other statistics to evaluate how good the predictions are. We show that the PIN prediction method performs very well in predicting the partition size compared to predicting the parallelization strategy. In this case, the accuracy of the overall scheme can be highly improved if we only use the partitioning size predicted by the PIN prediction scheme and then we use a search strategy to find the best parallelization strategy for that partition size. In this thesis, we use a detailed performance model to scan a large solution space for the best parameters for locality optimization of a set of graph problems. Using the M5 performance simulation we show gains of up to 20% vs. a naively picked baseline case. Our prediction scheme can achieve up to 100% of the best performance gains obtained using a search method on real hardware or performance simulation without running at all on the target hardware and up to 48% on average across all of our benchmarks and input sizes. There are several interesting aspects to this work. We are the first to devise and verify a performance model against detailed simulation results. We suggest and quantify that locality optimization and problem partitioning can increase sharing synergistically to achieve better performance overall. We have shown a new utilization for coherent reuse distance profiles as a helping tool for program developers and compilers to a optimize program's performance.