Abstracts:
Bin's:
The emergence of chip multiprocessors has enriched the computation resource available for large-scale scientific computing where Message Passing Interface (MPI) is the dominant programming model. Meanwhile, multi-core ar- chitecture raises new challenges to MPI applications. Memory system contention is a major problem and has been studied extensively. Recent advances in cache sharing models provide a opportunity to dissect the cause of the memory system contention, in particular, to separate the effect of MPI program parallelization from the effect of MPI task interaction.
In this task, we present two scaling models: one for MPI task partitioning, and the other for MPI cache sharing. We show that programs have different patterns in task partitioning but the patterns are consistent for the same program. The pattern can be exposed through training analysis, task normalization, and linear regression analysis. Using the 6 NAS benchmark programs, we show how well the models explain and predict the trend observed through experiments. A property of the cache sharing model is that MPI programs have the same performance with shared and with partitioned cache, which we call the neutrality conjecture. We verify the conjecture through experiments.
Zachary's:
The talk will be about Rubybob which is a an implementation of parallel programming by hints in the Ruby programming language. Paralell programming by hints is system in which the programming specifies possibly parallel regions of code and the code will execute correctly regardless of the correctness of the hints.