Many-core and PSPI: Mixing fine-grain and coarse-grain parallelism

Abstract:

Many of today's computer architectures are supported for fine-grain, rather than coarse-grain, parallelism. Conversely, many seismic imaging algorithms have been implemented using coarse-grain parallelization techniques. Sun's Niagara2 uses several processing threads per computational core, therefore the amount of memory per thread makes a strict coarse-grain approach to problems impractical. A strictly fine-grain parallelism approach can be problematic in algorithms that require frequent synchronization. We use a combination of fine-gain and coarse-grain parallelism in implementing a downward continuation based migration algorithm on the Niagara2. We show the best performance can be achieved by mixing these two programing styles.