A model for predicting the performance of sparse matrix vector multiply (SpMV) using memory bandwidth requirements and data locality
MetadataShow full item record
Sparse matrix vector multiply (SpMV) is an important computation that is used in many scientific and structural engineering applications. Sparse computations like SpMV require the use of special sparse data structures that efficiently store and access non-zeros. However, sparse data structures can tax the memory bandwidth of a machine and limit the performance of a computation, which for these computations is typically less than 10% of a processor's peak floating point performance. The goal of this thesis was to understand the trade-off between memory bandwidth needs and data locality for SpMV ...