A Framework for Analyzing the Performance of Sparse Matrix and Graph Operations

Thorough performance analysis is crucial for developing and optimizing algorithms, particularly for compute-intensive operations like those in scientific libraries such as the Basic Linear Algebra Subprograms (BLAS). While dense computations can predict performance based on dataset dimensions and strides, sparse matrix operations require more detailed analysis due to their complexity. Standard performance evaluations for sparse operations use a canonical set of matrices, but generalizing these results to new datasets is limited. This paper presents a framework to evaluate sparse matrix and graph operations by visualizing performance through parameterized graph models. It assesses different parameter sets and noise sources on performance, providing a modular and extensible approach, which includes system-wide performance considerations, allowing users to integrate various graph model generators, operation implementations, and noise types.