Large-Scale Visualization of Sparse Matrices

An efficient algorithm for parallel acquisition of visualization data for large sparse matrices is presented and evaluated both analytically and empirically. The algorithm was designed to be application-independent, i.e., it works with any matrix-processors mapping and with any sparse storage format/scheme. The empirical scalability study of the algorithm was carried on using multiple modern HPC systems. In our largest experiment, we utilized 262,144 processors for 73 seconds to gather and store to a file the visualization data for a matrix with 1.17x10^13 nonzero elements. Using the proposed algorithm, one can thus visualize large sparse matrices with a minimal runtime overhead imposed on executed HPC codes.