This paper presents two approaches using a Block Low-Rank (BLR) compressiontechnique to reduce the memory footprint and/or the time-to-solution of the sparse supernodalsolver PaStiX. This flat, non-hierarchical, compression method allows to take advantage of thelow-rank property of the blocks appearing during the factorization of sparse linear systems, whichcome from the discretization of partial differential equations. The first approach, called MinimalMemory, illustrates the maximum memory gain that can be obtained with the BLR compressionmethod, while the second approach, called Just-In-Time, mainly focuses on reducing the com-putational complexity and thus the time-to-solution. Singular Value Decomposition (SVD) andRank-Revealing QR (RRQR), as compression kernels, are both compared in terms of factorizationtime, memory consumption, as well as numerical properties. Experiments on a single node with24 threads and 128 GB of memory are performed to evaluate the potential of both strategies. Ona set of matrices from real-life problems, we demonstrate a memory footprint reduction of up to 4times using the Minimal Memory strategy and a computational time speedup of up to 3.5 timeswith the Just-In-Time strategy. Then, we study the impact of configuration parameters of theBLR solver that allowed us to solve a 3D laplacian of 36 million unknowns a single node, while thefull-rank solver stopped at 8 million due to memory limitation.
展开▼
