Artifacts for "Concurrent GPU Communications in Pipelined Preconditioned CG: Beyond Traditional Overlap"

This dataset contains computational artifacts related to the paper: “Concurrent GPU Communications in Pipelined Preconditioned CG: Beyond Traditional Overlap” The paper describes computational experiments that were conducted to evaluate the performance of multi-GPU iterative linear solver based on the pipelined preconditioned conjugate gradient (CG) method. The computational artifacts are located in several subdirectories: 'aCG/' contains the source code for the aCG software. This codebase implements the two- and three-stream multi-GPU Pipelined Preconditioned CG solver, which form the foundation of the performance benchmarks evaluated in this paper. 'partitions/' contains input files related to partitioning and distributing matrices that were used in the experiments. Partitions were computed using METIS (Karypis and Kumar, 1998), a multilevel graph partitioner. From the SuiteSparse Collection (Davis and Hu, 2011), six matrices were selected: Bump_2911, Geo_1438, Flan_1565, Queen_4147, Serena and audikw_1. For each matrix, partitions are provided for 4, 8, 16, 32 and 64 parts. 'scripts/' contains job scripts for submitting jobs on two clusters: MareNostrum 5 and Leonardo supercomputers. These scripts carry out performance measurements for the multi-GPU CG solver in aCG and PETSc, and were used to collect the results