Artifact for "Triple Overlap in Pipelined Krylov Methods on Multi-GPU Systems"
Authors/Creators
Description
Artifacts for "Triple Overlap in Pipelined Krylov Methods on Multi-GPU Systems"
===============================================
This directory contains computational artifacts related to our SC'26
workshop paper.
The paper describes computational experiments that were conducted to
evaluate the performance of pipelined BiCGStab and pipelined
preconditioned CG solvers in aCG, using NCCL for communication, and
compares them with the corresponding solvers in PETSc on two GPU
clusters: MareNostrum 5 (4 NVIDIA H100 GPUs per node) and Leonardo (4
NVIDIA A100 GPUs per node).
History:
--------
v1 (2026-09-25): Initial version
Artifacts:
----------
The computational artifacts are located in several subdirectories:
* 'src/aCG/' contains the source code for aCG
(https://github.com/ParCoreLab/aCG, commit 5ff8adb), which
implements the multi-GPU solvers that are used for the performance
benchmarks presented in the paper.
* 'data/' contains the partitions of the input matrices. From the
SuiteSparse Collection (Davis and Hu, 2011), six matrices were
selected: audikw_1, Bump_2911, Flan_1565, Geo_1438, Queen_4147 and
Serena. Row partitions computed with METIS (Karypis and Kumar,
1998) are provided for 4, 8, 16, 32 and 64 parts
('suitesparse/partitions/'). The matrices themselves are not
included; scripts/download-matrices.sh downloads them in Matrix
Market format from the SuiteSparse Matrix Collection into
'suitesparse/mtx/' and decompresses the partitions:
scripts/download-matrices.sh
The checksums of all files are
listed in data/SHA256SUMS; the download script verifies the
SuiteSparse files against them.
* 'scripts/' contains scripts for building aCG and job scripts for
submitting jobs on two clusters: MareNostrum 5 ('marenostrum5/')
and Leonardo ('leonardo/'). These scripts carry out performance
measurements for the pipelined solvers in aCG and PETSc. The README
in the subdirectory for each cluster describes how to configure the
environment, build aCG, run the jobs and collect the results.
'collect/' contains a script that turns the job output into a CSV
file with solver times, and 'figures/' contains the scripts that
produce the figures of the paper from that file.
References
----------
Davis, T. A. and Y. Hu. 2011. “The University of Florida Sparse Matrix Collection”.
ACM Transactions on Mathematical Software 38, 1, Article 1 (December 2011), 25 pages.
DOI: https://doi.org/10.1145/2049662.2049663
Karypis, G., and V. Kumar. 1998. “A fast and high quality multilevel scheme for partitioning irregular graphs”.
SIAM Journal on scientific Computing 20, 1, pp. 359–392.
DOI: https://doi.org/10.1137/S1064827595287997