hhhhhhejia/sciaccelbench-sa-0008

SciAccelBench sa-0008 — public practice graphs

The public practice inputs for sa-0008, "METIS — multilevel graph partitioning on GPU", in the SciAccelBench registry. The task page is at aitonomy.org/projects/sci-accel.

These are the graphs a solver develops against. Official scoring uses held-out graphs that are not in this repository and that differ from these in family, size, degree distribution and part count.

Contents

file	vertices	edges	what it is
graphs/mdual.graph	258,569	513,132	3D finite-element dual mesh; ships with METIS itself. Small enough to iterate on in a tenth of a second.
graphs/delaunay_n22.graph	4,194,304	12,582,869	Delaunay triangulation of random points in the unit square. Near-planar, bounded degree.
graphs/road_usa.graph	23,947,347	28,854,312	The US road network. Very high diameter and low average degree — the family multilevel partitioners find slowest.

Each graph appears twice:

• graphs/*.graph — METIS graph format, the text format gpmetis reads. This is what the task's agent container mounts at /app/data.
• csr/*.csr — the same graph as flat little-endian CSR (int64 n, int64 m, int32 indptr[n+1], int32 indices[2m]). The task's verifier memory-maps these to recompute the edge cut of a submitted partition, rather than spend minutes of its budget re-parsing text. Produced by authoring/graph2csr.c in the task package.

checksums.sha256 covers every data file. The task's Dockerfiles verify it at image build and fail the build on a mismatch — a corrupted graph is a different benchmark, and it should stop the build rather than quietly change the answer.

Provenance

mdual is redistributed from the METIS repository (KarypisLab/METIS, graphs/mdual.graph, tag v5.2.1), which is Apache-2.0.

delaunay_n22 and road_usa come from the SuiteSparse Matrix Collection, sets DIMACS10/delaunay_n22 (id 2477) and DIMACS10/road_usa (id 2459), both contributed for the 10th DIMACS Implementation Challenge. They are redistributed here under the terms of that collection; the original Matrix Market files carry their own provenance headers, which are preserved in the collection. Converted to METIS graph format by authoring/mtx2metis.c in the task package — symmetrised, self-loops dropped, duplicate edges removed, because METIS does not check for any of those and simply computes a wrong answer instead.

The conversion is reproducible from the task package: both converters are committed, and re-running them on the SuiteSparse originals reproduces these checksums.

Citation

If you use these graphs, cite the collection and the challenge:

• T. A. Davis and Y. Hu, "The University of Florida Sparse Matrix Collection", ACM Trans. Math. Softw. 38(1), 2011.
• D. A. Bader, H. Meyerhenke, P. Sanders, D. Wagner (eds.), Graph Partitioning and Graph Clustering: 10th DIMACS Implementation Challenge, AMS, 2013.

and, for METIS itself:

• G. Karypis and V. Kumar, "A Fast and High Quality Multilevel Scheme for Partitioning Irregular Graphs", SIAM J. Sci. Comput. 20(1):359–392, 1998.