Graph500 from OCaml-Multicore Perspective

TL;DR

This paper presents an implementation and performance analysis of Graph500 benchmarks using OCaml-Multicore, focusing on graph algorithms like construction, BFS, and shortest path, with insights into data structures, algorithms, and performance challenges.

Contribution

It provides a detailed OCaml-based implementation of Graph500 benchmarks, including algorithm choices, data structures, and performance analysis, highlighting challenges and future directions.

Findings

Performance metrics tested on Intel Xeon Gold 5120

Identified initial performance failures and bottlenecks

Guidance on data structures and algorithm choices for OCaml implementations

Abstract

OCaml is an industrial-strength, multi-paradigm programming language, widely used in industry and academia. OCaml was developed for solving numerical and scientific problems involving large scale data-intensive operations and one such classic application set is Graph Algorithms, which are a core part of most analytics workloads. In this paper, we aim to implement the graph benchmarks along with the performance analysis. Graph500 is one such serious benchmark which aims at developing data intensive applications requiring extreme computational power. We try to implement Graph Construction, BFS, Shortest-Path problems using the desired specifications and rules posed by graph500. This paper aims at providing a clear direction of choices of several data structures used, algorithms developed and pose a reason behind every step of program. The first few sections of the paper discusses a formal approach to the problem with a small guide for starters in OCaml. The latter sections describe the algorithms in detail with the possibilities of future exploration and several mistakes which we committed or encountered whilst approaching the solution. All performance metrics were tested on Intel(R) Xeon(R) Gold 5120 CPU @ 2.20GHz 24 core machine. Every section talks about the initial performance failures encountered, which will help analyse and prioritise our preferred implementation from a performance perspective.