A Benchmark to Evaluate InfiniBand Solutions for Java Applications

TL;DR

This paper introduces a benchmark to evaluate InfiniBand solutions for Java applications, focusing on throughput and latency, to aid developers in selecting suitable network solutions for high-performance computing environments.

Contribution

It provides the first comprehensive benchmark for InfiniBand in Java, comparing socket- and verbs-based libraries across different hardware configurations.

Findings

Socket- and verbs-based solutions show different performance trade-offs.

Benchmark results vary significantly between 56 Gbit/s and 100 Gbit/s NICs.

The benchmark aids developers in choosing optimal InfiniBand solutions.

Abstract

Low-latency network interconnects, such as InfiniBand, are commonly used in HPC centers and are even accessible with todays cloud providers offering equipped instances for rent. Most big data applications and frameworks are written in Java. But, the JVM environment alone does not provide interfaces to directly utilize InfiniBand networks. In this paper, we present the ``Java InfiniBand Benchmark'' to evaluate the currently available (and supported) ``low-level'' solutions to utilize InfiniBand in Java. It evaluates socket- and verbs-based libraries using typical network microbenchmarks regarding throughput and latency. Furthermore, we present evaluation results of the solutions on two hardware configurations with 56 Gbit/s and 100 Gbit/s InfiniBand NICs. With transparency often traded for performance and vice versa, the benchmark helps developers with studying the pros and cons of each solution and support them in their decision which solution is more suitable for their existing or new use-case.