Influence of atomic FAA on ParallelFor and a cost model for improvements

TL;DR

This paper investigates how atomic FAA operations impact ParallelFor's latency and proposes a cost model to optimize performance by reducing FAA influence across different platforms.

Contribution

It introduces a detailed analysis of FAA's effect on ParallelFor and presents a cost model for performance improvements.

Findings

FAA significantly affects ParallelFor latency

Performance varies across different hardware platforms

A cost model can guide optimization strategies

Abstract

This paper focuses on one of the most frequently visited multithreading library interfaces - ParallelFor. In this study, it is inferred that ParallelFor's end-to-end latency performance is noticeably affected by the frequency with which fetch-add-add (FAA) is called during program execution. This can be explained by ParallelFor's uniform semantics and the utilization of atomic FAA. To prove this assumption, a battery of tests was designed and conducted on diverse platforms. From the collected performance statistics and overall trends, several conclusions were drawn and a cost model is proposed to enhance performance by mitigating the influence of FAA.