Exploring Fine-grained Task Parallelism on Simultaneous Multithreading Cores

TL;DR

This paper analyzes the performance of fine-grained task parallelism on simultaneous multithreading cores and introduces Relic, a simple software framework that significantly improves speedups over existing parallel programming frameworks.

Contribution

The paper presents Relic, a novel software-only framework enabling extremely fine-grained tasking on SMT cores, with demonstrated performance improvements.

Findings

Relic achieves up to 33.2% higher speedup than X-OpenMP.

Performance improvements of 19.1% to 33.2% over various OpenMP implementations.

Performance gains are demonstrated on real-world fine-grained application kernels.

Abstract

Nowadays, latency-critical, high-performance applications are parallelized even on power-constrained client systems to improve performance. However, an important scenario of fine-grained tasking on simultaneous multithreading CPU cores in such systems has not been well researched in previous works. Hence, in this paper, we conduct performance analysis of state-of-the-art shared-memory parallel programming frameworks on simultaneous multithreading cores using real-world fine-grained application kernels. We introduce a specialized and simple software-only parallel programming framework called Relic to enable extremely fine-grained tasking on simultaneous multithreading cores. Using Relic framework, we increase performance speedups over serial implementations of benchmark kernels by 19.1% compared to LLVM OpenMP, by 31.0% compared to GNU OpenMP, by 20.2% compared to Intel OpenMP, by 33.2% compared to X-OpenMP, by 30.1% compared to oneTBB, by 23.0% compared to Taskflow, and by 21.4% compared to OpenCilk.