sRSP: GPUlarda Asimetrik Senkronizasyon Icin Yeni Olceklenebilir Bir Cozum

TL;DR

This paper introduces sRSP, a scalable GPU synchronization method that efficiently manages asymmetric data sharing by reducing heavyweight synchronization, resulting in significant performance improvements.

Contribution

The paper presents sRSP, a novel scalable implementation of Remote Scope Promotion for GPUs that optimizes synchronization for asymmetric sharing scenarios.

Findings

sRSP improves GPU performance by 29% on average.

sRSP reduces synchronization overhead in asymmetric sharing.

The evaluation used Gem5-APU simulator.

Abstract

Asymmetric sharing is a dynamic sharing model, where a shared data is heavily accessed by a (local) sharer, and rarely accessed by other (remote) sharers. On GPUs, without special support, asymmetric sharing requires heavily loaded synchronization on every access. With the introduction of Remote Scope Promotion (RSP), access to the local sharer is allowed with lightweight synchronization, while heavyweight synchronization is only used for remote accesses where it is rarely needed. RSP ensures data consistency by promoting local synchronizations on remote accesses. Unfortunately, the first implementation of RSP is not a scalable solution. We offer a more efficient and scalable RSP implementation. This new design, which we call sRSP, is based on the monitoring of the local sharer and the selective execution of heavyweight synchronization operations. We evaluated the sRSP with the time-detailed Gem5-APU simulator and the results show that the sRSP improves performance by an average of 29 percent on a 64 Compute Unit GPU.