Adaptive Data Path Selection for Durable Transaction in GPU Persistent Memory

TL;DR

This paper proposes AGPM, an adaptive data-path selection strategy for durable transactions in GPU persistent memory, significantly improving execution time over existing designs.

Contribution

It introduces a novel adaptive data-path selection approach for GPU persistent memory, enhancing performance of durable transactions.

Findings

Reduces GPU kernel execution time by at least 24.37%.

Achieves up to 66.44% performance improvement.

Demonstrates effectiveness of adaptive data-path selection in GPU persistent memory.

Abstract

The new non-volatile memory technology relies on data recoverability to achieve the promise of byte-addressable persistence in computer applications. The durable transaction (e.g. logging) is one of the major persistency programming models to provide recoverable data structures. To achieve performant failure-atomic transactional updates to PM, multi-data-path architectures that separate the data paths for persists are recently explored for CPUs. Considering the importance of GPU as a key computing platform for many application domains, we investigate the multi-data-path architecture for durable transactions to PM in GPU. Our solution, AGPM, exploits an adaptative data-path-selection strategy for the log updates to PM. AGPM reduces the GPU kernels' execution time by at least 24.37% (at most 66.44%) compared to the state-of-the-art designs.