Accelerating GPU Hardware Transactional Memory with Snapshot Isolation

Abstract

Snapshot Isolation (SI) is an established model in the database community, which permits write-read conflicts to pass and aborts transactions only on write-write conflicts. With the Write Skew anomaly correctly eliminated, SI can reduce the occurrence of aborts, save the work done by transactions, and greatly benefit long transactions involving complex data structures. GPUs are evolving towards a general-purpose computing device with growing support for irregular workloads, including transactional memory. The usage of snapshot isolation on transactional memory has proven to be greatly beneficial for performance. In this paper, we propose a multi-versioned memory subsystem for hardware-based transactional memory on the GPU, with a method for eliminating the Write Skew anomaly on the fly, and finally incorporate Snapshot Isolation with this system. The results show that snapshot isolation can effectively boost the performance of dynamically sized data structures such as linked lists, binary trees and red-black trees, sometimes by as much as 4.5x, which results in improved overall performance of benchmarks utilizing these data structures.