Scalable Software Transactional Memory for Global Address Space Architectures

This paper presents the challenges encountered in and potential solutions to designing scalable Software Transactional Memory (STM) for large-scale distributed memory systems with thousands of nodes. We introduce Global Transactional Memory (GTM), a generalized and scalable STM design supporting a dynamic programming model based on threadlevel parallelism, Single Process Multiple Data (SPMD) parallelism, and remote procedure call invocation. In addition to this, GTM is the first STM design to support the nonblocking abstraction, i.e. the control returns immediately from the STM procedure call instead of waiting on its entire operation to finish. This allows remote STM operations to execute asynchronously with respect to the caller, thus tolerating remote communication latency and exposing new opportunities for parallelism within the context of software transactions. Our experimental results for Red-Black Tree, Priority Queue, and Bank Transaction workloads exhibit good strong and weak scaling attributes on a wide range of node counts (up to 1024 quad-code nodes) and problem sizes (up to one billion transaction commits). The results also demonstrate the benefits of leveraging asynchronous protocols for tolerating latency, providing additional speedups of 1.8X, 1.75X, and 1.27X or greater over their synchronous counterparts for these workloads.