Software transactional memory for multicore embedded systems

Abstract

Embedded systems, like general-purpose systems, can benefit from parallel execution on a symmetric multicore platform. Unfortunately, concurrency issues present in general-purpose programming also apply to embedded systems, protection from which is currently only offered with performance-limiting coarse-grained locking or error-prone and difficult-to-implement fine-grained locking . Transactional memory offers relief from these mechanisms, but has primarily been investigated on general-purpose systems. In this paper, we present Embedded Software Transactional Memory (ESTM) as a novel solution to the concurrency problem in parallel embedded applications. We investigate common software transactional memory design decisions and discuss the best decisions for an embedded platform. We offer a full implementation of an embedded STM and test it against both coarse-grained and fine-grained locking mechanisms. We find that we can meet or beat the performance of fine-grained locking over a range of application characteristics, including size of shared data, time spent in the critical section, and contention between threads. Our ESTM implementation benefits from the effective use of L1 memory, a feature which is built into our STM model but which cannot be directly utilized by traditional locking mechanisms.