A Consistency Checking Optimization Algorithm for Memory-Intensive Transactions

Transactional memory (TM), a recent parallel programming concept, aims to simplify parallel programming while simultaneously maintaining performance benefits found in concurrent applications. Consistency checking, the manner in which memory conflicts are identified in transactional memory, is a critical aspect to TM system performance. We present a theoretical, analytical and empirical view of our novel consistency checking algorithm which is optimized for spatially wide transactional workloads. Initial tests show our algorithm yields super linear performance improvements over other alternatives as transaction size grows, resulting in performance gains between 5x − 250x for experimental benchmarks.