Hardware support for software controlled multithreading

Chip multi-processors have emerged as one of the most effective uses of the huge number of transistors available today and in the future, but questions remain as to the best way to leverage CMPs to accelerate single threaded applications. Previous approaches rely on significant speculation to accomplish this goal. Our proposal, NXA, is less speculative than previous proposals, relying heavily on software to guarantee thread correctness, though still allowing parallelism in the presence of ambiguous dependences. It divides a single thread of execution into multiple using the master-worker paradigm where some set of master threads execute code that spawns tasks for other, worker theads. The master threads generally consist of performance critical instructions that can prefetch data, compute critical control descisions, or compute performance critical dataflow slices. This prevents non-critical instructions from competing with critical instructions for processor resources, allowing the critical thread (and thus the workload) to complete faster. Empirical results from performance simulation show a 20% improvement in performance on a 2-way CMP machine, demonstrating that software controlled multithreading can indeed provide a benefit in the presence of hardware support.