Hardware/Software Techniques for Assisted Execution Runtime Systems

The increasing complexity of modern and future multi-core/multithreaded processors rises the question of how to best utilize processor resources. On one side, Amdahl’s Law limits the maximum theoretical speedup of parallel applications while, on the other side, the increasing complexity of runtime programming language may introduce implicit serialization points. Several studies demonstrated that it is often more convenient to use some of the hardware threads to assist execution than running supplementary application threads. Assisted execution approaches, however, may lead to low processor utilization: in this paper, we explore fine-grained hardware resource allocation techniques to assign hardware resources to application and auxiliary threads at runtime, according to their actual computing power demand. As a test case, we apply fine-grained resource allocation to STM2, the first parallel STM system that offload STM management operations to auxiliary threads. We implemented an integrated hardware/software solution in which each level performs well-defined tasks efficiently: 1) STM2 is enriched with a runtime mechanism to express computing power requirements of application and auxiliary threads; 2) the hardware enforces dynamic resource partitioning among running threads; 3) the operating system provides a simple and efficient interface between STM2 and the hardware resource allocation mechanism. In this paper, we leverage the IBM POWER7 hardware thread prioritization mechanism to bias the allocation of hardware resources in favor of computing intensive application threads or overloaded auxiliary threads. We test fine-grained resource allocation solutions on a real IBM POWER7 system running a simple and malleable TM benchmark (Eigenbench) and applications from the STAMP benchmark suite. Results show that the proposed integrated solution achieves up to 65% and 11% of performance improvement over the standard STM2 design for Eigenbench and STAMP applications, respectively.