Improving Resource Utilization by Curbing Speculative Trace Progression in Simultaneous Multi-Threading CPUs

張怡麟(YILIN ZHANG) ； 林維明(WEI-MING LIN)

《Journal of Information Science and Engineering》 32卷6期 (2016/11) Pp. 1487-1502

摘要

Simultaneous Multi-Threading (SMT) improves the overall performance of superscalar CPUs by allowing concurrent execution of multiple independent threads with sharing of key datapath components in order to better utilize the resources. Speculative executions help modern processors to exploit more Instruction-Level Parallelism. However, the performance penalty from a miss speculation is much more prominent in an SMT environment than a traditional multi-threading system due to the resulted waste of shared resources at clock-cycle level, versus thread level. In this paper, we show that instructions fetched due to incorrect prediction can be more than 30% of all instructions, which results in a huge waste of resources that could have been better used by other non-speculative threads. To minimize this waste of resources, a technique is proposed in this paper to control the amount of speculative instructions dispatched into Issue Queue (IQ), the most critically shared resource in the SMT pipeline. Simulation result shows the proposed technique can reduce the waste of resource due to miss-speculated traces by 38% and improve overall throughput by up to 17% in IPC.

關鍵字

computer architecture ； superscalar ； speculative execution ； simultaneous multithreading ； thread-level parallelism