Fast Task Submission in Software Thread Level Speculation Systems

Computation resources provided by modern multi-core processors make it possible to exploit thread-level parallelism in a specu-lative mode. In the past twenty years, tens of thread-level speculation (TLS) systems have been proposed and demonstrated the performance potential of TLS. Limited by the performance over-head caused by dynamic data dependency detection, speculative threads are usually confirmed sequentially in most TLS systems, and the execution of a speculative task may be delayed due to the rollback of tasks before it, even if there is no data dependency among them. In this paper, submission buffer, an aggressive TLS mechanism, is proposed to eliminate the unnecessary waiting caused by rollback. To keep the mis-speculation rate of this aggressive mechanism be at a lower value, some dynamic schedule strategies are also discussed. Experimental results indicate that the execution time could be shortened effectively under the same mis-speculation rate, and the overall performance is improved accordingly.