Research On The Sufficient Thread Submission Mechanism In Software Thread Level Speculation Systems

Computation resources provided by modern multi-core processors make it possible to exploit thread-level parallelism in a speculative 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 overhead 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 though there is no data dependency among them.The sequential confirmation mechanism decrease the execution efficiency of the threads,which in turn limits the practical enhancement of the system's performance.1)Based on the analyzing to the sequential confirmation mechanism of HEUSPEC speculation system,this paper undertakes a deep study on the limitations brought in by the sequential mechanism.The main work and innovations include: Most TLS systems obey the strict sequential rules during the confirmation of speculative results.These rules,which ensure the correctness of a program's execution,also result in unnecessary waiting caused by rollbacks.The waiting overhead will finally cut down the performance of the TLS systems.This paper proposes the Aggressive Task Submission Buffer Mechanism to get rid of the unnecessary waiting,which improves the execution efficiency of the threads and makes it possible to optimize the performance of HEUSPEC model.2)The SUB mechanism will assign more speculative tasks in advance than normal TLS systems due to the existence of a Data Buffer.As to the program whose data dependency is certain before its execution,these advance assignments increase the chances of rollbacks.A bigger Data Buffer can lead to a more serious rollback chain effects.To keep the rollbacks rate of this aggressive mechanism be at a lower value,two dynamic strategies,the Inter-Task Fetch and Dynamic Buffer Resizing,are discussed.Experimental results indicate that the execution time could be shortened effectively under the same rollback rate,and the overall performance is improved accordingly when SUB mechanism is used,and the peak performance enhancement is 28.32%.The ITF strategy is useful in controlling the rollback rate,and the average decrease is 57.47%,which gives an average rise of 19.77% in calculation performance.When DBR optimization strategy is brought in,along with the ITF strategy,a further performance enhancement is gained,6.73% on average.