PEWs: a decentralized dynamic scheduler for ILP processing

This paper presents a new execution model called PEWs (Parallel Execution Windows). It performs decentralized dynamic scheduling by splitting the instruction window among multiple hardware execution windows that can all execute code in parallel. While splitting the instruction stream, dependent instructions are generally allocated to the same execution window. Speculative execution is carried out with the help of branch prediction; speedy recovery from incorrect speculative execution is effected with the help of decentralized recovery actions. Other novel features of the execution model include register file queues to enforce register dependencies. The paper also presents the simulation results, to date, of our attempts to study the PEWs execution model. The results indicate goal potential for this execution model.