Skip to Main Content
The performance in simultaneous multithreading (SMT) processors is significantly influenced by how optimally the common resources are shared among all threads. However, the programs running on threads often take on changing complicated behavior at runtime. Existing resource distribution methods either rely mainly on front-end fetching policies, or make the resource distribution decision according to the restricted information from the pipeline instead of the direct in-flight performance. It's difficult for them to catch various resource requirements efficiently and obtain potential gain. This paper presents a dissipative resource distribution policy (DRD) for SMT processors, in which the resource allocations evolve targeting the direct runtime performance. The self-organization mechanism established for DRD and the introduction of additional chaos cooperate together to form the fluctuation in the distribution evolvement procedure, and maintain persistent evolvement in changing program behavior. Simulation results show that DRD benefits much from its dissipative structure, and obtains potential fairness gains as well as pleasing throughput IPC increments. The meaning and setting of important DRD parameters are also discussed.
Date of Conference: 14-18 April 2008