Miao Ju, Hun Jung, and Hao Che
Modeling and Simulation, Parallel Computer Architectures
This paper shows how a class of queuing network models with closed-form solutions can be used to characterize a large class of single-core multithreaded processors with various thread scheduling disciplines and resource access mechanisms. In particular, an explicit expression for the CPU idle probability as a function of the number of threads and other workload/system parameters is derived for this class of queuing network models. This expression leads to the identification of a surface in a multidimensional design space spanned by workload and system parameters, which splits the space into two distinct regions. In one region, multithreading can completely hide resource access latencies from the CPU, resulting in maximal throughput performance, whereas in the other, multithreading cannot. The surface can be easily identified with several workload and system parameter estimations, without having to resort to simulation or benchmark testing.
Important Links:
Go Back
