Clustered Communication for Efficient Pipelined Multithreading on Commodity MCPs

Low inter-core communication overheads are critical for pipelined multithreading (PMT) to using multi-core processors (MCPs) to improve the performance of general sequential applications. However, conventional software queue based communication mechanism will bring significant communication overheads, which limit the potential performance and hinder the wide commercial use. While dedicated intercore communication mechanism has been proposed, it demands chip redesign effort, costs so much and needs extensions to ISA. This paper addresses this problem and proposes a novel clustered communication mechanism to minimize the communication overheads from the average standpoint. We observe that the PMT performance is very sensitive to inter-core communication overheads, but is insensitive to amount of parallelisms. Based on the observation, we can achieve very low average communication overheads (ACOs) through sacrificing a certain amount of parallelisms. The principle of clustered communication mechanism and how to reduce the ACOs with this mechanism are presented in detail. A concurrent lockfree clustered software queue algorithm, which applies this mechanism, is given to support the pipelined communication. The algorithm is evaluated on the AMD Phenom four-core processor and experimental results show its communication performance is over 10x faster than that of conventional software queue, and significant PMT performance of real applications are, therefore, achieved. Keywords-Pipelined multithreading; commodity multi-core processors; software queue; clustered communication; low inter-core communication overheads.