A Performance Comparison of Homeless and Home-based Lazy Release Consistency Protocols in Software Shared Memory
5th International Symposium on High-Performance Computer Architecture, Orlando, FL, January 1999
In this paper, we compare the performance of two multiple-writer protocols based on lazy release consistency. In particular, we analyze the performance of Princeton's home-based protocol and TreadMarks' protocol on a 32-processor platform. We found the performance of the two protocols to be strikingly similar for most of our application suite. The homeless protocol performed better for two of the applications while the home-based protocol outperformed TreadMarks for one application. We show, however, that for the latter, the performance gap is not a consequence of the disparity in message count. It is, rather, a product of differing abilities of the protocols to balance communication among nodes. These results differ from a previous study conducted on the two protocols. We attribute the difference to: (1) a platform with a different ratio of memory bandwidth to network bandwidth and (2) lazy diffing and request overlapping, two features of the original TreadMarks implementation that were not considered in the previous study.