Design and evaluation of dynamic key message algorithms for cluster computing

The Key Message (KM) approach prioritizes communications along the critical path to speed up the execution of parallel applications in a cluster environment. In this article, we introduce a family of three dynamic KM algorithms that take into account of the dynamic nature of underlying networks to improve over their predecessor. These three dynamic KM algorithms are: Fixed Critical Path Dynamic KM (FDKM), Centralized Predictor Dynamic KM (CDKM), and Distributed Dynamic KM (DDKM). Our performance study based on simulation showed that in general CDKM performs better and is more stable than FDKM and DDKM under heavy traffic conditions. The performance of DDKM is close to that of CDKM and better than that of FDKM on average. By taking advantages of both FDKM and CDKM, DDKM can be expected to be a promising optimization in a real cluster environment