Developing a cost/benefit estimating service for dynamic resourcesharing in heterogeneous clusters: experience with SNL clusters

摘要

Heterogeneous or federated clusters are a significant special caseof grid computing. An important factor in efficiently executingapplications on such clusters is the mapping (or scheduling) ofapplication tasks onto cluster nodes. The cost/benefit estimatingservice aims at facilitating the search for efficient mappings bothinitially and during the evolution of a computation. This servicedynamically evaluates alternative mappings using system and applicationinformation. Building the necessary database for use with this serviceinvolves profiling applications and running benchmark suites on clusternodes. Recent data obtained at Sandia National Laboratories, NM,demonstrate the existence of differences in computation andcommunication speeds between the nodes of even simple homogeneousclusters, as well as the feasibility of using such data for predictingthe efficiency of application mappings