IPDPS '05 Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS'05) - Papers - Volume 01
| Hi-index | 0.00 |
Many large and complex computational applications canbe modeled as irregular graphs and are typically characterizedby a large number of vertices and edges. This paper proposesa new and fast mapping heuristic, called FastMap, tomap this class of applications onto heterogeneous metacomputingplatforms such as computational grids. While previousapproaches have delved into graph partitioning of the application,we attempt to solve this problem from the clusteringperspective. We exploit a hierarchical resource managementinfrastructure on the grid to distribute the overhead of mappingamong a tree of schedulers and develop a scheme thatproves to be almost linear in its scalability. Furthermore, weoptimize on the result of the mapping with the help of a geneticalgorithm at each scheduler node. Our experiments include a50,000-node application graph from NASA and several othersynthetically-generated graphs with as many as 100,000 vertices.Comparisons with another heterogeneous partitioner,MiniMax, show an improvement factor of over 100 on themapping time, yet with superior quality mapping.