A framework for reliable and efficient data placement in distributed computing systems
Journal of Parallel and Distributed Computing - Special issue: Design and performance of networks for super-, cluster-, and grid-computing: Part I
Performance Evaluation of Data Management Layer by Data Sharing Patterns for Grid RPC Applications
Euro-Par '08 Proceedings of the 14th international Euro-Par conference on Parallel Processing
| Hi-index | 0.00 |
High-performance peer-to-peer transfer between clusterswill be fundamental technology base for various Gridmiddleware, such as large-scale data transfer in DataGridsettings, or collective communication in Grid-wide MPIs.There, two major factors are involved: on one hand networkpipes with large RTT 脳 bandwidth typically becomedata-starved, resulting in bandwidth loss; on the other handwhen multiple nodes on the clusters attempt simultaneoustransfer, the network pipe could become saturated, resultingin packet loss which again may result in bandwidthdegradation in large RTT 脳 bandwidth networks. By dynamicallyand automatically adjusting transfer parametersbetween the two clusters, such as the number of networknodes, number of socket stripes, we could achieve optimalbandwidth even when the network is under heavy contention.In order to arrive at a proper performance modelfor automated adjustment, we have conducted several simulationsby which we have discovered that such automatictuning would beneficial, but the ideal number of networkpipes does not exactly match the simple transfer model oftraditional peer-to-peer settings between single nodes.