Tree-based Overlay Networks

Tree-based Overlay Networks (TBŌNs) are hierarchically organized process networks used for scalable, high-performance data communication and aggregation. Our open-source TBŌN prototype, MRNet, has an easy-to-use C++ API and supports features like flexible network topologies, customizable data aggregation filters, multiple concurrent data streams and scalable, responsive failure recovery mechanisms. MRNet runs on UNIX/Linux clusters and supercomputers like the IBM BlueGene and Cray XT. As a result of its scalability characteristics, flexibility and wide availability, MRNet has been integrated into a variety of tools and applications and is being used at research organizations throughout the world, including the Lawrence Livermore, Los Alamos and Sandia National Laboratories.

Autonomous Middleware

In developing highly-scalable, generic middleware services for tools and applications, this work seeks to answer two fundamental questions: (1) what are the performance limits of application-agnostic middleware infrastructure? and (2) how efficiently can a developer use middleware infrastructure without any performance considerations or knowledge of the system’s underlying mechanisms. In other words, how well can an infrastructure perform on behalf of an application if there is no application-specific knowledge embedded in the infrastructure, and the application developer is not a computer systems expert.

Our approach to this problem is to use lightweight mechanisms for autonomous middleware infrastructure. Such infrastructure entails monitoring components that compile events that are analyzed to detect functional and performance deficiencies. Decision processes must then be used to determine corrective actions and, in the case of performance failures, evaluate the costs and benefits of system reconfigurations. An important observation is that the definition of a performance failure is dataflow dependent; i.e. the rate of data transmission, the complexity of the aggregation operation, and the topology of the middleware infrastructure determine whether a deficiency exists -- generally defined as an under or over subscription of resources. Accordingly, corrective actions need to be considered on a per dataflow basis

This project is part of a UNM/Wisconsin collaboration.

MRNet website: MRNet


Loading publications...