Data Driven Communication and Synchronization in Non-Uniform Bandwidth Computing Clusters

Purdue Research Foundation, West Lafayette IN

Investigators

Abstract

This project proposes to develop efficient network- and application-aware data-driven communication and synchronization primitives for geographically distributed computing clusters connected over dynamically configured non-uniform bandwidth links. Synchronization and communication primitives are the backbone of any non trivial parallel computation and performance in cluster systems crucially depends on the efficiency of such primitives. The goal is to effectively incorporate diverse and dynamically varying network characteristics into the proposed solutions while taking advantage of the application characteristics. The solutions will be based on methods that detect and limit avoidable congestion and hotspots. The techniques employed will include collecting global information, sampling, and randomization. Examples of synchronization primitives to be investigated include barrier, eureka, and termination synchronizations for different application-motivated models. Data-driven communication primitives include data-dependent forms of many-to-many broadcast, multicast, and many-to-many transport operations. The experimental aspects of the proposed research will be carried out on a collection of high performance computing clusters and platforms spread across Purdue University, University of Illinois at Chicago, and NSF/DOD sponsored HPC centers, interconnected via different bandwidth links in the Internet. Data-intensive applications such as data mining/warehousing and multimedia will be used as a vehicle to study data-driven synchronization and communication issues. The proposed work will facilitate the design of such parallel applications on cluster systems.