A GPU/CPU Hybrid Cluster for Accelerating Network-of-X Research and Education
Abstract
In the age of AI and big data, networks provide powerful context to connect different things (e.g., entities, datasets, models) together. This phenomenon is referred to as Network-of-X in this project. By linking different entities, datasets and data mining models together, Network-of-X brings a unique opportunity to dramatically enrich the modeling power, boost the mining performance, and broaden the reasoning capabilities. On the flip side, Network-of-X faces two significant challenges in computation, including (1) the networked computing due to both the intertwined data in huge volume and algorithms with high complexities, and (2) the interactive computing which requires real-time or near real-time interactions between humans and computers to analyze the data and models across different granularities. The project plans to build an interactive networked infrastructure to accelerate Network-of-X research and education. The proposed infrastructure has three mutually complementary modules. The first module (Distributed Graph Database) consists of a storage-optimized cluster offering high disk throughput and a distributed graph database management system supporting fast transaction processing. The second module (Graph Computation Engine) consists of a hybrid CPU/GPU server that is computationally optimized. It supports two types of fundamental computations in Network-of-X, including the large sparse matrix computations and the deep learning computations. The third module (Interactive Graph Analytics) develops an integrated human-computer interactive platform that allows the end-users to effectively collaborate and coordinate with computers in decision making, cognitive modeling, and algorithm training and diagnosis. The proposed infrastructure will benefit a variety of Network-of-X applications, including network science of teams, modeling adversarial activities, robustness of inter-dependent infrastructure networks, isolated malicious activity analysis, fair network learning, etc. It will help distill the critical value from the massive, heterogeneous and dynamic data and models. It will further unveil subtle patterns that would be invisible if we look at the different datasets or data mining models separably. The success of this project will significantly improve the ability to perform basic research related to DoD interest, and train both graduate and undergraduate students in science and engineering. It will enhance DoDÕs capability for networked data modeling, mining and reasoning in terms of effectiveness, scalability and interactivity.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110088
Entities
People
- Hanghang Tong
Organizations
- Army Contracting Command
- United States Army
- University of Illinois Urbana–Champaign