Date: Thursday, April 21st, 2022

Time: 2pm - 4pm EDT

Location (virtual): https://gatech.zoom.us/j/94304662522

Kasimir Georg Gabert

PhD Candidate

School of Computer Science / School of Computational Science and Engineering

College of Computing

Georgia Institute of Technology

https://kasimir.co

Committee:

Dr. Ümit V. Çatalyürek (advisor), CSE, Georgia Institute of Technology

Dr. Srinivas Aluru, CSE, Georgia Institute of Technology

Dr. B. Aditya Prakash, CSE, Georgia Institute of Technology

Dr. Srijan Kumar, CSE, Georgia Institute of Technology

Dr. Ali Pınar, Data Science and Cyber Analytics, Sandia National Laboratories

Abstract:

Many of today's massive and rapidly changing graphs contain internal structure---hierarchies of locally dense regions---and finding and tracking this structure is key to detecting emerging behavior, exposing internal activity, summarizing for downstream tasks, identifying important regions, and more.  Existing techniques to track these regions fundamentally cannot handle the scale, rate of change, and temporal nature of today's graphs.  We identify the crucial missing piece as the need to address the significant variability in graph change rates, algorithm runtimes, temporal behavior, and dense structures themselves.

We tackle tracking dense regions in three parts.  First, we extend algorithms and theory around dense region computation.  We computationally unify nuclei into computing hypergraph cores, providing significant improvements over hand-tuned nuclei algorithms and enabling higher-order nuclei.  We develop new batch algorithms for maintaining core hierarchies.  We then define new temporal dense regions, called core chains, that build on nuclei hierarchy maintenance and enable effective and powerful dense region tracking.

Second, we scale up on shared-memory systems.  We provide a parallel input and output library that reduces start-up costs of all known graph systems.  We provide the first parallel scalable core and hypergraph core maintenance algorithms, building on the connection between h-indices and cores.  This addresses computation on rapidly changing graphs during bursty periods with large numbers of graph changes.

Third, we address scaling out to support massive graphs.  We develop the first parallel h-index algorithm, the key kernel for tracking dense regions.  We identify that system elasticity is imperative to handle large bursts of changes.  We develop a dynamic and elastic graph system, using consistent hashing and sketches, and demonstrate competitive performance against static, inelastic graph systems while enabling new, dynamic applications.

By addressing variability directly---in algorithm and system design---we break through previous barriers and bring dense region tracking to massive, rapidly changing graphs.