Modeling computations as iterative graph algorithms has gained wide acceptance across recent analytics techniques due to its effectiveness in interacting with complex data relationships. While many parallel graph algorithms have been developed to perform useful analyses, the large sizes of real-world graphs necessitate developing efficient custom solutions that can fully exploit the available compute and memory capacity to process these large graphs.
In this talk, I will first introduce my work that exploits the inherent asynchrony available in various graph algorithms to accelerate large-scale graph processing while simultaneously providing correctness and fault tolerance guarantees. Subsequently, I will give a flavor of the challenges involved and the techniques developed across different graph processing problems by diving deeper into my contributions for two graph processing scenarios. In context of distributed graph processing, I will present a confined recovery strategy, to handle machine failures, that exploits the algorithmic asynchrony to reconstruct a valid graph execution. For streaming graph processing, I will show a dynamic dependence based correction strategy to provide fast and accurate results for real-time queries. The talk will conclude with a brief discussion on future research directions.
Keval Vora is a Ph.D. candidate in the Department of Computer Science and Engineering at the University of California, Riverside where he is advised by Prof. Rajiv Gupta. His research interests lie in the broad area of Parallel and Distributed Computing including their Programmability, Performance, Scalability and Fault Tolerance. In particular, his work addresses challenges involved in processing large-scale static and dynamic graphs across distributed and single machine based processing environments. The results of above research have been published in top venues including USENIX ATC, ASPLOS, OOPSLA, HPDC and TACO.