Title:  Near-data Processing for Dynamic Graph Analytics

Committee: 

Dr. Conte, Advisor       

Dr. Yalamanchili, Chair

Dr. Bader

Abstract:

The objective of the proposed research is to develop novel hardware and software solutions that allow dynamic graph analytics workloads to fully utilize the performance gains of emerging high-bandwidth memory architectures. The exponential growth in processing power and storage capacity of high-performance computing systems has enabled rapid advances in the field of graph analytics, delivering breakthroughs in the fields of cybersecurity, simulation, and social media analysis. The data structures that enable streaming graph analytics pose unique challenges for HPC system designers. When the sorted, contiguous arrays of static graphs are replaced with the fragmented, linked data structures of dynamic graphs, workloads struggle to reach the memory bandwidth saturation point. Because these graphs are constantly changing, they are difficult to evenly distribute across memory banks. Behaviors such as pointer-chasing and poor spatial locality expose the true latency of modern memory devices, which has not kept up with processor clock rates.  Emerging systems target these problems with high-bandwidth memory and near-data processing paradigms. While these techniques have proven effective for a wide range of memory-intensive applications, most experimental evaluations only consider static graphs. This proposal aims to use the DynoGraph benchmark suite to evaluate the suitability of high-bandwidth memory architectures for streaming graph processing, culminating in the design of a custom near-memory accelerator for dynamic data structures.