High-Performance Graph Processing

We design computation- and memory-optimized scalable parallel and distributed graph algorithms and processing models. This research sits at the intersection between high-performance computing, computer architecture, and computer networks. Some of this work uses approximation, or transprecise computing.

ParaGrapher [2023-2024] is a high-performance graph loading API and library that helps HPC researchers with immediate access to a wide range of real-world graphs, especially the large ones that are published in compressed formats and require deploying other frameworks and programming languages.

MS-BioGraphs Datasets [2022 – 2023] are a family of trillion-scale public real-world graph datasets and the largest public real-world graphs at the time of publication (Aug. 2023). The datasets have been created by matching similarity between 1.7 billion proteins and contains up to 2.5 trillion edges.

Subgraph Isomorphism [2021 – ongoing] studies aims to enhance the efficiency and effectiveness of subgraph isomorphism search, by adopting heuristics dynamically, embracing vector-based embeddings, and harnessing parallelism.

Approximate Maximum Weighted Clique [2021 – ongoing] is the search for an efficient (polynomial time) algorithm to solve the maximum weighted clique problem while finding near-optimal weighted cliques.

Graptor [2019 – ongoing] enhances GraphGrind and VEBO by providing Single Instruction Multiple Data (SIMD) vectorization for graph analytics using the SSE, AVX2 and AVX-512 instruction sets families.

LaganLighter  [2019 – 2022]: Having the experience of GraphGrind and VEBO, we understood the requirement of re-studying graph algorithms based on the implications imposed by the structure of datasets into the execution of graph analytics. LaganLighter introduces structure-aware high-performance graph algorithms.

VEBO [2017 – 2018] provides load balancing and better CPU utilization by applying a new graph relabeling and partitioning algorithm. The assumption behind VEBO is that load balance is a more important consideration for shared-memory graph partitioning than memory locality. VEBO has been implemented for GraphGrind and also for Polymer.

GraphGrind [2014 – 2017] optimizes Ligra and Polymer for systems with Non-Uniform Memory-Architecture (NUMA) by adapting the parallelisation approach (vertices and edges partitioning and scheduling algorithms) to the graph algorithm, by compressing of topology data, and by introducing a new NUMA-aware work-stealing algorithm that extends Cilk.