Semester of Graduation

Spring 2026

Degree

Master of Science (MS)

Department

Computer Science

Document Type

Thesis

Abstract

Graph analytics are central to applications, ranging from scientific computing to social-network analysis. Many large graph workloads exceed the memory capacity and performance capabilities of a single machine. Existing distributed graph systems often force a trade-off between programmability and performance: high-level platforms can incur substantial runtime overheads, while lower-level distributed libraries frequently expose complex or fragile programming models. This thesis investigates whether modern C++ can support generic distributed graph algorithms that remain both expressive and efficient. We present a prototype distributed graph library in C++ built on the HPX runtime system and guided by the abstractions of the NWGraph library. The work extends NWGraph's concept-based design into distributed-memory by introducing partitioned graph data structures, and locality-aware access patterns using hpx::partitioned_vector. These abstractions preserve a generic programming style while making data ownership and remote access explicit enough to support efficient implementations. Using this framework, we develop distributed versions of Breadth-First Search, PageRank, and Triangle Counting that rely on asynchronous remote actions, partition-local computation, and fine-grained shared-memory parallelism within each node. Experimental results show that this approach can scale to large synthetic and benchmark graphs while remaining competitive with, and often substantially faster than, existing systems. Our results suggest that HPX provides a practical foundation for distributed graph processing in modern C++, and that NWGraph-style generic abstractions can be extended into distributed memory without necessarily sacrificing performance.

Date

3-26-2026

Committee Chair

Kaiser Hartmut

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Share

COinS