Optimizing Data Locality for Scalable Graph Processing and Graph Learning Systems

Speaker: Nikolai Merkel
Technical University of Munich (TUM)

Title: Optimizing Data Locality for Scalable Graph Processing and Graph Learning Systems

Date: Wednesday, 19 March 2025

Time: 2:00pm - 3:00pm

Venue: Room 3598 (via lift 27/28), HKUST

Abstract:

Graphs are a fundamental data structure used in various applications, including web search, recommendation systems, fraud detection, and social network analysis. Processing large-scale graphs efficiently often requires distributed execution, making graph partitioning a crucial pre-processing step. However, selecting the best partitioner for a given workload and input graph is a challenging task. In this talk, we will first present a machine learning-based approach for selecting the most suitable graph partitioner. Then, we will explore how graph reordering and graph partitioning can improve data locality and accelerate (distributed) Graph Neural Network (GNN) training. These techniques help reduce communication overhead and enhance computational efficiency in large-scale graph learning systems.

The talk draws from the following publications:

• Partitioner Selection with EASE to Optimize Distributed Graph Processing, ICDE 23, https://arxiv.org/pdf/2304.04976
• Can Graph Reordering Speed Up Graph Neural Network Training? An Experimental Study, VLDB 2024, https://www.vldb.org/pvldb/vol18/p293-merkel.pdf
• An Experimental Comparison of Partitioning Strategies for Distributed Graph Neural Network Training, EDBT 2025, https://openproceedings.org/2025/conf/edbt/paper-22.pdf

Biography:

Nikolai Merkel is a research associate at the Technical University of Munich (TUM), Germany, specializing in graph processing and graph learning systems. His research focuses on improving data locality and computational efficiency in large-scale graph processing and Graph Neural Network (GNN) training. His work has been published in top-tier venues such as VLDB and ICDE.