HashOrder: Accelerating Graph Processing Through Hashing-based Reordering

Tianyi Zhang, Aditya Desai, Gaurav Gupta, Anshumali Shrivastava

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: graph processing, graph reordering, efficiency, hashing

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Graph processing systems are a fundamental tool across various domains such as machine learning, and their efficiency has become increasingly crucial due to the rapid growth in data volume. A major bottleneck in graph processing systems is poor cache utilization. Graph reordering techniques can mitigate this bottleneck and significantly speed up graph workloads by improving the data locality of the graph memory layout. However, since existing approaches use greedy algorithms or simple heuristics to find good orderings, they suffer from either high computational overhead or suboptimal ordering quality. To this end, we propose HashOrder, a probabilistic algorithm for graph reordering based on randomized hashing. We theoretically show that hashing-based orderings have quality guarantees under reasonable assumptions. HashOrder produces high-quality orderings while being lightweight and parallelizable. We empirically show that HashOrder beats the efficiency-quality tradeoff curve of existing algorithms. Evaluations on various graph processing workloads and GNN data loaders reveal that HashOrder is competitive with or outperforms the existing best method while being 592$\times$ more efficient in reordering, speeding up PageRank by up to 2.49$\times$ and GNN data loaders by up to 2.33$\times$.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8097