Lin Hu’s paper of graph pattern mining framework on GPUs “A Graph Pattern Mining Framework for Large Graphs on GPU” has been accepted by VLDBJ 2024. This is the extension version of Lin Hu’s paper “GAMMA：A Graph Pattern Mining Framework for Large Graphs on GPU” on ICDE 2023.

Graph pattern mining (GPM) is getting increasingly important recently. There are many parallel frameworks for GPM, many of which suffer from performance. GPU is a powerful option for graph processing, which has excellent potential for performance improvement; however, parallel GPM algorithms produce a large number of intermediate results, limiting GPM implementations on GPU. In this paper, we present GAMMA, an out-of-core GPM framework on GPU, and it makes full use of host memory to process large graphs. Specifically, GAMMA adopts a self-adaptive implicit host memory access manner to achieve high bandwidth, which is transparent to users. GAMMA provides flexible and effective interfaces for users to build their algorithms. We also propose several optimizations over primitives provided by GAMMA in the out-of-core GPU system. Experimental results show that GAMMA has scalability advantages in graph size over the state-of-the-art by an order of magnitude, and is also faster than existing GPM systems.

Compared with its conference version, this paper features new optimizations proposed for GPU-accelerated GPM. We extended GAMMA to support multi-GPU scenario based on GAMMA’s prediction model for data hotness and implicit data transfer; we proposed a parallel strategy for multi-set intersection on GPUs, which is common in GPM; we optimized the data layout on GPUs’ shared memory to reduce bank conflicts. Through experimental analysis, we prove the effectiveness of these optimizations.