Graphs are continually growing in size, and processing complex queries, such as multi-hop pattern queries, on them is becoming increasingly important. Although GPUs have received significant attention recently, there is still a notable shortage of efficient GPU-based out-of-memory methods for handling these queries. Three key issues arise when processing multi-hop queries on large-scale graphs using GPUs: the need for an efficient graph format, effective scheduling of accesses to graph partitions on storage, and dynamic buffer management on both the host and GPUs. To address these issues, we propose an efficient GPU-based out-of-memory multi-hop query processing framework called GFlux. Through extensive experiments, we have demonstrated that GFlux significantly improves both the speed and scalability compared to existing state-of-the-art methods.

Recently, full-graph Graph Neural Networks (GNNs) have gained prominence by addressing complex problems such as weather forecasting and material discovery. Existing full-graph training methods do not fully manage intermediate data generated during training and rely on rigid inter-GPU communication, limiting both training speed and scale. We propose FlexGNN, which fully manages intermediate data and adaptively performs inter-GPU communication by generating and optimizing best-effort training execution plans. Extensive experiments demonstrate that FlexGNN significantly outperforms existing full-graph GNN methods in both training speed and scale. Specifically, it is up to 5.4X faster than HongTu and up to 95.5X faster than NeutronStar.

Operator fusion is essentially and widely used in a large number of matrix computation systems in science and industry. The existing distributed operator fusion methods focus on only either low communication cost with the risk of out of memory or large-scale processing with high communication cost. We propose a distributed elastic fused operator called Cuboid-based Fused Operator (CFO) that achieves both low communication cost and large-scale processing. We also propose a novel fusion plan generator called Cuboid-based Fusion plan Generator (CFG) that finds a fusion plan to fuse more operators including large-scale matrix multiplication. We implement a fast distributed matrix computation engine called FuseME by integrating both CFO and CFG seamlessly. FuseME outperforms the state-of-the-art systems including SystemDS by orders of magnitude.

The concept of OLAP query processing is now being widely adopted in various applications. The number of complex queries containing the joins between non-unique keys (called FK-FK joins) increases in those applications. However, the existing in-memory OLAP systems tend not to handle such complex queries efficiently since they generate a large amount of intermediate results or incur a huge amount of probe cost. In this paper, we propose an effective query planning method for complex OLAP queries. It generates a query plan containing n-ary join operators based on a cost model. The plan does not generate intermediate results for processing FK-FK joins and significantly reduces the probe cost. We also propose an efficient processing method for n-ary join operators. We implement the prototype system SPRINTER by integrating our proposed methods into an open-source in-memory OLAP system. Through experiments using the TPC-DS benchmark, we have shown that SPRINTER outperforms the state-of-the-art OLAP systems for complex queries.

Matrix computation, in particular, matrix multiplication is time-consuming, but essentially and widely used in a large number of applications in science and industry. The existing distributed matrix multiplication methods only focus on either low communication cost (i.e., high performance) with the risk of out of memory or large-scale processing with high communication overhead. We propose a distributed elastic matrix multiplication method called CuboidMM that achieves both high performance and large-scale processing. We also propose a GPU acceleration method that can be combined with CuboidMM. CuboidMM partitions matrices into cuboids for optimizing the network communication cost with considering memory usage per task, and the GPU acceleration method partitions a cuboid into subcuboids for optimizing the PCI-E communication cost with considering GPU memory usage. We implement a fast and elastic matrix computation engine called DistME by integrating CuboidMM with GPU acceleration on top of Apache Spark. Through extensive experiments, we have demonstrated that CuboidMM and DistME significantly outperform the state-of-the-art methods and systems, respectively, in terms of both performance and data size.

As the amount of data to process increases, a scalable and efficient horizontal database partitioning method becomes more important for OLAP query processing in parallel database platforms. Existing partitioning methods have a few major drawbacks such as a large amount of data redundancy and not supporting join processing without shuffle in many cases despite their large data redundancy. We elucidate the drawbacks arise from their tree-based partitioning schemes and propose a novel graph-based database partitioning method called GPT that improves query performance with lower data redundancy. Through extensive experiments using three benchmarks, we show that GPT significantly outperforms the state-of-the-art method in terms of both storage overhead and query performance.