FaScalSQL: A Fast and Scalable GPU-Accelerated SQL Query Engine for Out-of-Memory Tables

Graphics Processing Units (GPUs) are promising for analytical SQL query processing, but their limited memory capacity hinders processing large input tables exceeding the GPU memory. The existing engines either 1) statically split input columns into chunks and iteratively perform host-to-GPU transfer and relational operations (streaming engines), or 2) maintain a static and fixed-size cache on GPU memory and distribute input columns and their query workloads to the host CPU and a GPU (CPU-GPU distributive engines). However, we find that they suffer from two primary bottlenecks which eventually lead the engines to severe GPU underutilization: excessive host-to-GPU data movement and CPU-GPU load imbalance. We find that they arise from a conflict between their static input data placement and the dynamic progressive filtering of analytical queries. This conflict leads the engines either to transfer column values that are eventually discarded or to assign a large amount of the workload to the host CPU as the input table size scales. In this paper, we present FaScalSQL, a fast and scalable GPU-accelerated SQL query engine that overcomes the severe GPU underutilization of query processing on out-of-memory tables. FaScalSQL introduces a new type of on-demand CPU-GPU coprocessing engine which exploits both GPU-initiated data transfer and CPU-GPU co-processing capability. It replaces the static large unfiltered chunks with a dynamic GPU-initiated on-demand fetching of necessary input data, guided by the host CPU's prefiltering. We evaluate FaScalSQL with Star Schema Benchmark (SSB) and TPC-H. Using SSB with scale factors of 100 and 200, FaScalSQL achieves geometric mean speedups of 2.60× and 2.20× over the existing streaming and CPU-GPU distributive engines.