PG-Strom

PG-Strom is an open-source extension module for PostgreSQL, designed for version 15 or later, that offloads SQL workload processing to GPU devices to accelerate data analytics and batch processing on large datasets. It targets users who need to run heavy analytical queries on PostgreSQL without migrating to a separate data warehouse, such as data engineers, analysts, and organizations dealing with big data. By leveraging GPU devices with thousands of cores per chip, PG-Strom aims to speed up operations like SCAN (WHERE-clause evaluation), JOIN, and GROUP BY, transparently replacing PostgreSQL's vanilla execution plans when GPU processing offers an advantage. The extension is developed by HeteroDB and distributed under the PostgreSQL License, with active community support through GitHub discussions and issue trackers.

PG-Strom's core technology includes an automatic GPU code generator that produces GPU programs directly from SQL commands, and an asynchronous parallel execution engine that runs these workloads on the GPU. A standout feature is GPUDirect SQL, which bypasses CPU and RAM to read data directly from NVMe SSDs or NVMe-oF storage to the GPU, maximizing bandwidth. Version 3.0 added support for NVIDIA GPUDirect Storage, enabling software-defined storage over NVMe-oF and shared filesystems. The extension offers two storage options: PostgreSQL's native heap storage for row-based data, and Apache Arrow files for efficient columnar data processing via a foreign data wrapper. Version 5.0 overhauled the architecture, switching from a multi-process to a multi-threaded model and replacing CUDA C++ native code with portable pseudo-code for better performance and stability. PG-Strom also supports execution of PostGIS functions and GiST index searches on the GPU, and includes a GPU cache for frequently updated table contents.

In the PostgreSQL extension ecosystem, PG-Strom occupies a unique niche by providing GPU acceleration, a capability not offered by standard extensions like pg_stat_statements or PostGIS. Compared to Cypher or TimescaleDB, which focus on graph or time-series workloads, PG-Strom targets general analytical SQL acceleration. Its main competition comes from proprietary solutions like BryteFlow or Amazon Redshift Spectrum, but PG-Strom is open source and integrates directly into PostgreSQL, avoiding data migration. The CMU Database Group's extension analysis notes PG-Strom uses extensibility types including Functions, Types, Index Access Methods, Storage Managers, Client Authentication, and Query Processing, with 81,381 lines of code and low duplication (0.46%). It supports PostgreSQL versions 15 and 16.

The honest trade-offs with PG-Strom are significant. It requires specific hardware: at least one NVIDIA GPU with compute capability 7.5 (Turing generation) or later, and for GPUDirect SQL, NVMe SSDs under the same PCIe root complex as the GPU. The software stack demands CUDA Toolkit 12.2 update1 or later, a supported Linux distribution (Red Hat Enterprise Linux or Rocky Linux 8.x/9.x/10.x), and for GPUDirect Storage, the nvidia-fs driver and Mellanox OFED. IOMMU must be disabled for GPU-Direct SQL, which can be a security or configuration concern. PG-Strom is limited to PostgreSQL, so users on other databases cannot benefit. Setup and tuning require technical expertise, and the extension's performance gains are workload-dependent—ideal for large analytical queries but not for simple transactional operations. The community is active but smaller than mainstream PostgreSQL extensions, and documentation is primarily in English with some Japanese support.

Open source under the PostgreSQL License, free to use, modify, and distribute.