GPU-Accelerated Presto for Low-Latency Analytics

7/8/2026

This post demonstrates the performance of GPU-accelerated Presto on NVIDIA DGX B200 and GB200 NVL72 systems for low-latency analytical workloads. It highlights performance improvements of up to 8x lower latency compared to CPU-based Presto on TPC-H benchmarks using single-node DGX B200. It further details scaling out to GB200 NVL72, showcasing the benefits of NVIDIA GPUDirect Storage (GDS) with IBM Storage Scale for high I/O throughput. The post details specific optimizations like increasing I/O task size, using more I/O threads, and query rewrites (e.g., Q11) that led to a 64% reduction in query runtimes. It also mentions the importance of NVLink for GPU-to-GPU communication and UcxExchange for high-performance communications.

6/30/2026

This post introduces NVIDIA GQE (GPU Query Engine), a reference architecture for executing SQL queries on large datasets using modern NVIDIA hardware. It details GQE's three-layer architecture (query, data, execution) and its reliance on cuDF and other CUDA-X libraries. Key technical contributions include: 1. Optimized in-memory table layout with row groups and non-contiguous columnar partitions for efficient transfer. 2. Pipelined host-to-device transfer orchestration that overlaps scheduling, H2D transfer, decompression, and CUDA kernel execution across concurrent streams. 3. Data transfer optimizations through compression using NVIDIA nvCOMP (Cascaded and LZ4 algorithms) and dedicated Blackwell Decompression Engine, and aggressive partition pruning using zone maps. These techniques collectively achieve a 7.5x aggregate speedup over state-of-the-art CPU databases on TPC-H SF1000.