TechPrufer
Get started
NVIDIA
NVIDIA logoNVIDIA · Feature trail

GPU-Accelerated Presto for Low-Latency Analytics

2 posts · Jun 2026 – Jul 2026

What this trail covers

This feature thread tracks the evolution of GPU-accelerated query engines, focusing on optimizing data layout, transfer orchestration, and compression for low-latency analytics. Initial efforts demonstrated significant performance gains with GPU-accelerated Presto on NVIDIA DGX B200 systems. Subsequent developments, exemplified by NVIDIA GQE, introduce a reference architecture leveraging modern NVIDIA hardware like HBM, NVLink-C2C, and dedicated decompression engines in GB200 NVL72. This post details GQE's architecture, data layout, and transfer optimizations, including hybrid compression strategies and pipelined data movement, achieving substantial speedups on benchmarks.

All NVIDIA trails
  1. Running Low-Latency Analytical Workloads with GPU-Accelerated Presto on NVIDIA GB200 NVL72 | NVIDIA Technical BlogThis post details the performance improvements achieved by GPU-accelerated Presto on NVIDIA GB200 NVL72 systems, demonstrating up to 8x lower latency compared to CPU-based Presto clusters. It highlights the critical role of NVIDIA GPUDirect Storage (GDS) for high I/O throughput when paired with IBM Storage Scale, and optimizations in data transfer (increasing I/O task size, using more I/O threads) and communication (UcxExchange configuration). The post also showcases the impact of query rewrites, specifically for Q11, which reduced runtime from 50 seconds to 2 seconds by utilizing a GPU-based parquet writer for efficient result storage.
  2. Designing GPU-Accelerated Query Engines with NVIDIA GQE | NVIDIA Technical BlogThis post introduces NVIDIA GQE (GPU Query Engine), a reference architecture for accelerating SQL query execution 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 contributions include: 1) An in-memory data layout organized into row groups and partitions for efficient transfer. 2) Pipelined host-to-device transfer orchestration that overlaps scheduling, H2D transfer, decompression, and CUDA kernel execution across concurrent CUDA streams. 3) Data transfer optimizations using GPU-accelerated compression via nvCOMP (balancing compression ratio, compression speed, and GPU decompression speed with algorithms like Cascaded and LZ4) and aggressive partition pruning using zone maps. The post demonstrates a 7.5x aggregate speedup on TPC-H SF1000 compared to CPU databases.