The Bottleneck in Modern Analytics
Running analytical queries on massive datasets is the bread and butter of data engineering. But as datasets grow into terabytes, traditional CPU-bound query engines start to choke. The bottleneck is often not just compute, but also data movement—between CPU, GPU, storage, and across nodes.
Enter GPU-accelerated Presto. By offloading heavy lifting to NVIDIA GPUs and using technologies like NVLink for high-speed GPU-to-GPU communication and GPUDirect Storage (GDS) for direct storage-to-GPU data paths, Presto can deliver dramatic performance improvements.
This post breaks down the key findings from NVIDIA's recent benchmarks on the GB200 NVL72 system, showing exactly how these technologies work together to deliver up to 8x faster query runtimes compared to traditional CPU clusters.
The Architecture: NVLink, cuDF, and GDS in Action
Single-Node Performance: DGX B200
A single DGX B200 node packs 8 GPUs connected via NVLink 5.0 with 1,800 GB/s bidirectional bandwidth. This allows GPU-accelerated Presto workers to share data at speeds that would be impossible over traditional networking.
Key Benchmark Results (TPC-H derived, 1 TB dataset):
| Configuration | Runtime | Speedup vs 8-node CPU |
|---|---|---|
| 8-node CPU (Intel Xeon) | Baseline | 1x |
| 1 GPU (B200) | 2.5x faster | 2.5x |
| 8 GPU (B200) | 8.2x faster | 8.2x |
At 3 TB scale, the story is similar: 8 GPUs deliver 7.8x speedup over a 10-node CPU cluster. The secret? cuDF algorithms running natively on GPU memory, with zero data copying overhead between GPUs thanks to NVLink.
Multi-Node Scaling: GB200 NVL72 + IBM Storage Scale
When you need to go beyond a single node, the GB200 NVL72 is the answer. This system connects 18 nodes (each with 2 Grace CPUs, 4 B200 GPUs, and 4 ConnectX-7 NICs) through NVLink, forming a single, massive GPU pool.
The I/O Optimization Journey (8 nodes, 32 GPUs, scale factor 10K):
- Baseline (POSIX reads, 4 MiB tasks): Slow, due to bounce buffers and NUMA crossing.
- + GDS + 16 MiB tasks: 30% faster. GPUDirect Storage bypasses the CPU entirely.
- + 16 I/O threads: 17% more speed. Better NVLink saturation.
- + Rebatching + Q11 rewrite: 35% additional gain. Reduced GPU idle time.
Total improvement: 64% faster query runtimes.
The GDS Advantage
GPUDirect Storage (GDS) is a game-changer. Instead of reading data into a CPU buffer and then copying to GPU memory (POSIX path), GDS uses RDMA to move data directly from the storage device (IBM Storage Scale) to GPU memory. This eliminates bounce buffers and NUMA penalties.
In a two-node, eight-GPU test at scale factor 10K, GDS cold reads were ~2x faster than POSIX cold reads. For cost-performance, GDS is the clear winner.
Limitations and Caveats
While the benchmarks are impressive, it's important to keep a few things in mind:
- Workload Dependency: Not all queries benefit equally. Queries with heavy data shuffling or small result sets may not see the same gains. The Q11 rewrite (SELECT to INSERT) is a workaround for a specific bottleneck in result transfer.
- Hardware Requirements: The performance numbers are tied to specific hardware (GB200 NVL72, IBM Storage Scale). Your mileage will vary with different storage systems or network fabrics.
- Cold Cache vs Hot Cache: The benchmarks use 'cold' cache for GDS comparisons. In production with hot caches, the gap might narrow.
- Complexity: Setting up GDS, tuning I/O threads, and optimizing batch sizes requires deep expertise. It's not a plug-and-play solution.
Next Steps for Developers
Ready to try GPU-accelerated Presto? Here’s how to get started:
- Test with Docker: Use the
prestodb/presto:gpu-nightlytag on Docker Hub for a quick trial. - Explore the Build Scripts: NVIDIA provides comprehensive build and deployment scripts in the rapidsai/velox-testing GitHub repo.
- Understand Your Data Path: If you're using a parallel file system like IBM Storage Scale or Lustre, investigate GDS compatibility.
- Profile Your Queries: Not all queries are GPU-friendly. Start with the TPC-H derived benchmark suite to understand the gains.
For a deeper dive into the communication layer between GPU workers, check out our earlier post on Accelerating Large-Scale Data Analytics with GPU-Native Velox and NVIDIA cuDF.
The Future of GPU-Accelerated Analytics
NVIDIA's benchmarks on the GB200 NVL72 show that GPU-accelerated Presto is not just a lab experiment—it's a production-ready solution for cutting query latency by up to 8x. The combination of cuDF, NVLink, and GDS creates a data path that is both faster and more efficient than traditional CPU clusters.
As datasets continue to grow, the ability to scale out GPU compute with technologies like the GB200 NVL72 will become a competitive advantage for organizations running interactive dashboards, nightly ETL, or ad-hoc analytics.
The bottom line: If you're running Presto on CPU and your queries are taking too long, it's time to seriously evaluate GPU acceleration.