AI Practices AI实践 2d ago Updated 2d ago 更新于 2天前 56

Running Low-Latency Analytical Workloads with GPU-Accelerated Presto on NVIDIA GB200 NVL72 在NVIDIA GB200 NVL72上使用GPU加速Presto运行低延迟分析工作负载

GPU-accelerated Presto on NVIDIA GB200 NVL72 achieves up to 8x lower latency than multinode CPU clusters for TPC-H analytical benchmarks. Performance gains are driven by NVIDIA cuDF for query execution, NVLink 5.0 for high-bandwidth GPU-to-GPU communication, and GPUDirect Storage for direct data transfer. Cluster-level optimizations, including increased I/O task sizes and query rewriting, yield up to 64% faster query runtimes by reducing CPU involvement and NUMA penalties. Single-node DGX B200 c GPU加速的Presto在NVIDIA GB200 NVL72及DGX B200硬件上运行TPC-H基准测试时,相比多节点CPU集群可实现最高8倍的延迟降低。 性能提升主要得益于NVIDIA cuDF查询执行引擎、NVLink高速GPU间通信以及GPUDirect Storage (GDS)实现的存储到GPU内存的直接数据传输。 通过增加I/O任务大小、优化线程配置及重写查询以适配GPU,整体I/O和通信效率提升带来高达64%的运行时间缩短。 单节点DGX B200(8 GPU)在1TB数据集上比8节点CPU集群快8.2倍,在3TB数据集上比10节点CPU集群快7.8倍。 在GB200 NVL

75
Hot 热度
85
Quality 质量
80
Impact 影响力

Analysis 深度分析

TL;DR

  • GPU-accelerated Presto on NVIDIA GB200 NVL72 achieves up to 8x lower latency than multinode CPU clusters for TPC-H analytical benchmarks.
  • Performance gains are driven by NVIDIA cuDF for query execution, NVLink 5.0 for high-bandwidth GPU-to-GPU communication, and GPUDirect Storage for direct data transfer.
  • Cluster-level optimizations, including increased I/O task sizes and query rewriting, yield up to 64% faster query runtimes by reducing CPU involvement and NUMA penalties.
  • Single-node DGX B200 configurations outperform 8-10 node Intel Xeon CPU clusters, demonstrating the efficiency of consolidating workloads onto fewer, more powerful GPU nodes.
  • Scaling to the GB200 NVL72 system (18 nodes) with IBM Storage Scale enables high-throughput processing of multi-terabyte datasets with minimal I/O bottlenecks.

Why It Matters

This development signals a paradigm shift in data analytics infrastructure, proving that GPU-accelerated engines can significantly outperform traditional CPU-based distributed systems for large-scale analytical workloads. For AI practitioners and data engineers, it highlights the critical importance of optimizing the entire stack—from storage protocols like GPUDirect Storage to interconnect technologies like NVLink—to unlock the full potential of modern AI hardware. This efficiency gain reduces operational costs and latency, enabling faster iteration cycles for data-intensive applications and real-time analytics.

Technical Details

  • Hardware Architecture: Benchmarks were conducted on NVIDIA DGX B200 (single-node, 8 GPUs) and GB200 NVL72 (multinode, 18 nodes, 72 B200 GPUs total). The GB200 NVL72 features NVLink connectivity between all GPUs and ConnectX-7 NICs for 400 Gbps networking.
  • Software Stack: Utilizes GPU-accelerated Presto with NVIDIA cuDF for query execution and Velox for data processing. Communication between GPU workers is optimized using UcxExchange.
  • Storage Integration: Paired with IBM Storage Scale (formerly GPFS), leveraging NVIDIA GPUDirect Storage (GDS) to enable Remote Direct Memory Access (RDMA) from storage directly to GPU memory, bypassing host CPU and system memory.
  • Benchmark Methodology: Used TPC-H derived queries (22 analytical queries) on Parquet datasets at scale factors 1K (~1 TB), 3K (~3 TB), 10K, and 30K. Runtimes included SQL parsing, plan optimization, worker execution, and result return, averaged over multiple runs.
  • Optimization Techniques: Key improvements came from increasing I/O task size, utilizing more I/O threads, rewriting queries for GPU efficiency, and tuning UcxExchange configurations for inter-node communication.

Industry Insight

  • Consolidation Opportunity: Organizations can reduce infrastructure complexity and cost by replacing large CPU clusters with fewer, high-performance GPU nodes without sacrificing throughput or latency.
  • End-to-End Optimization: Maximizing performance requires holistic optimization across the stack, particularly focusing on I/O paths (GDS) and interconnects (NVLink), rather than just raw compute power.
  • Real-Time Analytics Viability: The demonstrated latency reductions make interactive, real-time analytics on massive datasets feasible, opening new possibilities for AI agents and data-driven decision-making systems.

TL;DR

  • GPU加速的Presto在NVIDIA GB200 NVL72及DGX B200硬件上运行TPC-H基准测试时,相比多节点CPU集群可实现最高8倍的延迟降低。
  • 性能提升主要得益于NVIDIA cuDF查询执行引擎、NVLink高速GPU间通信以及GPUDirect Storage (GDS)实现的存储到GPU内存的直接数据传输。
  • 通过增加I/O任务大小、优化线程配置及重写查询以适配GPU,整体I/O和通信效率提升带来高达64%的运行时间缩短。
  • 单节点DGX B200(8 GPU)在1TB数据集上比8节点CPU集群快8.2倍,在3TB数据集上比10节点CPU集群快7.8倍。
  • 在GB200 NVL72大规模集群中,结合IBM Storage Scale与GDS技术,有效解决了大规模数据分析中的I/O瓶颈和NUMA惩罚问题。

为什么值得看

本文展示了GPU原生加速引擎在超大规模数据分析场景下的极致性能,证明了单节点多GPU系统可超越传统多节点CPU集群,为构建超低延迟的分析型数据仓库提供了新范式。对于追求实时AI代理交互和高并发查询响应的企业而言,理解如何利用NVLink和GDS优化I/O路径是提升系统效能的关键。

技术解析

  • 硬件与架构优势:测试基于NVIDIA DGX B200(单节点8 GPU,NVLink 5.0带宽1,800 GB/s)和GB200 NVL72(18节点,每节点4 GPU,全互联NVLink)。GB200 NVL72还配备了ConnectX-7网卡,支持计算与存储网络流量复用。
  • 关键软件栈:采用GPU原生的Presto执行引擎,底层依赖NVIDIA cuDF进行高性能查询执行,并使用Velox作为通用执行框架。通信层利用UcxExchange优化GPU Worker间的数据交换。
  • I/O优化技术:引入NVIDIA GPUDirect Storage (GDS),配合IBM Storage Scale并行文件系统,实现数据从存储设备直接传输至GPU显存,绕过主机CPU和系统内存缓冲区,显著减少CPU负载和NUMA访问延迟。
  • 基准测试细节:使用TPC-H衍生基准(22个分析查询),数据源为Parquet文件。测试涵盖了Scale Factor 1K (~1TB)、3K (~3TB)、10K和30K等不同规模,测量包括SQL解析、计划优化、Worker执行及结果返回的全链路耗时。
  • 集群级调优策略:通过增加I/O任务大小、启用更多I/O线程以及针对GPU利用率重写查询逻辑,实现了显著的集群级性能提升,特别是在大规模扩展场景下优化了通信开销。

行业启示

  • 分析型工作负载的硬件重构:传统基于CPU的多节点横向扩展模式在超低延迟分析场景中可能不再是最优解,单节点高密度GPU集群凭借内部高带宽互联(如NVLink)可提供更具性价比和性能密度的替代方案。
  • 存储与计算边界的融合:GDS等技术的成熟使得“存算一体”或“近存计算”成为可能,企业应重新评估其数据湖架构,优先选择支持RDMA和直接GPU内存映射的存储解决方案,以消除I/O瓶颈。
  • 软件栈的GPU原生化转型:为了充分发挥新一代AI基础设施(如GB200)的性能,数据分析引擎必须从CPU-centric转向GPU-native设计(如使用cuDF/Velox),开发者需关注查询重写和I/O参数调优以适应新的硬件拓扑。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPU GPU Benchmark 基准测试 Inference 推理 Research 科学研究