GPU数据洗牌框架实现超高吞吐量 RapidsMPF优化内存溢出机制

RapidsMPF分布式数据洗牌库在NVIDIA的DGX B200系统上实现了约1.8 TiB/s的全局吞吐量1，可处理超过GPU显存容量的海量数据。该框架通过智能内存溢出机制解决显存压力问题，允许接收端将数据直接存储在主机内存中以避免内存溢出1。

在DGX B200的典型配置下——配备8个Blackwell GPU、每个180GB显存——基础测试使用160GB总数据量1。当设备内存被限制为32GB时，吞吐量降至477 GiB/s，受PCIe Gen 5传输速率约32 GiB/s的制约1。更严苛的场景中，当内存限制降至20GB（等于单个输入大小）时，真正的设备到主机溢出机制才会触发，此时copy-pinned_host-to-device操作的耗时从171.8毫秒增至1170毫秒1。

RapidsMPF支持UCXX加速传输协议，兼容NVLink、InfiniBand、EFA、TCP等多种网络拓扑1。该框架已被多个主流项目采纳，其中NeMo-Curator进行了实验性集成、Ray Data进行试用，cuDF Polars则已正式使用1。

NVIDIA's RapidsMPF framework has demonstrated significant advances in GPU-accelerated data shuffling, achieving approximately 1.8 TiB/s global throughput on a single DGX B200 system 1. The distributed data shuffling library is designed to handle datasets exceeding GPU memory capacity while preventing unnecessary data movement through intelligent spillover mechanisms 1.

The framework leverages the computational power of eight Blackwell GPUs, each equipped with 180GB of memory, in baseline testing configurations using 160GB of total data 1. When device memory is constrained to 32GB, throughput decreases to 477 GiB/s due to PCIe Gen 5 transfer rate limitations of approximately 32 GiB/s 1. More severe memory restrictions—specifically when limited to 20GB, equivalent to a single input size—trigger genuine device-to-host spillover, with copy-pinned host-to-device operations extending from 171.8 milliseconds to 1170 milliseconds 1.

RapidsMPF supports accelerated cross-device transfers through UCXX, enabling communication across multiple network architectures including NVLink, InfiniBand, EFA, and TCP 1. The library has achieved adoption across several major projects: it is being experimentally integrated into NeMo-Curator, piloted by Ray Data, and formally implemented in cuDF Polars 1.