1.11 Third-party storage parameters

Key points

1. GDS (GPUDirect Storage) lets storage move data straight into GPU memory, skipping a copy through CPU memory. NVIDIA's configuration guide covers local (DAS) and network (NAS) storage and benchmarks them with gdsio, the tool installed with GDS.

   What NVIDIA says (2)

   “Local drive configurations (Direct Attached Storage - DAS) and Network storage (Network Attached Storage - NAS) are covered.”

   — GPUDirect Storage Configuration and Benchmarking Guide

   “The benchmarking tool included when GDS is installed, gdsio, is covered and its use demonstrated.”

   — GPUDirect Storage Configuration and Benchmarking Guide

2. RDMA (remote direct memory access) moves data between machines without the CPU copying it. GDS uses the listed client addresses to choose RDMA devices for an NFS mount.

   What NVIDIA says (2)

   “fs:nfs:rdma_dev_addr_list [] Provides the list of IPv4 addresses for all the interfaces a single NFS mount can use.”

   — GPUDirect Storage Configuration and Benchmarking Guide

   “This property is used by the cuFile dynamic routing feature to infer preferred RDMA devices.”

   — GPUDirect Storage Configuration and Benchmarking Guide

3. The storage fabric carries file data between DGX nodes and storage systems. NVIDIA wants high bandwidth per node and fabric features such as adaptive routing (AR). AR spreads traffic across paths.

   What NVIDIA says (2)

   “This is because the I/O per-node for the DGX SuperPOD must exceed 40 GBps.”

   — DGX SuperPOD H100 Reference Architecture: Network Fabrics

   “The storage devices are connected at a 1:1 port to uplink ratio.”

   — DGX SuperPOD H100 Reference Architecture: Network Fabrics

Key terms

• GPUDirect Storage: A direct data path between storage and GPU memory that bypasses a CPU bounce buffer.

Try it

• GPUDirect Storage: simulator lab

Sample question

You attach third-party storage and want to measure GPUDirect Storage throughput. Which tool ships with GDS for this?