引导式测验 · 硬 等级

InfiniBand 和 RDMA 网络测验

测试 RDMA 数据路径、注册缓冲区、队列对、NDR 速率范围、RoCE、流控制和 CPU 设置与批量传输。

相关引导路径Infiniband 和 RDMA 网络
问题 1 的 8

What does an RDMA-capable bulk data path aim to reduce?

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. What does an RDMA-capable bulk data path aim to reduce?
    • Every CPU operation including setup
    • The need to register or protect memory
    • All application failure modes
    • Repeated CPU copying and kernel-path work during buffer transfers
  2. How can an RDMA-capable InfiniBand data path reduce bulk-transfer overhead?
    • By compressing every application buffer
    • By offloading transfers and reducing repeated host copying and kernel processing
    • By removing the need for memory permissions
    • By guaranteeing lower latency for every application
  3. NDR InfiniBand 提供的每链路带宽大致是多少？
    • 40Gb/秒
    • 100Gb/秒
    • 400Gb/秒
    • 10Gb/秒
  4. 在 RDMA 中，“队列对”有什么用？
    • 保存网络适配器处理的发送和接收工作请求
    • 存储每个文件的两个备份副本
    • 将两个 GPU 配对为一个逻辑设备
    • 按顺序加密和解密数据包
  5. 什么是RoCE？
    • 一种新的 GPU 编程语言
    • 基于融合以太网的 RDMA，为以太网带来 RDMA 优势
    • 一种存储压缩格式
    • 一种光缆连接器
  6. InfiniBand 如何使其结构有效无损？
    • 通过丢弃任何太大的数据包
    • 通过将每个数据包发送三次
    • 通过禁用所有错误检查
    • 通过使用基于信用的链路层流量控制来防止缓冲区溢出
  7. Why must an RDMA application manage registered buffer lifetime?
    • The adapter needs valid authorized memory while the transfer runs
    • Registration proves the data are correct
    • Buffers never need completion checks
    • Registration replaces all queue pairs
  8. Does RDMA mean a CPU never participates in networking?
    • Yes; CPUs are removed from the servers
    • Yes; every control operation is a GPU kernel
    • No; software and CPU work remain for setup and management
    • No; RDMA requires every byte to be copied repeatedly by the CPU