Posted in: NVIDIA, Rivermax

Nvidia Mellanox ConnectX Rivermax Version 1.20 released

Nvidia Mellanox ConnectX Rivermax Version 1.20 released

赠人玫瑰,手留余香。


用Mellanox CX系列网卡作为视音频IP传输的都知道Rivermax,其提供一个实现视音频IP传输的完整的实现方案。在Mellanox并入Nvidia之后,Rivermax SDK中并入了Nvidia GPU加速计算技术,实现网卡和显卡的直接数据交换,减少了对CPU的依赖,提高了吞吐,降低了延迟。

GPUDirect看起来不错,实现数据在网卡和显卡之间的交换,这个技术本事需要使用高速PCIe,不知道和RDMA的结果孰强孰弱, 不过已经被Nvidia认为是新颖“The novelty of RoCE/RDMA UDP/IP Multicast for media and data streaming”

No GPUDirect GPUDirect
PCIe Bandwidth 21Gbps 10.5Gbps
CPU Load* 31% 11%

file
file
file

file

Nvidia Rivermax SDK官方地址:
https://developer.nvidia.com/networking/rivermax

file
file

Release Highlights:

  • JT-NM Tested – Aug-22 – SMPTE ST2110 and NMOS –
    https://www.jt-nm.org/jt-nm-tested

  • Performance Improvements

    • Dynamic Header Data Split (allows separating non-constant sized headers from payload)
    • Memory registration API—allows app to create only a single memory key
  • Example Code

    • media_sender:
    • Added example code for sending media streams using GPUDirect
    • Added example code for sending SMPTE ST2022-7 redundant streams using GPUDirect
  • Functionality

    • Supports devices querying—provides list of Rivermax-supported devices
    • Rivermax log outputs to a file without printing to stdout
    • DOCA-Rivermax Rx support on BlueField-2 DPU – Early Access
    • Rx only – Basic API code samples
    • Basic performance tool will be uploaded to NVIDIA/Rivermax GitHub*
    • Full validity checking of API function parameters – disabled by default
  • Supported NVIDIA SDK

    • DeepStream
    • Clara Holoscan Sdk
    • DOCA-Rivermax
    • SDKManager (Easy deployment of Rivermax)
  • GitHub for Add-on samples and extra code examples
    Rivermax GitHub* https://github.com/NVIDIA/Rivermax
    *The Github will be available by mid of Dec-22

  • Integrated with DOCA-Firefly Time Service

    • Rivermax’s strict timing requirements on Windows and Virtualized environment are provided using DOCA-Firefly time service
    • DOCA-Firefly Time Service is now available on public NGC
      • Updated DPU deployment guide is available on the SDK page

Next Release OS support:

  • Linux:
    Ubuntu 20.04, Ubuntu 22.04 LTSnew, 20.04 Canonicalnew, RH8.4new, RH8.6new , Oracle Linux 8.6new
    provided but untested – Ubuntu 18.04, RH 8.1, RH 8.3, Oracle Linux 8.4.
  • Windows:
    Windows Server 2019, Windows Server 2022 , Windows 10 PRO, Windows 11 PRO , Windows 20H2
  • Supported Virtualization Hypervisors (only SRIOV mode):
    VMware ESXi 7.0-U3, Linux KVM (CentOS 8.4) and Linux Guest (for versions see “Supported OSs”), Window client on KVM
  • Obsolete:
    RH 7.8, RH 7.9, CentOS 7.7, CentOS 7.9, CentOS 8.1

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to Top