3. Getting Started#

3.1. Hardware Requirements#

  • x86_64 or ARM64 (Grace CPU only) CPU architecture

  • Mellanox ConnectX-5 and above InfiniBand HCA or RoCE-capable Ethernet adapter

  • NVIDIA GPU with CUDA Compute Capability 6.0 or higher (for GPU memory operations)

  • Minimum 16GB system RAM recommended

3.2. Software Requirements#

  • Linux operating system (tested on Ubuntu 22.04/24.04, RHEL 9/10)

  • CUDA Toolkit 13.1.1 or later

  • NVIDIA GPU Driver compatible with CUDA 13.1.1

  • InfiniBand/RDMA drivers and libraries (libibverbs, librdmacm)

  • CUDA runtime libraries for GPU memory support

  • GPUDirect Storage (GDS) libraries (for GPU/system memory support)

  • C++14 or later compatible compiler

3.3. Installation#

  1. Install DOCA and RDMA Libraries

    Use the DOCA installation guide to install DOCA.

  2. Install cuObject client Libraries

    cuObject client libraries are part of cuda-toolkit. Follow the instructions from cuobject client library downloads

    Install libcuobjclient and libcufile from the same CUDA Toolkit release. Combining an updated cuObject client with an older libcufile is unsupported and can cause library loading failures.

    Configure the client’s local RDMA interfaces or IP addresses in /etc/cuobj.json, or select an alternate configuration file using CUOBJ_ENV_PATH_JSON. Set the configuration environment before the first runtime initialization.

    Applications that previously used CUFILE_ENV_PATH_JSON to select a cuObject configuration must use CUOBJ_ENV_PATH_JSON. The cuObject client does not use CUFILE_ENV_PATH_JSON or /etc/cufile.json for configuration-file selection.

  3. Verify Installation

    # Check library presence
    $ ldconfig -p | grep cuobj
    $ ldconfig -p | grep cufile
    
    # Verify cuObjClient installation
    $ rpm -qa | grep cuobjclient    # RHEL/CentOS
    $ dpkg -l | grep cuobjclient    # Ubuntu/Debian
    
    # Verify GDS support
    $ /usr/local/cuda/gds/tools/gdscheck -p