Unified Collective Communication (UCC) provides a flexible, feature-rich API for collective communication operations, targeting HPC, AI/ML, and I/O workloads. It aims to deliver highly scalable and performant collectives across various programming models and runtimes, supporting nonblocking operations, flexible resource allocation, and hardware-specific collectives.

How It Works

UCC is built on a component architecture, leveraging UCX (Unified Communication X) for its transport layer. It supports hardware collectives as first-class citizens, enabling optimized communication patterns for GPUs and other accelerators. The design emphasizes flexibility in resource management and synchronization, allowing for repetitive collective operations with a single initialization.

Quick Start & Requirements

• Install: Compile from source using autogen.sh, configure --prefix=<ucc-install-path> --with-ucx=<ucx-install-path>, and make -j install.
• Prerequisites: UCX is required. CUDA (11.0+) or HIP are optional for GPU support. Doxygen is used for documentation generation.
• Setup: Requires compiling UCX and potentially Open MPI if integrating with MPI.
• Links: UCX, Open MPI

Highlighted Details

• Supports UCX/UCP transports including InfiniBand, RoCE, Shared Memory, and hardware offload via SHARP.
• Integrates with CUDA (NCCL) and HIP (RCCL) for GPU-accelerated collectives.
• Enables UCC collectives within Open MPI and OpenSHMEM applications via specific MCA parameters.
• Offers a flexible synchronous model and supports repetitive collective operations.

Maintenance & Community

The project is associated with the openucx organization. Further community and contribution details are available in the CONTRIBUTING file.

Licensing & Compatibility

UCC is BSD-style licensed, permitting commercial use and integration with closed-source applications.

Limitations & Caveats

Compilation requires building UCX and potentially Open MPI from source. Specific hardware transports and GPU support depend on correctly installing and configuring CUDA or HIP.