Intel® oneAPI Collective Communications Library

Intel® oneAPI Collective Communications Library (oneCCL) provides an efficient implementation of communication patterns used in deep learning.

oneCCL features include:

• Built on top of lower-level communication middleware – Intel® MPI Library and libfabrics.

• Optimized to drive scalability of communication patterns by allowing to easily trade off compute for communication performance.

• Works across various interconnects: InfiniBand*, Cornelis Networks*, and Ethernet.

• Provides common API sufficient to support communication workflows within Deep Learning / distributed frameworks (such as PyTorch*, Horovod*).

oneCCL package comprises the oneCCL Software Development Kit (SDK) and the Intel® MPI Library Runtime components.

Get Started

• Release Notes
• Installation Guide
  • System Requirements
  • Installation using Command Line Interface
  • Learn More
• Sample Application
  • Build details
  • Run the sample
• Use oneCCL package from CMake
  • oneCCLConfig files generation

Developer Guide

• Programming Model
  • Host Communication
  • Device Communication
  • Limitations
• General Configuration
  • Execution of Communication Operations
  • Transport Selection
• Advanced Configuration
  • Selection of Collective Algorithms
  • Low-precision Datatypes
  • Caching of Communication Operations
  • Prioritization of Communication Operations
  • Fusion of Communication Operations
  • Enabling OFI/verbs/dmabuf Support

Developer Reference

• oneCCL API
  • Initialization
  • oneCCL Concepts
  • Communication Operations
  • Generic workflow
  • Error Handling
• Environment Variables
  • Collective Algorithms Selection
  • Workers
  • ATL
  • Multi-NIC
  • Inter Process Communication (IPC)
  • Low-precision datatypes
  • CCL_LOG_LEVEL
  • CCL_ITT_LEVEL
  • Fusion
  • CCL_PRIORITY
  • CCL_MAX_SHORT_SIZE
  • CCL_SYCL_OUTPUT_EVENT
  • CCL_ZE_LIBRARY_PATH
  • Point-To-Point Operations
  • CCL_ZE_TMP_BUF_SIZE

Benchmark User Guide

• oneCCL Benchmark Collective User Guide
  • Build oneCCL Benchmark
  • Run oneCCL Benchmark
  • Example
• oneCCL Benchmark Point-to-Point User Guide
  • Build oneCCL Benchmark
  • Run oneCCL Point-to-Point Benchmark
  • Examples