Developer Guide and Reference

  • 2021.6
  • 04/11/2022
  • Public Content
Contents

Single Program Multiple Data

This section includes concepts and descriptions of objects that support distributed computations using SPMD model.

Distributed computation using SPMD model

In a typical usage scenario, a user provides a communicator object as a first parameter of a free function to indicate that the algorithm can process data simultaneously. All internal inter-process communications at sync points are hidden from the user.
General expectation is that input dataset is distributed among processes. Results are distributed in accordance with the input.
Example of SPMD Flow in oneDAL
Typical SPMD flow

Supported Collective Operations

The following collective operations are supported:
  • bcast
    Broadcasts data from specified process.
  • allreduce
    Reduces data among all processes.
  • allgatherv
    Gathers data from all processes and shares the result among all processes.
  • sendrecv_replace
    Sends and receives data using a single buffer.

Backend-specific restrictions

  • oneCCL: Allgetherv does not support arbitrary displacements. The result is expected to be closely packed without gaps.
  • oneMPI: Collective operations in oneMPI do not support asynchronous executions. They block the process till completion.

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.