Contents

Distributed SPMD model

Refer to Developer Guide: SPMD.

Programming interface

All types and functions in this section are declared in the oneapi::dal::spmd::preview namespace and are available via inclusion of the header file from specified backend.

SPMD distributed model consists of the following components:

  1. Additional train, infer, and compute methods that accept communicator object as the first parameter. Those methods are expected to be called on all ranks to start distributed simulations.

  2. The communicator class that contains methods to perform collective operations among all ranks.

  3. Free functions to create a communicator using a specified communicator backend. Available backends are ccl and mpi.

Usage example

The following listings provide a brief introduction on how to create a particular communicator.

MPI backend

#ifndef ONEDAL_DATA_PARALLEL
#define ONEDAL_DATA_PARALLEL
#endif

#include "oneapi/dal/algo/kmeans.hpp"
#include "oneapi/dal/spmd/mpi/communicator.hpp"

kmeans::model<> run_training(const table& data,
                           const table& initial_centroids) {
   const auto kmeans_desc = kmeans::descriptor<float>{}
      .set_cluster_count(10)
      .set_max_iteration_count(50)
      .set_accuracy_threshold(1e-4);

   auto comm = dal::preview::spmd::make_communicator<dal::preview::spmd::backend::mpi>(queue);
   auto rank_id = comm.get_rank();

   const auto result_train = dal::preview::train(comm, kmeans_desc, local_input);

   if(rank_id == 0) {
      print_table("centroids", result.get_model().get_centroids());
      print_value("objective", result.get_objective_function_value());
   }
   return result.get_model();
}

CCL backend

#ifndef ONEDAL_DATA_PARALLEL
#define ONEDAL_DATA_PARALLEL
#endif

#include "oneapi/dal/algo/kmeans.hpp"
#include "oneapi/dal/spmd/ccl/communicator.hpp"

kmeans::model<> run_training(const table& data,
                           const table& initial_centroids) {
   const auto kmeans_desc = kmeans::descriptor<float>{}
      .set_cluster_count(10)
      .set_max_iteration_count(50)
      .set_accuracy_threshold(1e-4);

   auto comm = dal::preview::spmd::make_communicator<dal::preview::spmd::backend::ccl>(queue);
   auto rank_id = comm.get_rank();

   const auto result_train = dal::preview::train(comm, kmeans_desc, local_input);

   if(rank_id == 0) {
      print_table("centroids", result.get_model().get_centroids());
      print_value("objective", result.get_objective_function_value());
   }
   return result.get_model();
}

previous

Distributed Model: Single Process Multiple Data

next

Communicators