Function construct_io_thread_pool

src/utils/lbann_library.cpp:365–399

Setup I/O thread pool that is shared across all models

// Setup I/O thread pool that is shared across all models
std::unique_ptr<thread_pool> construct_io_thread_pool(lbann_comm* comm,
                                                      bool serialized_io)
{
  int max_io_threads = num_free_cores_per_process(comm);
  // Allow the trainer to override the command-line option or environment
  // variable
  if (serialized_io) {
    max_io_threads = 1;
  }

  auto& arg_parser = global_argument_parser();
  int req_io_threads = arg_parser.get<int>(LBANN_OPTION_NUM_IO_THREADS);
  int max_io_rng_banks = arg_parser.get<int>(LBANN_OPTION_MAX_IO_RNG_BANKS);
  // Limit the number of I/O threads to:
  //   < number of available free cores per process
  //   < number of RNG banks provisioned
  // and at least one
  int num_io_threads = std::max(
    std::min(max_io_rng_banks, std::min(max_io_threads, req_io_threads)),
    1);

  auto io_threads_offset = free_core_offset(comm);

  if (comm->am_world_master()) {
    std::cout << "\tNum. I/O Threads: " << num_io_threads
              << " (Limited to # Unused Compute Cores [" << max_io_threads
              << "] # of RNG banks [" << max_io_rng_banks
              << "] or 1) at offset " << io_threads_offset << std::endl;
  }

  auto io_thread_pool = std::make_unique<thread_pool>();
  io_thread_pool->launch_pinned_threads(num_io_threads, io_threads_offset);

  return io_thread_pool;
}

std::unique_ptr<model> build_model_from_prototext(
  int argc,