Apply Memory Advice with a Specific Device ID

When using unified memory on systems with CUDA GPUs, various types of memory advice can be applied to modify how the CUDA runtime moves this memory around between the CPU and GPU. When applying memory advice, a device ID can be used to specific which device the advice relates to. One type of advice that can be applied is “preferred location”, and you can specificy where you want the preferred location of the memory to be. This can be useful for ensuring that the memory is kept on the GPU.

By passing a specific device id when constructing an umpire::strategy::AllocationAdvisor, you can ensure that the advice will be applied with respect to that device

            "preferred_location_device_2", allocator, "PREFERRED_LOCATION",

The complete example is included below:

// Copyright (c) 2016-20, Lawrence Livermore National Security, LLC and Umpire
// project contributors. See the COPYRIGHT file for details.
// SPDX-License-Identifier: (MIT)
#include <iostream>

#include "umpire/Allocator.hpp"
#include "umpire/ResourceManager.hpp"
#include "umpire/strategy/AllocationAdvisor.hpp"
#include "umpire/util/Exception.hpp"

int main(int, char**)
  auto& rm = umpire::ResourceManager::getInstance();

  auto allocator = rm.getAllocator("UM");

   * Create an allocator that applied "PREFFERED_LOCATION" advice to set a
   * specific GPU device as the preferred location.
   * In this case, device #2.
  const int device_id = 2;

  try {
    auto preferred_location_allocator =
            "preferred_location_device_2", allocator, "PREFERRED_LOCATION",

    void* data = preferred_location_allocator.allocate(1024);

  } catch (umpire::util::Exception& e) {
    std::cout << "Couldn't create Allocator with device_id = " << device_id
              << std::endl;

    std::cout << e.message() << std::endl;

  return 0;