Round-Robin Policy
The dynamic selection API is an experimental feature in the Intel® oneAPI DPC++ Library (oneDPL) that selects an execution resource based on a chosen selection policy. There are several policies provided as part of the API. Policies encapsulate the logic and any associated state needed to make a selection.
The round-robin policy cycles through the set of resources at each selection. round_robin_policy is useful for offloading kernels of similar cost to devices of similar capabilities. In those cases, a round-robin assignment of kernels to devices will achieve a good load balancing.
namespace oneapi::dpl::experimental {
template<typename Backend = sycl_backend>
class round_robin_policy {
public:
// useful types
using resource_type = typename Backend::resource_type;
using wait_type = typename Backend::wait_type;
class selection_type {
public:
round_robin_policy<Backend> get_policy() const;
resource_type unwrap() const;
};
// constructors
round_robin_policy(deferred_initialization_t);
round_robin_policy();
round_robin_policy(const std::vector<resource_type>& u);
// deferred initializer
void initialize();
void initialize(const std::vector<resource_type>& u);
// queries
auto get_resources() const;
auto get_submission_group();
// other implementation defined functions...
};
}
This policy can be used with all the dynamic selection functions, such as select, submit, and submit_and_wait. It can also be used with policy_traits.
Example
The following example demonstrates a simple approach to send work to each queue in a set of queues, and then wait for all devices to complete the work before repeating the process. A round_robin_policy is used rotate through the available devices.
#include <oneapi/dpl/dynamic_selection>
#include <sycl/sycl.hpp>
#include <iostream>
const std::size_t N = 10000;
namespace ex = oneapi::dpl::experimental;
void f(sycl::handler& h, float* v);
int round_robin_example(std::vector<sycl::queue>& similar_devices,
std::vector<float*>& usm_data) {
ex::round_robin_policy p{similar_devices}; // (1)
auto num_devices = p.get_resources().size();
auto num_arrays = usm_data.size();
// (2)
auto submission_group_size = (num_arrays < num_devices) ? num_arrays : num_devices;
std::cout << "Running with " << num_devices << " queues\n"
<< " " << num_arrays << " usm arrays\n"
<< "Will perform " << submission_group_size << " concurrent offloads\n";
for (int i = 0; i < 100; i += submission_group_size) { // (3)
for (int j = 0; j < submission_group_size; ++j) { // (4)
ex::submit(p, [&](sycl::queue q) { // (5)
float* data = usm_data[j];
return q.submit([=](sycl::handler &h) { // (6)
f(h, data);
});
});
}
ex::wait(p.get_submission_group()); // (7)
}
return 0;
}
The key points in this example are:
A round_robin_policy is constructed that rotates between the CPU and GPU queues.
The total number of concurrent offloads, submission_group_size, will be limited to the number of USM arrays or the number of queues, whichever is smaller.
The outer i-loop iterates from 0 to 99, stepping by the submission_group_size. This number of submissions will be offload concurrently.
The inner j-loop iterates over submission_group_size submissions.
submit is used to select a queue and pass it to the user’s function, but does not block until the event returned by that function completes. This provides the opportunity for concurrency across the submissions.
The queue is used in a function to perform an asynchronous offload. The SYCL event returned from the call to submit is returned. Returning an event is required for functions passed to submit and submit_and_wait.
wait is called to block for all the concurrent submission_group_size submissions to complete.
Selection Algorithm
The selection algorithm for round_robin_policy rotates through the elements of the set of available resources. A simplified, expository implementation of the selection algorithm follows:
template<typename ...Args>
selection_type round_robin_policy::select(Args&&...) {
if (initialized_) {
auto& r = resources_[next_context_++ % num_resources_];
return selection_type{*this, r};
} else {
throw std::logic_error(“selected called before initialization”);
}
}
where resources_ is a container of resources, such as std::vector of sycl::queue, next_context_ is a counter that increments at each selection, and num_resources_ is the size of the resources_ vector.
Constructors
round_robin_policy provides three constructors.
Signature |
Description |
|---|---|
round_round_policy(deferred_initialization_t); |
Defers initialization. An initialize function must be called prior to use. |
round_robin_policy(); |
Initialized to use the default set of resources. |
round_robin_policy(const std::vector<resource_type>& u); |
Overrides the default set of resources. |
Deferred Initialization
A round_robin_policy that was constructed with deferred initialization must be initialized by calling one its initialize member functions before it can be used to select or submit.
Signature |
Description |
|---|---|
initialize(); |
Initialize to use the default set of resources. |
initialize(const std::vector<resource_type>& u); |
Overrides the default set of resources. |
Queries
A round_robin_policy has get_resources and get_submission_group member functions.
Signature |
Description |
|---|---|
std::vector<resource_type> get_resources(); |
Returns the set of resources the policy is selecting from. |
auto get_submission_group(); |
Returns an object that can be used to wait for all active submissions. |
Reporting Requirements
If a resource returned by select is used directly without calling submit or submit_and_wait, it may be necessary to call report to provide feedback to the policy. However, the round_robin_policy does not require any feedback about the system state or the behavior of the workload. Therefore, no explicit reporting of execution information is needed, as is summarized in the table below.
execution_info |
is reporting required? |
|---|---|
task_submission |
No |
task_completion |
No |
task_time |
No |
In generic code, it is possible to perform compile-time checks to avoid reporting overheads when reporting is not needed, while still writing code that will work with any policy, as demonstrated below:
auto s = select(my_policy);
if constexpr (report_info_v<decltype(s), execution_info::task_submission_t>)
{
s.report(execution_info::task_submission);
}