Progetto di esempio Activity Coordinator

Questo semplice esempio per Activity Coordinator illustra come l'API può essere usata per ripetere il training di un modello in background quando vengono soddisfatte le condizioni di sistema.

Panoramica del progetto di esempio

Si consideri il caso di un'app di modifica della musica. Questa app ha attività in background ad alta priorità richieste dall'utente, ad esempio la pubblicazione di contenuto nell'archiviazione cloud. Esistono anche attività in background con priorità bassa che supportano l'interazione dell'utente, ad esempio fornendo raccomandazioni automatiche per migliorare una composizione durante la modifica. Infine, esistono un set di attività posticipate che non devono essere eseguite in un momento specifico senza la richiesta dell'utente, che è il nostro obiettivo in questo esempio. In particolare, si vuole ripetere periodicamente il training del modello di raccomandazione quando l'impatto dell'utente è minimo. A tale scopo, è possibile usare l'API Activity Coordinator.

Per questo scenario, si vuole ripetere il training del modello quando l'utente non è presente. Il flusso di lavoro di ripetizione del training in questo scenario è anche un consumer GPU, quindi si vuole eseguire anche quando è un buon momento per usare la GPU. È possibile specificare questi requisiti usando i criteri di Activity Coordinator. L'API Activity Coordinator userà i criteri per determinare quando vengono soddisfatti i requisiti e inviare notifiche per quando avviare o interrompere l'esecuzione del lavoro.

In questo caso, il modello di criteri GOOD soddisfa la maggior parte delle esigenze in quanto tiene traccia della CPU, della memoria, del disco di sistema, dell'alimentazione e dell'inattività dell'utente. È sufficiente impostare in modo esplicito una condizione per la GPU. È importante ricordare che anche se il carico di lavoro utilizzerà principalmente la GPU, l'esecuzione dell'attività consuma ancora in modo intrinseco CPU, memoria, disco e potenza. L'impatto su queste risorse può variare notevolmente anche tra le configurazioni di sistema. Ad esempio, una GPU più veloce potrebbe comportare una maggiore quantità di tempo per l'alimentazione della GPU con dati, che potrebbe quindi comportare la lettura o il salvataggio di più dati su disco. La velocità di questo disco può influire anche sull'utilizzo della CPU in modo simile. Configurando tutte le risorse interessate, è possibile assicurarsi di non interferire accidentalmente con l'esperienza utente o ridurre le prestazioni del sistema. Inoltre, il lavoro stesso è stato suddiviso per accadere in piccoli blocchi, in modo da poter rispondere adeguatamente alle notifiche di coordinamento per evitare l'esecuzione al di fuori delle condizioni desiderate.

Per dimostrare in che modo gli sviluppatori possono modificare o effettuare il downgrade dei criteri, si aggiunge anche il requisito che si vuole completare il training entro 48 ore. Le prime 24 ore, la scadenza flessibile, si tenta di eseguire con i criteri ideali e le ultime 24 ore si effettua il downgrade a un criterio minore.

Codice di progetto di esempio

Il codice seguente è l'applicazione di esempio di modifica della musica. Usa l'API Activity Coordinator per eseguire attività in background, come descritto nella panoramica.

#include <chrono>
#include <mutex>
#include <condition_variable>
#include <Windows.h>
#include <ActivityCoordinator.h>
#include <wil/resource.h>

// To use ActivityCoordinator, we must link to the OneCoreUAP library.

#pragma comment(lib, "OneCoreUAP.lib")

using namespace std;
using namespace chrono;
using namespace wil;

// Declare RAII wrappers for the Activity Coordinator policy and subscription.
// These behave like traditional smart pointers and will call their associated
// API cleanup functions when they go out of scope.

typedef wil::unique_any<
        ACTIVITY_COORDINATOR_POLICY,
        decltype(&DestroyActivityCoordinatorPolicy),
        DestroyActivityCoordinatorPolicy>
    unique_policy;

typedef wil::unique_any<
        ACTIVITY_COORDINATOR_SUBSCRIPTION,
        decltype(&UnsubscribeActivityCoordinatorPolicy),
        UnsubscribeActivityCoordinatorPolicy>
    unique_subscription;

struct WORKER_CONTEXT {
    mutex ContextLock;
    unique_threadpool_work Worker;
    bool ShouldRun;
    bool IsRunning;
    bool IsComplete;
    std::condition_variable CompletionSignal;
};

_Requires_lock_held_(workerContext->ContextLock)
void
ResumeWorker(
    _In_ WORKER_CONTEXT* workerContext
    )
{
    workerContext->ShouldRun = true;
    if (!workerContext->IsRunning && !workerContext->IsComplete) {

        // No active workers, so start a new one.

        workerContext->IsRunning = true;
        SubmitThreadpoolWork(workerContext->Worker.get());
    }
}

void
DeferredWorkEventCallback(
    _In_ ACTIVITY_COORDINATOR_NOTIFICATION notificationType,
    _In_ void* callbackContext
    )
{
    WORKER_CONTEXT* workerContext = reinterpret_cast<WORKER_CONTEXT*>(callbackContext);

    // Use this callback thread to dispatch notifications to a worker thread
    // about whether or not it should process the next chunk of deferred work.

    // Note: Do not use this thread to perform your activity's workload.

    lock_guard<mutex> scopedLock(workerContext->ContextLock);
    switch (notificationType) {
    case ACTIVITY_COORDINATOR_NOTIFICATION_RUN:

        // Allow deferred work to be processed.

        ResumeWorker(workerContext);

        break;

    case ACTIVITY_COORDINATOR_NOTIFICATION_STOP:

        // Stop processing deferred work.

        workerContext->ShouldRun = false;

        break;

    default:
        FAIL_FAST();
        break;
    }
}

bool
TrainNextModelChunk(
    )
{
    //
    // Returns true if all work is completed, or false if there is more work.
    //

    return false;
}

void
DeferredModelTrainingWorker(
    _Inout_ PTP_CALLBACK_INSTANCE callbackInstance,
    _Inout_opt_ PVOID callbackContext,
    _Inout_ PTP_WORK work
    )
{
    // Threadpool callback instance and work are not needed for this sample.

    UNREFERENCED_PARAMETER(callbackInstance);
    UNREFERENCED_PARAMETER(work);

    WORKER_CONTEXT* workerContext = reinterpret_cast<WORKER_CONTEXT*>(callbackContext);
    bool workComplete = false;

    // Keep processing work until being told to stop or all work has been completed.

    while (true) {
        {
            lock_guard<mutex> scopedLock(workerContext->ContextLock);

            if (workComplete) {
                workerContext->IsComplete = true;
            }

            if (!workerContext->ShouldRun || workerContext->IsComplete) {
                workerContext->IsRunning = false;
                break;
            }
        }

        // TrainNextModelChunk returns true when there is no more work to do.

        workComplete = TrainNextModelChunk();
    }

    workerContext->CompletionSignal.notify_all();
}

int
__cdecl
wmain(
    )
{
    WORKER_CONTEXT workerContext;
    workerContext.ShouldRun = false;
    workerContext.IsRunning = false;
    workerContext.IsComplete = false;

    // Create the worker that will be started by our subscription callback.

    workerContext.Worker.reset(CreateThreadpoolWork(
        DeferredModelTrainingWorker,
        &workerContext,
        nullptr));
    RETURN_LAST_ERROR_IF_NULL(workerContext.Worker);

    // Allocate a policy suited for tasks that are best run when unlikely
    // to cause impact to the user or system performance.

    unique_policy policy;
    RETURN_IF_FAILED(CreateActivityCoordinatorPolicy(
        ACTIVITY_COORDINATOR_POLICY_TEMPLATE_GOOD,
        &policy));

    // The model training in this sample consumes GPU. As per the MSDN docs, the
    // GOOD policy template doesn't currently include the GPU resource. We
    // therefore customize the policy to include good GPU conditions to minimize
    // the impact of running our work.

    RETURN_IF_FAILED(SetActivityCoordinatorPolicyResourceCondition(
        policy.get(),
        ACTIVITY_COORDINATOR_RESOURCE_GPU,
        ACTIVITY_COORDINATOR_CONDITION_GOOD));

    // Subscribe to the policy for coordination notifications.

    unique_subscription subscription;
    RETURN_IF_FAILED(SubscribeActivityCoordinatorPolicy(
        policy.get(),
        DeferredWorkEventCallback,
        &workerContext,
        &subscription));;

    // Destroy the policy because we no longer need it.

    policy.reset();

    // We want our task to complete within 48h, so we allocate 24h under our
    // ideal policy and before falling back to a downgraded policy.

    bool workerCompleted;

    {
        unique_lock<mutex> scopedLock(workerContext.ContextLock);
        workerCompleted = workerContext.CompletionSignal.wait_for(
            scopedLock,
            hours(24),
            [&workerContext] { return workerContext.IsComplete; });
    }

    if (workerCompleted) {

        // Since our work is complete, we should clean up our subscription by
        // unsubscribing. This would normally be handled quietly by our RAII
        // types, but we release them explicitly to demonstrate API flow for
        // developers manually managing resources.

        subscription.reset();
        return S_OK;
    }

    // We passed our soft deadline, so downgrade the policy and wait the
    // remaining 24h until our hard deadline has been reached. Since
    // Subscriptions and policies are independent of each other, we need to
    // create a new subscription with our downgraded policy to receive
    // notifications based on its configuration.
    // 
    // The downgraded policy uses medium conditions for all needed resources.
    // This gives us the best chance to run while helping to prevent us from
    // critically degrading the user experience, which we are more likely to do
    // when falling back to manual execution.

    RETURN_IF_FAILED(CreateActivityCoordinatorPolicy(
        ACTIVITY_COORDINATOR_POLICY_TEMPLATE_MEDIUM,
        &policy));

    RETURN_IF_FAILED(SetActivityCoordinatorPolicyResourceCondition(
        policy.get(),
        ACTIVITY_COORDINATOR_RESOURCE_GPU,
        ACTIVITY_COORDINATOR_CONDITION_MEDIUM));

    subscription.reset();
    RETURN_IF_FAILED(SubscribeActivityCoordinatorPolicy(
        policy.get(),
        DeferredWorkEventCallback,
        &workerContext,
        &subscription));

    {
        unique_lock<mutex> scopedLock(workerContext.ContextLock);
        workerCompleted = workerContext.CompletionSignal.wait_for(
            scopedLock,
            hours(24),
            [&workerContext] { return workerContext.IsComplete; });
    }

    // We passed our deadline, so unsubscribe and manually resume our task.

    subscription.reset();
    ResumeWorker(&workerContext);

    // We destroyed our subscription, so we wait indefinitely for completion as
    // there's nothing to pause execution of our task.

    unique_lock<mutex> scopedLock(workerContext.ContextLock);
    workerContext.CompletionSignal.wait(
        scopedLock,
        [&workerContext] { return workerContext.IsComplete; });

    return S_OK;
}

Panoramica dell'API Activity Coordinator

API e terminologia di Activity Coordinator

Scelta dei criteri di Activity Coordinator corretti