具有Boost::Fiber的多个共享工作池

Question

我一直在研究boost::fibers作为处理数据处理和IO问题的一种方法。特别是shared_work调度器看起来很有希望，因为它允许我为每个数据处理源划分一个数据处理任务，然后让它们根据需要在几个线程之间相互分发。

然而，这就引出了我问题的根源:看起来每个进程只能有一个shared_work“池”。如果我希望在4个线程之间共享处理数据中的一组12条光纤，而同时，另一组12条纤维正在写入处理后的数据以在其他4个线程之间共享，那么我该怎么办？

类似于：

#include<string>
#include<iostream>
#include<vector>
#include<mutex>
#include<thread>
#include<random>
#include<map>
#include<sstream>
#include<boost/bind.hpp>
#include<boost/fiber/all.hpp>

typedef boost::fibers::fiber FiberType;
typedef std::unique_lock<boost::fibers::mutex> LockType;


static const int fiberIterationCount = 5000;
static const int fiberCount          = 12;
static const int threadCount         = 4;
static const int distLowerLimit      = 50;
static const int distUpperLimit      = 500;

static boost::fibers::mutex firstMutex{};
static boost::fibers::mutex secondMutex{};
static boost::fibers::condition_variable firstCondition{};
static boost::fibers::condition_variable secondCondition{};
static boost::fibers::barrier synchronize{2*threadCount};
static int typeOneFibersFinished{0};
static int typeTwoFibersFinished{0};

static std::mt19937 typeOneGenerators[fiberCount];
static std::mt19937 typeTwoGenerators[fiberCount];

static std::mutex typeMapMutex;//lock for writing unnecessary for reads
static std::map<std::thread::id, std::string> threadTypeMap;


//simple function to give a heavy cpu load of variable duration
unsigned long long findPrimeNumber(int n)
{
    int count=0;
    unsigned long long a = 2;
    while(count<n)
    {
        bool isPrime = true;
        for(unsigned long long b = 2; (b * b) <= a; ++b)
        {
            if((a % b) == 0)
            {
                isPrime = false;
                break;
            }
        }
        if(isPrime)
        {
            count++;
        }
        ++a;
    }
    return (a - 1);
}

void fiberTypeOne(int fiberNumber)
{
    std::cout<<"Starting Type One Fiber #"<<fiberNumber;
    std::uniform_int_distribution<int> dist(distLowerLimit, distUpperLimit);
    for(int i=0; i<fiberIterationCount; ++i)
    {
        //generate a randomish load on this fiber so that it does not take a regular time slice
        int tempPrime = dist(typeOneGenerators[fiberNumber]);
        unsigned long long temp = findPrimeNumber(tempPrime);
        std::cout << "T1 fiber #"<<fiberNumber<<" running on "<<threadTypeMap[std::this_thread::get_id()]
                  <<"\n    Generated: "<<tempPrime<<", "<<temp;
        boost::this_fiber::yield();
    }

    {
        LockType lock(firstMutex);
        ++typeOneFibersFinished;
    }
    firstCondition.notify_all();
}

void threadTypeOne(int threadNumber)
{
    //make a shared work scheduler that associates its fibers with "fiber pool 0"
    boost::fibers::use_scheduling_algorithm< multi_pool_scheduler<0> >();
    std::cout<<"Starting Type One Thread #"<<threadNumber<<" With Thread ID: "<<std::this_thread::get_id();

    {
        std::unique_lock<std::mutex> lock{typeMapMutex};
        std::ostringstream gen;
        gen<<"Thread Type 1 - Number: "<<threadNumber<<" with id: "<<std::this_thread::get_id();
        threadTypeMap[std::this_thread::get_id()] = gen.str();
    }
    if(threadNumber == 0)
    { //if we are thread zero, create the fibers then join them to take ourselves off the "fiber list"
        std::cout<<"Spawning Type One Fibers";
        for(int fiberNumber=0; fiberNumber<fiberCount; ++fiberNumber)
        {//create the fibers and instantly detach them
            FiberType(boost::bind(&fiberTypeOne, fiberNumber)).detach();
        }
    }
    synchronize.wait();
    std::cout<<"T1 Thread preparing to wait";
    //now let the fibers do their thing
    LockType lock(firstMutex);
    firstCondition.wait(lock, [](){return (typeOneFibersFinished == fiberCount);});
}

void fiberTypeTwo(int fiberNumber)
{
    std::cout<<"Starting Type Two Fiber #"<<fiberNumber;
    std::uniform_int_distribution<int> dist(distLowerLimit, distUpperLimit);
    for(int i=0; i<fiberIterationCount; ++i)
    {
        //generate a randomish load on this fiber so that it does not take a regular time slice
        int tempPrime = dist(typeTwoGenerators[fiberNumber]);
        unsigned long long temp = findPrimeNumber(tempPrime);
        std::cout << "T2 fiber #"<<fiberNumber<<" running on "<<threadTypeMap[std::this_thread::get_id()]
                  <<"\n    Generated: "<<tempPrime<<", "<<temp;
        boost::this_fiber::yield();
    }

    {
        LockType lock(secondMutex);
        ++typeTwoFibersFinished;
    }
    secondCondition.notify_all();
}

void threadTypeTwo(int threadNumber)
{
    //make a shared work scheduler that associates its fibers with "fiber pool 1"
    boost::fibers::use_scheduling_algorithm< multi_pool_scheduler<1> >();
    std::cout<<"Starting Type Two Thread #"<<threadNumber<<" With Thread ID: "<<std::this_thread::get_id();
    {
        std::unique_lock<std::mutex> lock{typeMapMutex};
        std::ostringstream gen;
        gen<<"Thread Type 2 - Number: "<<threadNumber<<" with id: "<<std::this_thread::get_id();
        threadTypeMap[std::this_thread::get_id()] = gen.str();
    }
    if(threadNumber == 0)
    { //if we are thread zero, create the fibers then join them to take ourselves off the "fiber list"
        std::cout<<"Spawning Type Two Fibers";
        for(int fiberNumber=0; fiberNumber<fiberCount; ++fiberNumber)
        {//create the fibers and instantly detach them
            FiberType(boost::bind(&fiberTypeTwo, fiberNumber)).detach();
        }
    }
    synchronize.wait();
    std::cout<<"T2 Thread preparing to wait";
    //now let the fibers do their thing
    LockType lock(secondMutex);
    secondCondition.wait(lock, [](){return (typeTwoFibersFinished == fiberCount);});
}

int main(int argc, char* argv[])
{
    std::cout<<"Initializing Random Number Generators";
    for(unsigned i=0; i<fiberCount; ++i)
    {
        typeOneGenerators->seed(i*500U - 1U);
        typeTwoGenerators->seed(i*1500U - 1U);
    }

    std::cout<<"Commencing Main Thread Startup Startup";
    std::vector<std::thread> typeOneThreads;
    std::vector<std::thread> typeTwoThreads;
    for(int i=0; i<threadCount; ++i)
    {
        typeOneThreads.emplace_back(std::thread(boost::bind(&threadTypeOne, i)));
        typeTwoThreads.emplace_back(std::thread(boost::bind(&threadTypeTwo, i)));
    }
    //now let the threads do their thing and wait for them to finish with join
    for(unsigned i=0; i<threadCount; ++i)
    {
        typeOneThreads[i].join();
    }
    for(unsigned i=0; i<threadCount; ++i)
    {
        typeTwoThreads[i].join();
    }
    std::cout<<"Shutting Down";
    return 0;
}

如果不编写自己的光纤调度程序，这是可能的吗？如果是这样的话，是怎么做的？

Answer 1

我决定我确实需要编写我自己的调度程序。然而，实际工作量很小。boost::fibers::shared_work调度程序使用一个由静态互斥保护的静态队列管理线程之间共享的纤维列表。还有另一个队列控制每个线程的主光纤(因为每个线程都有自己的调度程序)，但它是类实例的本地队列，而不是以静态成员的方式在类的所有实例之间共享的。

因此，为了防止静态队列和锁在不同的线程集合之间共享，解决方案是在类的前面放置一个模板参数，这个参数通常是无用的。然后，每个线程向该模板传递一个不同的参数。以这种方式，由于您为模板的每个专门化获得了不同的对象，因此对于每个具有不同池号的实例化，您得到了不同的静态变量集。

下面是这个解决方案的实现(主要是一个boost::fiber::shared_work的副本，其中有几个变量和类型更清楚地命名，并添加了模板参数)。

#include <condition_variable>
#include <chrono>
#include <deque>
#include <mutex>
#include <boost/config.hpp>
#include <boost/fiber/algo/algorithm.hpp>
#include <boost/fiber/context.hpp>
#include <boost/fiber/detail/config.hpp>
#include <boost/fiber/scheduler.hpp>
#include <boost/assert.hpp>
#include "boost/fiber/type.hpp"

#ifdef BOOST_HAS_ABI_HEADERS
#  include BOOST_ABI_PREFIX
#endif

#ifdef _MSC_VER
# pragma warning(push)
# pragma warning(disable:4251)
#endif

/*!
* @class SharedWorkPool
* @brief A scheduler for boost::fibers that operates in a manner similar to the
* shared work scheduler, except that it takes a template parameter determining
* which pool to draw fibers from. In this fashion, one group of threads can share
* a pool of fibers among themselves while another group of threads can work with
* a completely separate pool
* @tparam PoolNumber The index of the pool number for this thread
*/
template <int PoolNumber>
class SharedWorkPool : public boost::fibers::algo::algorithm
{
    typedef std::deque<boost::fibers::context * >      ReadyQueueType;
    typedef boost::fibers::scheduler::ready_queue_type LocalQueueType;
    typedef std::unique_lock<std::mutex>               LockType;

public:
    SharedWorkPool() = default;
    ~SharedWorkPool() override {}

    SharedWorkPool( bool suspend) : suspendable{suspend}{}

    SharedWorkPool( SharedWorkPool const&) = delete;
    SharedWorkPool( SharedWorkPool &&) = delete;

    SharedWorkPool& operator=(const SharedWorkPool&) = delete;
    SharedWorkPool& operator=(SharedWorkPool&&) = delete;

    void awakened(boost::fibers::context* ctx) noexcept override;

    boost::fibers::context* pick_next() noexcept override;

    bool has_ready_fibers() const noexcept override
    {
        LockType lock{readyQueueMutex};
        return ((!readyQueue.empty()) || (!localQueue.empty()));
    }

    void suspend_until(const std::chrono::steady_clock::time_point& timePoint) noexcept override;

    void notify() noexcept override;

private:
    static ReadyQueueType readyQueue;
    static std::mutex     readyQueueMutex;

    LocalQueueType          localQueue{};
    std::mutex              instanceMutex{};
    std::condition_variable suspendCondition{};
    bool                    waitNotifyFlag{false};
    bool                    suspendable{false};

};

template <int PoolNumber>
void SharedWorkPool<PoolNumber>::awakened(boost::fibers::context* ctx) noexcept
{
    if(ctx->is_context(boost::fibers::type::pinned_context))
    { // we have been passed the thread's main fiber, never put those in the shared queue
        localQueue.push_back(*ctx);
    }
    else
    {//worker fiber, enqueue on shared queue
        ctx->detach();
        LockType lock{readyQueueMutex};
        readyQueue.push_back(ctx);
    }
}


template <int PoolNumber>
boost::fibers::context* SharedWorkPool<PoolNumber>::pick_next() noexcept
{
    boost::fibers::context * ctx = nullptr;
    LockType lock{readyQueueMutex};
    if(!readyQueue.empty())
    { //pop an item from the ready queue
        ctx = readyQueue.front();
        readyQueue.pop_front();
        lock.unlock();
        BOOST_ASSERT( ctx != nullptr);
        boost::fibers::context::active()->attach( ctx); //attach context to current scheduler via the active fiber of this thread
    }
    else
    {
        lock.unlock();
        if(!localQueue.empty())
        { //nothing in the ready queue, return main or dispatcher fiber
            ctx = & localQueue.front();
            localQueue.pop_front();
        }
    }
    return ctx;
}

template <int PoolNumber>
void SharedWorkPool<PoolNumber>::suspend_until(const std::chrono::steady_clock::time_point& timePoint) noexcept
{
    if(suspendable)
    {
        if (std::chrono::steady_clock::time_point::max() == timePoint)
        {
            LockType lock{instanceMutex};
            suspendCondition.wait(lock, [this](){return waitNotifyFlag;});
            waitNotifyFlag = false;
        }
        else
        {
            LockType lock{instanceMutex};
            suspendCondition.wait_until(lock, timePoint, [this](){return waitNotifyFlag;});
            waitNotifyFlag = false;
        }
    }
}

template <int PoolNumber>
void SharedWorkPool<PoolNumber>::notify() noexcept
{
    if(suspendable)
    {
        LockType lock{instanceMutex};
        waitNotifyFlag = true;
        lock.unlock();
        suspendCondition.notify_all();
    }
}

template <int PoolNumber>
std::deque<boost::fibers::context*> SharedWorkPool<PoolNumber>::readyQueue{};

template <int PoolNumber>
std::mutex SharedWorkPool<PoolNumber>::readyQueueMutex{};

注意，如果尝试在不同的编译单元中使用声明中的相同池号，我不完全确定会发生什么。但是，在正常情况下，即您只在单个位置为每个boost::fibers::use_scheduling_algorithm< Threads::Fibers::SharedWorkPool<WorkPoolNumber> >();编写了WorkPoolNumber，它工作得非常完美。分配给给定一组线程的光纤总是在同一组线程中运行，而不是由不同的一组线程运行。