gpt4 book ai didi

c++ - 如何让 boost::object_pool 线程安全?

转载 作者:塔克拉玛干 更新时间:2023-11-03 02:18:04 24 4
gpt4 key购买 nike

在询问 Is there a faster heap allocation/deallocation mechanism available than boost::object_pool? 后,我得到反馈说这个对象池不是线程安全的。

所以我写了一个 ObjectFactory 包装 boost::object_pool 并添加互斥锁:

#include <memory>
using std::shared_ptr;

#include <boost/pool/object_pool.hpp>
#include <boost/thread/mutex.hpp>

template <typename T>
class ObjectFactory
{
private:
struct SharedDeleter
{
ObjectFactory<T>* m_pFact;

SharedDeleter(ObjectFactory<T>* fact) : m_pFact(fact) {}

inline void operator()(T* p) const
{
m_pFact->destroy(p);
}
};


boost::object_pool<T> m_Pool;
boost::mutex m_PoolMutex;

SharedDeleter m_Deleter;


public:
ObjectFactory() : m_Deleter(this)
{
}

template<typename TType = T, typename... TArgs>
inline TType* create(TArgs&&... mArgs)
{
boost::unique_lock<boost::mutex> scoped_lock(m_PoolMutex);
return m_Pool.construct(mArgs...);
}

inline void destroy(T* mObj)
{
boost::unique_lock<boost::mutex> scoped_lock(m_PoolMutex);
m_Pool.destroy(mObj);
}


template<typename TType = T, typename... TArgs>
inline std::shared_ptr<TType> make_shared(TArgs&&... mArgs)
{
return std::shared_ptr<TType>(this->create(mArgs...), m_Deleter);
}

};

没有互斥锁的计时结果:

 With WithObjectFactory!: 
Start time: 1381682855810868000 nanoseconds since Jan 1, 1970
End time: 1381682863375427000 nanoseconds since Jan 1, 1970
Duration: 7.56456 seconds

With WithObjectFactory and std::shared_ptr!:
Start time: 1381682863375476000 nanoseconds since Jan 1, 1970
End time: 1381682879114065000 nanoseconds since Jan 1, 1970
Duration: 15.7386 seconds

使用互斥锁的计时结果:

 With WithObjectFactory!: 
Start time: 1381683562246086000 nanoseconds since Jan 1, 1970
End time: 1381683574399319000 nanoseconds since Jan 1, 1970
Duration: 12.1532 seconds

With WithObjectFactory and std::shared_ptr!:
Start time: 1381683574399378000 nanoseconds since Jan 1, 1970
End time: 1381683595103438000 nanoseconds since Jan 1, 1970
Duration: 20.7041 seconds

您看到的是互斥锁定占用了超过 20% 的时间,恕我直言,boost::object_pool 仅可用于单线程应用程序。

我的问题:我是否使用了正确的锁定机制?destroy 是否也需要互斥锁?上面的代码中有没有我没有看到的错误?

谢谢

编辑:使用 std::map 进行测试,但速度较慢。发现 boost::thread_specific_ptr 并且它似乎有效:

#include <memory>
using std::shared_ptr;

#include <boost/pool/object_pool.hpp>
#include <boost/thread.hpp>

template <typename T>
class ObjectFactory
{
private:
struct SharedDeleter
{
ObjectFactory<T>* m_pFact;

SharedDeleter(ObjectFactory<T>* fact) : m_pFact(fact) {}

inline void operator()(T* p) const
{
m_pFact->destroy(p);
}
};

boost::thread_specific_ptr<boost::object_pool<T>> m_tpPool;
SharedDeleter m_Deleter;

public:
ObjectFactory() : m_Deleter(this)
{
if ( !m_tpPool.get() )
m_tpPool.reset(new boost::object_pool<T>());
}

template<typename TType = T, typename... TArgs>
inline TType* create(TArgs&&... mArgs)
{
return m_tpPool->construct(mArgs...);
}

inline void destroy(T* mObj)
{
m_tpPool->destroy(mObj);
}

template<typename TType = T, typename... TArgs>
inline std::shared_ptr<TType> make_shared(TArgs&&... mArgs)
{
return std::shared_ptr<TType>(this->create(mArgs...), m_Deleter);
}
};

上面代码的结果:

 With WithObjectFactory!: 
Start time: 1381714923605177000 nanoseconds since Jan 1, 1970
End time: 1381714934202228000 nanoseconds since Jan 1, 1970
Duration: 10.5971 seconds

With WithObjectFactory and std::shared_ptr!:
Start time: 1381714934202285000 nanoseconds since Jan 1, 1970
End time: 1381714950900537000 nanoseconds since Jan 1, 1970
Duration: 16.6983 seconds

最佳答案

  • 使用 std::mutex 和 boost:thread_specific_ptr 来降低速度
  • std::atomic_load(std::shared_ptr<...>*) 似乎不适用于 GCC 4.6
  • 使用 this 回退到良好的旧 pthread 调用解决方案

结果:在带有 GCC 4.6 的 Ubuntu LTS 12.04 上

 With OUT smartptrs (new and delete): 
Start time: 1381746876399819258 nanoseconds since Jan 1, 1970
End time: 1381746881851990579 nanoseconds since Jan 1, 1970
Duration: 5.45217 seconds

With smartptrs (boost::shared_ptr withOUT make_shared):
Start time: 1381746881852079492 nanoseconds since Jan 1, 1970
End time: 1381746889453586405 nanoseconds since Jan 1, 1970
Duration: 7.60151 seconds

With smartptrs (boost::shared_ptr with make_shared):
Start time: 1381746889453642790 nanoseconds since Jan 1, 1970
End time: 1381746896396534068 nanoseconds since Jan 1, 1970
Duration: 6.94289 seconds

With STD smart_ptr (std::shared_ptr with make_shared):
Start time: 1381746896396596314 nanoseconds since Jan 1, 1970
End time: 1381746902544346880 nanoseconds since Jan 1, 1970
Duration: 6.14775 seconds

With UniquePtr (boost::unique_ptr):
Start time: 1381746902544386766 nanoseconds since Jan 1, 1970
End time: 1381746907842640751 nanoseconds since Jan 1, 1970
Duration: 5.29825 seconds

With STD UniquePtr (std::unique_ptr):
Start time: 1381746907842679994 nanoseconds since Jan 1, 1970
End time: 1381746913141429138 nanoseconds since Jan 1, 1970
Duration: 5.29875 seconds

With Object Pool (boost::object_pool<>):
Start time: 1381746913141469017 nanoseconds since Jan 1, 1970
End time: 1381746917062689541 nanoseconds since Jan 1, 1970
Duration: 3.92122 seconds


With Thread Safe ObjectFactory<TestClass>...
Start time: 1381746917062729671 nanoseconds since Jan 1, 1970
End time: 1381746921388452186 nanoseconds since Jan 1, 1970
Duration: 4.32572 seconds

With Thread Safe ObjectFactory<TestClass> and std::shared_ptr...
Start time: 1381746921388491395 nanoseconds since Jan 1, 1970
End time: 1381746928808481617 nanoseconds since Jan 1, 1970
Duration: 7.41999 seconds

因此线程安全 ObjectFactory 仍然比使用普通的旧的 new 和 delete 快 1 秒以上。

大家有更好的建议欢迎补充!

感谢所有有用的反馈!

编辑:使用 __thread gnu 选项也可以快速工作:

#include <memory>
#include <boost/pool/object_pool.hpp>

template <typename T>
class ObjectFactory
{
private:
struct SharedDeleter
{
ObjectFactory<T>* m_pFact;

SharedDeleter(ObjectFactory<T>* fact) : m_pFact(fact) {}

inline void operator()(T* p) const
{
m_pFact->destroy(p);
}
};

static __thread boost::object_pool<T>* m_tlsPool;
SharedDeleter m_Deleter;

public:
ObjectFactory() : m_Deleter(this)
{
m_tlsPool = new boost::object_pool<T>();
}

virtual ~ObjectFactory()
{
delete m_tlsPool;
}

template<typename TType = T, typename... TArgs>
inline TType* create(TArgs&&... mArgs)
{
return m_tlsPool->construct(mArgs...);
}

inline void destroy(T* mObj)
{
m_tlsPool->destroy(mObj);
}

template<typename TType = T, typename... TArgs>
inline std::shared_ptr<TType> make_shared(TArgs&&... mArgs)
{
return std::shared_ptr<TType>(this->create(mArgs...), m_Deleter);
}
};

template<typename T>
/*static*/__thread boost::object_pool<T>* ObjectFactory<T>::m_tlsPool(NULL);

但是 m_tlsPool 不需要是静态的吗?

关于c++ - 如何让 boost::object_pool 线程安全?,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/19347890/

24 4 0
Copyright 2021 - 2024 cfsdn All Rights Reserved 蜀ICP备2022000587号
广告合作:1813099741@qq.com 6ren.com