Boost池分配器比新的慢

Question

所以我memory_pools基于boost池创建了这个容器分配器类:

memory_pools.hpp

#ifndef MEMORY_POOL_HPP
# define MEMORY_POOLS_HPP

// boost
# include <boost/pool/pool.hpp>
# include <boost/unordered_map.hpp>

template<typename ElementType>
class   memory_pools
{
public:
  template <typename>
  friend class memory_pools;

private:
  using pool = boost::pool<>;

public:
  using value_type = ElementType;
  using pointer = value_type*;
  using const_pointer = const value_type*;
  using reference = value_type&;
  using const_reference = const value_type&;
  using size_type = pool::size_type;
  using difference_type = pool::difference_type;

public:

  template<typename OtherElementType>
  struct rebind
  {
    using other = memory_pools<OtherElementType>;
  };

public:
  memory_pools();

  template<typename SourceElement>
  memory_pools(const memory_pools<SourceElement>&);

public:
  pointer   allocate(const size_type n);
  void  deallocate(const pointer ptr, const size_type n);

  template<typename... Args>
  void  construct(pointer, Args...);
  void  destroy(pointer);

public:
  bool  operator==(const memory_pools&);
  bool  operator!=(const memory_pools&);

private:
  using pools_map = boost::unordered_map<std::size_t, std::shared_ptr<pool>>;

private:
  std::shared_ptr<pools_map>      pools_map_;
  std::shared_ptr<pool>           pool_;
};

# include <memory_pools.ipp>

#endif

memory_pools.ipp

#ifndef MEMORY_POOLS_IPP
# define MEMORY_POOLS_IPP

template<typename ElementType>
memory_pools<ElementType>::memory_pools()
  :
  pools_map_(std::make_shared<pools_map>
             (pools_map
             {
               std::make_pair
                 (sizeof(ElementType),
                  make_shared<pool>(sizeof(ElementType)))
             })),
  pool_(pools_map_->at(sizeof(ElementType)))
{
}

template<typename ElementType>
template<typename SourceElement>
memory_pools<ElementType>::memory_pools
(const memory_pools<SourceElement>& rebinded_from)
  :
  pools_map_(rebinded_from.pools_map_),
  pool_(pools_map_->insert
        (std::make_pair(sizeof(ElementType),
                        make_shared<pool>(sizeof(ElementType)))).first->second)
  {
  }

template<typename ElementType>
typename memory_pools<ElementType>::pointer memory_pools<ElementType>::allocate
(const size_type n)
{
  pointer ret = static_cast<pointer>(pool_->ordered_malloc(n));

  if ((!ret) && n)
    throw std::bad_alloc();

  return (ret);
}

template<typename ElementType>
void        memory_pools<ElementType>::deallocate
(const pointer ptr, const size_type n)
{
  pool_->ordered_free(ptr, n);
}

template<typename ElementType>
template<typename... Args>
void        memory_pools<ElementType>::construct(pointer ptr, Args... args)
{
  new (ptr) ElementType(std::forward<Args>(args)...);
}

template<typename ElementType>
void        memory_pools<ElementType>::destroy(pointer ptr)
{
  ptr->~ElementType();
}

template<typename ElementType>
bool        memory_pools<ElementType>::operator==(const memory_pools& rhs)
{
  return (pools_map_ == rhs.pools_map_);
}

template<typename ElementType>
bool        memory_pools<ElementType>::operator!=(const memory_pools& rhs)
{
  return (pools_map_ != rhs.pools_map_);
}

#endif

---

然后我用它测试它:

#include <memory_pools.hpp>

int     main(void)
{
  using pools_type = memory_pools<std::pair<const int, int>>;
  pools_type    pools;

  boost::unordered_map<int, int, boost::hash<int>, std::equal_to<int>, pools_type>      map;
  //boost::unordered_map<int, int, boost::hash<int>, std::equal_to<int>>      map;

  for (unsigned int i = 0; i < 20000; ++i)
    {
      map[i] = i + 1;
    }

  return (0);
}

使用macOSX 10.10上的clang3.5,我得到了:

$ time ./a.out

real    0m1.873s
user    0m1.850s
sys     0m0.009s

---

而我发布的时候:

#include <memory_pools.hpp>

int     main(void)
{
  using pools_type = memory_pools<std::pair<const int, int>>;
  pools_type    pools;

  //boost::unordered_map<int, int, boost::hash<int>, std::equal_to<int>, pools_type>      map;
  boost::unordered_map<int, int, boost::hash<int>, std::equal_to<int>>      map;

  for (unsigned int i = 0; i < 20000; ++i)
    {
      map[i] = i + 1;
    }

  return (0);
}

我有:

$ time ./a.out

real    0m0.019s
user    0m0.016s
sys     0m0.002s

---

题

使用boost pool的内存分配应该是那么慢还是我的测试由于某种原因无效？

---

编辑

在Carmeron的评论之后,我添加了-O3和-DNDEBUG旗帜,现在我有:

$time ./a.out

real    0m0.438s
user    0m0.431s
sys     0m0.003s

对于memory_pools版本,和:

$ time ./a.out

real    0m0.008s
user    0m0.006s
sys     0m0.002s

对于标准分配器版本.

题

这个问题仍然存在,它是否正常更慢？

Answer 1

我从未使用过Boost的池代码,甚至没有使用它.但我一般都知道有关内存池的一些事情,我不希望测试中的内存池优于malloc.

要理解这一点,您必须首先了解如果您还没有实现malloc和free.这个问题的答案似乎提供了一个非常好的总结:malloc()和free()如何工作？

内存碎片是一个很难的问题malloc()和free(),而且也没有简单,快速的解决方案.但是,如果你可以保证所有的分配大小都相同,那就更容易了:这就是内存池可以获胜的方式.但是你的测试不涉及大量的内存碎片,并且可能根本没有释放大量内存.所以在这个测试中,malloc()胜利和池都输了.要优化您的测试,您可能会混合一堆删除,例如:

// Allocate 10,000 things
// Loop 10 times:
//   Allocate 1,000 things
//   Delete 1,000 things

说了这么多,如果你真的想知道为什么一段特定的代码按照它的方式执行,你应该对它进行分析.思考一段代码为什么表现出某种特定方式的理论是有用的,但你也必须测试你的理论.