为什么迭代器调试在调试版本中会降低std :: unordered_map 200x的速度？

Question

I understand that code will be slower, but why so much? How do I code to avoid this slowdown?

std::unordered_map uses other containers internally and those containers use iterators. When built debug, _ITERATOR_DEBUG_LEVEL=2 by default. This turns on iterator debugging . Sometimes my code is not affected much, and sometimes it runs extremely slowly.

I can speed my example up by setting _ITERATOR_DEBUG_LEVEL=0 in my project properties >> C++ >> Preprocessor >> Preprocessor definitions. But as this link suggests, I cannot do so in my real project. In my case, I get conflicts with MSVCMRTD.lib, which contains std::basic_string built with _ITERATOR_DEBUG_LEVEL=2. I understand I can work around the problem by statically linking to the CRT. But I would prefer not to if I can fix the code so the problem does not arise.

I can make changes that improve the situation. But I am just trying things out without understanding why they work. For example, as is, the first 1000 inserts work at full speed. But if I change O_BYTE_SIZE to 1, the first inserts are as slow as everything else. This looks like a small change (not necessarily a good change.)

This , this , and this also shed some light, but don't answer my question.

I am using Visual Studio 2010 (This is legacy code.) I created a Win32 console app and added this code.

Main.cpp

#include "stdafx.h"


#include "OString.h"
#include "OTHashMap.h"

#include <cstdio>
#include <ctime>
#include <iostream>

// Hash and equal operators for map
class CRhashKey {
public:
   inline unsigned long operator() (const OString* a) const { return a->hash(); }
};

class CReqKey {
public:
    inline bool operator() (const OString& x, const OString& y) const { return strcmp(x.data(),y.data()) != 0; }
    inline bool operator() (const OString* x, const OString& y) const { return operator()(*x,y); }
    inline bool operator() (const OString& x, const OString* y) const { return operator()(x,*y); }
    inline bool operator() (const OString* x, const OString* y) const { return operator()(*x,*y); }
};


int _tmain(int argc, _TCHAR* argv[])
{
    const int CR_SIZE = 1020007;

    CRhashKey h;
    OTPtrHashMap2<OString, int, CRhashKey, CReqKey> *code_map = 
        new OTPtrHashMap2 <OString, int, CRhashKey, CReqKey>(h, CR_SIZE);

    const clock_t begin_time = clock();

    for (int i=1; i<=1000000; ++i)
    {
        char key[10];
        sprintf(key, "%d", i);

        code_map->insert(new OString(key), new int(i));

        //// Check hash values
        //OString key2(key);
        //std::cout << i << "\t" << key2.hash() << std::endl;

        // Check timing
        if ((i % 100) == 0)
        {
            std::cout << i << "\t" << float(clock() - begin_time) / CLOCKS_PER_SEC << std::endl;
        }
    }

    std::cout << "Press enter to exit" << std::endl;
    char buf[256];
    std::cin.getline(buf, 256);

    return 0;
}

OTHashMap.h

#pragma once

#include <fstream>
#include <unordered_map>    

template <class K, class T, class H, class EQ>
class OTPtrHashMap2
{
    typedef typename std::unordered_map<K*,T*,H,EQ>                     OTPTRHASHMAP_INTERNAL_CONTAINER;
    typedef typename OTPTRHASHMAP_INTERNAL_CONTAINER::iterator          OTPTRHASHMAP_INTERNAL_ITERATOR;

public:
    OTPtrHashMap2(const H& h, size_t defaultCapacity) : _hashMap(defaultCapacity, h) {}

    bool insert(K* key, T* val)
    {
        std::pair<OTPTRHASHMAP_INTERNAL_ITERATOR,T> retVal = _hashMap.insert(std::make_pair<K*,T*>(key, val));
        return retVal.second != NULL;
    }

    OTPTRHASHMAP_INTERNAL_CONTAINER _hashMap;

private:
};

OString.h

#pragma once

#include <string>

class OString
{
public:
    OString(const std::string& s) : _string (s) { } 
    ~OString(void) {}

    static unsigned hash(const OString& s) { return unsigned (s.hash()); }
    unsigned long hash() const
    {
        unsigned hv = static_cast<unsigned>(length());
        size_t i = length() * sizeof(char) / sizeof(unsigned);
        const char * p = data();
        while (i--) {
            unsigned tmp;
            memcpy(&tmp, p, sizeof(unsigned));
            hashmash(hv, tmp);
            p = p + sizeof(unsigned);
        } 
        if ((i = length() * sizeof(char) % sizeof(unsigned)) != 0)  {
            unsigned h = 0;
            const char* c = reinterpret_cast<const char*>(p);
            while (i--)
            {
                h = ((h << O_BYTE_SIZE*sizeof(char)) | *c++);
            }
            hashmash(hv, h);
        }
        return hv; 
    }

    const char* data() const { return _string.c_str(); }
    size_t length() const    { return _string.length(); }


private:
    std::string _string;

    //static const unsigned O_BYTE_SIZE = 1;
    static const unsigned O_BYTE_SIZE = 8;
    static const unsigned O_CHASH_SHIFT = 5;

    inline void hashmash(unsigned& hash, unsigned chars) const
    {
        hash = (chars ^
                ((hash << O_CHASH_SHIFT) |
                 (hash >> (O_BYTE_SIZE*sizeof(unsigned) - O_CHASH_SHIFT))));
    }
};

Answer 1

I found enough of an answer. Collisions are the source of slowing.

Edit 2 : -- Another fix is to add this around the #include in main.cpp --

// Iterator debug checking makes the Microsoft implementation of std containers 
// *very* slow in debug builds for large containers. It must only be undefed around 
// STL includes. Otherwise we get linker errors from the debug C runtime library, 
// which was built with _ITERATOR_DEBUG_LEVEL set to 2. 
#ifdef _DEBUG
#undef _ITERATOR_DEBUG_LEVEL
#endif

#include <unordered_map>

#ifdef _DEBUG
#define _ITERATOR_DEBUG_LEVEL 2
#endif

Edit : -- The fix is switch to boost::unordered_map. --

std::unordered_map is defined in < unordered_map >. It inherits from _Hash, defined in < xhash >.

_Hash contains this (highly abbreviated)

template<...> 
class _Hash
{
    typedef list<typename _Traits::value_type, ...> _Mylist;
    typedef vector<iterator, ... > _Myvec;

    _Mylist _List;  // list of elements, must initialize before _Vec
    _Myvec _Vec;    // vector of list iterators, begin() then end()-1
};

All values are stored in _List.

_Vec is a vector of iterators into _List. It divides _List into buckets. _Vec has an iterator to the beginning and end of each bucket. Thus, if the map has 1M buckets (distinct key hashes), _Vec has 2M iterators.

When a key/value pair is inserted into the map, usually a new bucket is created. The value is pushed onto the beginning of the list. The hash of the key is the location in _Vec where two new iterators are put. This is quick because they point to the beginning of the list.

If a bucket already exists, the new value must be inserted next to the existing value in _List. This requires inserting an item in the middle of the list. Existing iterators must be updated. Apparently this requires a lot of work when iterator debugging is enabled. The code is in < list >, but I did not step through it.

---

To get an idea of how much work, I used some nonsense hash functions that would be terrible to use, but give lots of collisions or few collisions when inserting.

Added to OString.h

static unsigned hv2;

// Never collides. Always uses the next int as the hash
unsigned long hash2() const
{
    return ++hv2;
}

// Almost never collides. Almost always gets the next int. 
// Gets the same int 1 in 200 times. 
unsigned long hash3() const
{
    ++hv2;
    unsigned long lv = (hv2*200UL)/201UL;
    return (unsigned)lv;
}

// A best practice hash
unsigned long hash4() const
{
    std::hash<std::string> hasher;
    return hasher(_string);
}

// Always collides. Everything into bucket 0. 
unsigned long hash5() const
{
    return 0;
}

Added to main.cpp

// Hash and equal operators for map
class CRhashKey {
public:
   //inline unsigned long operator() (const OString* a) const { return a->hash(); }
   //inline unsigned long operator() (const OString* a) const { return a->hash2(); }
   //inline unsigned long operator() (const OString* a) const { return a->hash3(); }
   //inline unsigned long operator() (const OString* a) const { return a->hash4(); }
   inline unsigned long operator() (const OString* a) const { return a->hash5(); }
};

unsigned OString::hv2 = 0;

The results were dramatic. No realistic hash is going to work.

• hash2 - Never collide - 1M inserts in 15.3 sec
• hash3 - Almost never - 1M inserts in 206 sec
• hash4 - Best practice - 100k inserts in 132 sec, and getting slower as collisions became more frequent. 1M inserts would take > 1 hour
• hash5 - Always collide - 1k inserts in 48 sec, or 1M inserts in ~13 hours

My choices are

• Release build, debug symbols, optimization off as Retired Ninja suggests
• Statically link to MSVCMRTD so I can turn off _ITERATOR_DEBUG_LEVEL. Also solve some other similar issues.
• Change from unordered_map to a sorted vector.
• Something else. Suggestions welcome.