MikeMirzayanov's blog

By MikeMirzayanov, 10 years ago, translation, In English

Hello, Codeforces!

In late December and early January, I wrote a proof-of-concept of separate service in C ++, to take out heavy data from the Codeforces Java-code to C++. Long time I have not written in C ++, experienced a funny sense of immersion into another world.

I was surprised to find the lack of open-addressing hashmap in the C++ standard library, and indeed in the boost, and in other popular libraries. It is strange somehow, because often open addressing should be better than separate chaining both in time and memory. And since I intend to keep small objects, then sure.

I quickly sketched a prototype that actually shows that open addressing in 2-3 times faster than the standard std::unordered_map. Here's the output of the benchmark on my laptop:

std :: map takes 15779 ms
std :: unordered_map takes 4698 ms
oaht :: hash_map takes 1473 ms

I think that there is no good implementation of such container in stl-style with the support of C++11 (move semantics). Maybe someone will show, but I have not found.

Here is my prototype on github: https://github.com/MikeMirzayanov/open_addressing_hash_table Unfortunately, I'm not cool C ++ expert, and I have no time to implement.

On the other hand, on Codeforces where were many C ++- discussions, and seems there many users who understand modern C ++. Also algorithms are water and air of Codeforces! I have hope that the universal mind Codeforces and straight hands of someone in the community will help bring to life such a container. Of course, we want maximum compatibility with std::unordered_map (API and semantics should be equivalent where it is possible). Here is a sample list of features that I want to see:

  1. One more template type parameter class Pred = std::equal_to<Key>
  2. Member function at
  3. Member function clear
  4. Member function empty
  5. Member function insert
  6. Member function equal_range
  7. Member function swap
  8. Member function operator=
  9. Member functions load_factor, max_load_factor, rehash, reserve
  10. Constant members
  11. Member types: difference_type, hasher, key_type, mapped_type, reference, size_type, value_type, pointer, const_reference, key_equal
  12. Iterators: member types iterator, const_iterator, member functions begin, end, cbegin, cend
  13. Member function emplace
  14. Member function emplace_hint
  15. Relational operators
  16. Non-member function overload: swap
  17. Move semantics

In addition, I want to keep it working on popular C ++ 11 compilers: g ++, Visual Studio 13+, clang. Please try to follow the code style of oaht.h.

I'll be happy to delegate the management of this project to someone who has the enthusiasm, knowledge of C ++ and have proven experience in the problems. I'm looking for volunteer! Codeforces t-shirt and +5 to your honor guaranteed.

P.S. As a result we want to have a good implementation, so there is a desire to put into the project only correct and well-written code.

  • Vote: I like it
  • +140
  • Vote: I do not like it

»
10 years ago, # |
  Vote: I like it +64 Vote: I do not like it

Will we be able to use oaht::hash_map h; after #include "codeforces.h" on codeforces in the future? It sounds very interesting :)

»
10 years ago, # |
Rev. 3   Vote: I like it +9 Vote: I do not like it

Have you tried SGI hash_map?

Unfortunately I couldn't find the collision resolution method used in it, but I've tried your benchmark.cpp on my laptop and it shows a better performance (than the std::unordered_map). So you might want to check it out.

Header file: <ext/hash_map>

Namespace: __gnu_cxx

Edit1: It's not SGI's hash_map, I misinterpreted this comment in <ext/hash_map>

/** @file backward/hash_map
 *  This file is a GNU extension to the Standard C++ Library (possibly
 *  containing extensions from the HP/SGI STL subset).
 */

Edit2: After some digging into the header files, I found that it uses closed addressing method for collision resolution.

You can find that in the header file <backward/hashtable.h> function name ... insert_equal_noresize(...).

  • »
    »
    10 years ago, # ^ |
      Vote: I like it +8 Vote: I do not like it

    It seems it has worst performance than open addressing hash map proof-of-concept:

    n oaht::hash_map __gnu_cxx::hash_map
    500000 204 ms 374 ms
    1000000 408 ms 841 ms
    2000000 860 ms 1760 ms
    4000000 1731 ms 3787 ms
    8000000 3956 ms 8250 ms
»
10 years ago, # |
  Vote: I like it +8 Vote: I do not like it

Have you tried Google Dense Hashmap? Though it might not 100% matching your specs, its a lot faster then stl containers and really nicely written.

https://code.google.com/p/sparsehash/

  • »
    »
    10 years ago, # ^ |
      Vote: I like it +3 Vote: I do not like it

    Looks promising, thanks. But I've tried test<GOOGLE_NAMESPACE::sparse_hash_map<int64_t,int64_t>>("GOOGLE_NAMESPACE::sparse_hash_map"); and it works ~2000 ms while open addressing hash map proof-of-concept ~400 ms. Right now I can't use it MinGW (can't compile), using in MS VS 2013.

    • »
      »
      »
      10 years ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      I think dense_hash_map is more time efficient while sparse_hash_map is more space efficient. Can you try dense_hash_map?

    • »
      »
      »
      10 years ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      I checked and default dense_hash_map is about 2x slower on my machine.

      I wonder how much of this difference is due to the fact that your proof of concept supports few operations, namely only inserts but not deletes.

      • »
        »
        »
        »
        10 years ago, # ^ |
          Vote: I like it 0 Vote: I do not like it

        Did you use benchmark code from my github repository? So dense_hash_map wasn't tested on erases as well. I do not see that support of more functionallity in oaht will reduce performance.

        Also in my case I will not use erases often.

        How to use dense_hash_map? I've got Assertion failed: settings.use_empty(), file google\sparsehash/internal/densehashtable.h, line 476.

        • »
          »
          »
          »
          »
          10 years ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          Yes, I cloned your github code.

          Also had that error with dense_hash_map, but it is explained in the readme:

             dense_hash_map requires you to set aside one key value as the
             'empty bucket' value, set via the set_empty_key() method.  This
             *MUST* be called before you can use the dense_hash_map.  It is
             illegal to insert any elements into a dense_hash_map whose key is
             equal to the empty-key.
          

          So, I created a new test with

          // Init
          google::dense_hash_map<int64_t, int64_t> m;
          m.set_empty_key(1LL << 62);
          // do operations
          
        • »
          »
          »
          »
          »
          10 years ago, # ^ |
            Vote: I like it 0 Vote: I do not like it

          Sorry for the multiple replies. Just noticed that dense_hash_map also uses open addressing with quadratic probing for collisions resolution.

          So, I think these efficiency differences are more about the particular data set or hash function used, rather than the hash map implementation.

          • »
            »
            »
            »
            »
            »
            10 years ago, # ^ |
              Vote: I like it 0 Vote: I do not like it

            I believe that +1 strategy could be better than quadratic probing or double hashing because of better cpu cache usage.

            • »
              »
              »
              »
              »
              »
              »
              10 years ago, # ^ |
                Vote: I like it +5 Vote: I do not like it

              I think it depends on the data set. Linear probing has the risk of clumping, having too many adjacent buckets occupied.

              If you feel this is unlikely to occur in the data you'll use, then it's probably better.

              My feeling is that libraries like dense_hash_map are designed to minimize the probability of an adversarial input, which may not be the best solution for your case.

              • »
                »
                »
                »
                »
                »
                »
                »
                10 years ago, # ^ |
                  Vote: I like it 0 Vote: I do not like it

                It depends on load factor. Linear probing has worse function cluster_length(load_factor). Because of hashing it doesn't depend much on input (except anti-hash cases).

                Currently, I'm looking into Walrus's implementation on https://code.google.com/p/morzhovets/source/browse It looks as a good trade off between time and memory. Much better than std::unordered_map.

          • »
            »
            »
            »
            »
            »
            10 years ago, # ^ |
              Vote: I like it 0 Vote: I do not like it

            I remember a long time ago I tried to implement my own oaht and it was also considerably faster than dense_hash_map for all benchmarks. I was using double hashing to resolve collisions.

            For some reason, in the case of hash maps, it seems the more generic the code is, the less-performant it becomes. Maybe it has something to do with working well against most datasets, or is really just a load factor issue, I don't know.

»
10 years ago, # |
  Vote: I like it -26 Vote: I do not like it

How to use it on Codeforces? What compiler should be chosen and #include what file(s) I got lots of CE……