std::unordered_map - how to “track” max/min key at any time

后端 未结 2 391
栀梦
栀梦 2021-01-16 13:36

I have std::unordered_map. I do not want to use other structures like tree or anything else cause latency requirements. But at any time I need t

相关标签:
2条回答
  • 2021-01-16 13:57

    Your exact requirements cannot be met: std::unordered_map is fast (i.e. O(1) insertion/erasure/lookup) because it doesn't order its elements. This means that finding the minimum/maximum takes O(N).

    The price of ordering paid by std::map that insertion/erasure/lookup (including finding the minimum/maximum) all become O(log N).

    If you don't mind using Boost, you could take a look at the Boost.MultiIndex library. This allows you to access your data by multiple interfaces simultaneously. It is an extremely versatile and high-performance library that can also be used for MRU cache data structures (which also combine fast access and tracking of an ordering).

    Your primary std::unordered_map interface is provided by the hashed_unique index, where you use an identity function object (provided by Boost.MultiIndex). Your secondary interface would mimic a std::set. This allows finding the minimum/maximum value in O(1) time as *begin() and *rbegin().

    You can do that by adding a ordered_unique index with a user-defined MyOrder function object to compute whatever criterion you like to use to order them. Small code example (omitting the boost:: namespace everywhere)

    using MyContainer = multi_index_container<
        Item,
        indexed_by<
            hashed_unique<identity<Item>>, // O(1) erasure/lookup, O(log N) insertion
            ordered_unique<MyOrder<Item>>  // track minimum/maximum in O(log N)
        >
    >;
    

    Behind the scenes, Boost.MultiIndex implements this roughly as a std::set<Value> with a std::unordered_map<Key, set<Value>::iterator>. This means both lookups and erasures are O(1). The erasure can be O(1) because the unordered_map::erase(value) returns an iterator into the set and std::set<Value>::erase(iterator) is O(1).

    However, insertions remain O(log N), because you cannot avoid finding the rank of a newly inserted value within an ordered sequence in less time.

    This improvement compared to std::map for lookup/erasure costs only a few pointers per element space overhead.

    0 讨论(0)
  • 2021-01-16 14:10

    You can write a wrapper. With that, you get every insertion/deletion and can retain min / max values.

    #pragma once
    
    #include <unordered_map>
    
    using std::unordered_map;
    
    class my_unordered_map
    {
    public:
        void insert(int key, int value)
        {
            _map[key] = value;
            if (_max < value)
                _max = value;
            if (_min> value)
                _min = value;
        }
        int get(int key)
        {
            auto it = _map.find(key);
            if (it != _map.end())
                return it->second;
            return 0; // or some error handling here.
        }
        int getMin()
        {
            return _min;
        }
        int getMax()
        {
            return _max;
        }
        // more functionality here.
    private:
        unordered_map<int, int> _map;
        int _min = 0;
        int _max = 0;
    };
    
    0 讨论(0)
提交回复
热议问题