Alternative for std::set without memory reallocation?

Question

In an application i generate a lot of subproblems exhaustively and solve them using "std::set" operations. For this I need to "insert" and "find" elements and also "iterate" over the sorted list.

The problem is that for each of the millions of subproblems the "std::set" implementation allocates new memory each time I insert an element in the set which makes the whole application very slow:

{   // allocate a non-value node
    _Nodeptr _Pnode = this->_Getal().allocate(1); // <- bottleneck of the program

Is there some stl-structure that allows me to to the above operations in "O(log(n))" while not reallocating any memory?

Answer 1

Using a custom allocator seems a way to reduce the amount of time spent for building and releasing a std::set<...>. Below is a complete demo of a simple allocator together with a program profiling the resulting times.

#include <algorithm>
#include <chrono>
#include <cstdlib>
#include <iostream>
#include <iterator>
#include <memory>
#include <set>
#include <vector>

// ----------------------------------------------------------------------------

template <typename T, std::size_t pool_size = 1024>
class pool_allocator
{
private:
    std::vector<T*> d_pools;
    T*              d_next;
    T*              d_end;
public:
    template <typename O>
    struct rebind {
        typedef pool_allocator<O, pool_size> other;
    };
    pool_allocator(): d_next(), d_end() {}
    ~pool_allocator() {
        std::for_each(this->d_pools.rbegin(), this->d_pools.rend(),
                      [](T* memory){ operator delete(memory); });
    }
    typedef T value_type;
    T*   allocate(std::size_t n) {
        if (std::size_t(this->d_end - this->d_next) < n) {
            if (pool_size < n) {
                // custom allocation for bigger number of objects
                this->d_pools.push_back(static_cast<T*>(operator new(sizeof(T) * n)));
                return this->d_pools.back();
            }
            this->d_pools.push_back(static_cast<T*>(operator new(sizeof(T) * pool_size)));
            this->d_next = this->d_pools.back();
            this->d_end  = this->d_next + pool_size;
        }
        T* rc(this->d_next);
        this->d_next += n;
        return rc;
    }
    void deallocate(T*, std::size_t) {
        // this could try to recycle buffers
    }
};

// ----------------------------------------------------------------------------

template <typename Allocator>
void time(char const* name, std::vector<int> const& random) {
    std::cout << "running " << name << std::flush;
    using namespace std::chrono;
    high_resolution_clock::time_point start(high_resolution_clock::now());

    std::size_t size(0);
    {
        std::set<int, std::less<int>, Allocator> values;
        for (int value: random) {
            values.insert(value);
        }
        size = values.size();
    }

    high_resolution_clock::time_point end(high_resolution_clock::now());
    std::cout << ": size=" << size << " time="
              << duration_cast<milliseconds>(end - start).count() << "ms\n";
}

// ----------------------------------------------------------------------------

int main()
{
    std::cout << "preparing..." << std::flush;
    std::size_t count(10000000);
    std::vector<int> random;
    random.reserve(count);
    std::generate_n(std::back_inserter(random), count, [](){ return std::rand(); });
    std::cout << "done\n";

    time<std::allocator<int>>("default allocator      ", random);
    time<pool_allocator<int, 32>>("custom allocator (32)  ", random);
    time<pool_allocator<int, 256>>("custom allocator (256) ", random);
    time<pool_allocator<int, 1024>>("custom allocator (1024)", random);
    time<pool_allocator<int, 2048>>("custom allocator (2048)", random);
    time<pool_allocator<int, 4096>>("custom allocator (4096)", random);
    time<std::allocator<int>>("default allocator      ", random);
}

// results from clang/libc++:
// preparing...done
// running default allocator      : size=10000000 time=13927ms
// running custom allocator (32)  : size=10000000 time=9260ms
// running custom allocator (256) : size=10000000 time=9511ms
// running custom allocator (1024): size=10000000 time=9172ms
// running custom allocator (2048): size=10000000 time=9153ms
// running custom allocator (4096): size=10000000 time=9599ms
// running default allocator      : size=10000000 time=13730ms

// results from gcc/libstdc++:
// preparing...done
// running default allocator      : size=10000000 time=15814ms
// running custom allocator (32)  : size=10000000 time=10868ms
// running custom allocator (256) : size=10000000 time=10229ms
// running custom allocator (1024): size=10000000 time=10556ms
// running custom allocator (2048): size=10000000 time=10392ms
// running custom allocator (4096): size=10000000 time=10664ms
// running default allocator      : size=10000000 time=17941ms