set vs unordered_set for fastest iteration

问题:

In my application I have the following requirements -

1. The data structure will be populated just once with some values (not key/value pairs). The values may be repeated but I want the data structure to store them once only.

2. I will be iterating 100s of times through all the elements of the data structure created above. The order in which the elements appear in the iteration is immaterial.

Constraint 1 suggests that I will either have to use set or unordered_set as data is not in the form of key value pairs.

Now set insertion is costlier than unordered_set insertion but the data structure is populated only once in the beginning of my program.

I believe the deciding factor will be how fast I can iterate through all the elements of the data structure. I am not sure whether set or unordered_set will be faster for this purpose. I believe the standard makes no mention of this fact as this operation will be O(n) for either data structure. But I wonder for which data structure iterator.next() will be faster.

回答1:

There are several approaches.

1. The comments to your question suggest keeping a std::unordered_set that has the fastest O(1) lookup/insertion and O(N) iteration (as has every container). If you have data that changes a lot, or requires a lot of random lookups, this is probably the fastest. But test.
2. If you need to iterate 100s of times without intermediate insertions, you can do a single O(N) copy to a std::vector and gain from contiguous memory layout 100s of times. Test whether this is faster than a regular std::unordered_set.
3. If you have a small number of intermediate insertions between iterations, it could pay to use a dedicated vector. If you can use Boost.Container, try boost::flat_set which offers a std::set interface with a std::vector storage back-end (i.e. a contiguous memory layout that is very cache- and prefetch friendly). Again, test whether this gives a speed-up to the other two solutions.

For the last solution, see the Boost documentation for some of the tradeoffs (it's good to be aware of all the other issues like iterator invalidation, move semantics and exception safety as well):

Boost.Container flat_[multi]map/set containers are ordered-vector based associative containers based on Austern's and Alexandrescu's guidelines. These ordered vector containers have also benefited recently with the addition of move semantics to C++, speeding up insertion and erasure times considerably. Flat associative containers have the following attributes:

• Faster lookup than standard associative containers
• Much faster iteration than standard associative containers
• Less memory consumption for small objects (and for big objects if shrink_to_fit is used)
• Improved cache performance (data is stored in contiguous memory)
• Non-stable iterators (iterators are invalidated when inserting and erasing elements)
• Non-copyable and non-movable values types can't be stored
• Weaker exception safety than standard associative containers (copy/move constructors can throw when shifting values in erasures and insertions)
• Slower insertion and erasure than standard associative containers (specially for non-movable types)

NOTE: with faster lookup, it is meant that a flat_set does O(log N) on contiguous memory rather than O(log N) pointer chasing of a regular std::set. Of course, a std::unordered_set does O(1) lookup, which will faster for large N.

回答2:

I'd suggest you to use either set or unordered_set for "filtration" and when you are done, move data to vector of fixed size

回答3:

If building of the data-structure does not factor into the performance-concerns (or at least only marginally), consider saving your data into a std::vector: There's nothing beating it.

For speeding up the initial building of the data-structure, you might first insert into a std::unordered_set or at least use one for checking existence before insertion.

In the second case it need not contain the elements, but could contain e.g. indices.

std::vector<T> v;
auto h = [&v](size_t i){return std::hash<T>()(v[i]);};
auto c = [&v](size_t a, size_t b){return v[a] == v[b];};
std::unordered_set<size_t, decltype(h), decltype(c)> tester(0, h, c);

回答4:

I highly recommend you not to use in such this case. set is binary tree, and unordered_set is hash table - so they use lots of memory, and have slow iteration speed and bad locality of reference. If you have to insert/remove/find data frequently, set or unordered_set good choice, but now you need to just read, store, sort data once and only use data many times.

In this case, sorted vector can be such a good choice. vector is dynamic array, so it has low overhead.

Just directly, see the code.

std::vector<int> data;

int input;
for (int i = 0; i < 10; i++)
{
    std::cin >> input;
    data.push_back(input); // store data
}

std::sort(data.begin(), data.end()); // sort data

That's all. All your data is ready.

If you need to remove duplicates like set, just use unique - erase after sorting.

data.erase(
    std::unique(data.begin(), data.end()),
    data.end()
    );

Notice that you should use lower_bound, upper_bound and equal_range rather than find or find_if to use the benefits of sorted data.

回答5:

An unordered-set uses a hash-table to provide near O(1) time searching. This is done by using a hash of the key to calculate the offset of the element-you-are-seeking (keys) from the beginning of the dataset. Unless your dataset is small (like chars) different keys may have the same hash (a collision).

To minimize collisions a unordered-set will have to keep the data-store fairly sparsely populated. This means that finding a key will most be O(1) time (unless there is a collision).

However when iterating through a hash-table our iterator will encounter a lot of unused space in our datastore which will slow down the finding of the next element by our iterator. We could link adjacent elements in the hash-table with extra pointers but I do not think an unordered-set does so.

In light of the above, I will suggest you use a sorted vector for your "set". Using bisections you can search the store in O(log n) time and iterating through the list is trivial. A vector has the added advantage that the memory is contiguous so you are less likely to experience cache misses.