array-hash

A C++ implementation of a fast and memory efficient hash map/set for strings

Cache conscious hash map and hash set for strings based on the "Cache-conscious collision resolution in string hash tables." (Askitis Nikolas and Justin Zobel, 2005) paper. You can find some details regarding the structure here.

Thanks to its cache friendliness, the structure provides fast lookups while keeping a low memory usage. The main drawback is the rehash process which is a bit slow and need some spare memory to copy the strings from the old hash table to the new hash table (it can’t use std::move as the other hash tables using std::string as key).

Four classes are provided: tsl::array_map, tsl::array_set, tsl::array_pg_map and tsl::array_pg_set. The first two are faster and use a power of two growth policy, the last two use a prime growth policy instead and are able to cope better with a poor hash function. Use the prime version if there is a chance of repeating patterns in the lower bits of your hash (e.g. you are storing pointers with an identity hash function). See GrowthPolicy for details.

A benchmark of tsl::array_map against other hash maps can be found here. This page also gives some advices on which hash table structure you should try for your use case (useful if you are a bit lost with the multiple hash tables implementations in the tsl namespace). You can also find another benchmark on the tsl::hat-trie page.

Overview

• Header-only library, just add the include directory to your include path and you are ready to go. If you use CMake, you can also use the tsl::array_hash exported target from the CMakeLists.txt.
• Low memory usage with good performances, see the benchmark for some numbers.
• Support for move-only and non-default constructible values.
• Strings with null characters inside them are supported (you can thus store binary data as key).
• If the hash is known before a lookup, it is possible to pass it as parameter to speed-up the lookup (see precalculated_hash parameter in API).
• Support for efficient serialization and deserialization (see example and the serialize/deserialize methods in the API for details).
• By default the maximum allowed size for a key is set to 65 535. This can be raised through the KeySizeT template parameter (see API for details).
• By default the maximum size of the map is limited to 4 294 967 296 elements. This can be raised through the IndexSizeT template parameter (see API for details).

Differences compared to std::unordered_map

tsl::array_map tries to have an interface similar to std::unordered_map, but some differences exist:

• Iterator invalidation doesn't behave in the same way, any operation modifying the hash table invalidate them (see API for details).
• References and pointers to keys or values in the map are invalidated in the same way as iterators to these keys-values.
• Erase operations have an amortized runtime complexity of O(1) for tsl::array_map. An erase operation will delete the key immediately but for the value part of the map, the deletion may be delayed. The destructor of the value is only called when the ratio between the size of the map and the size of the map + the number of deleted values still stored is low enough. The method shrink_to_fit may be called to force the deletion.
• The key and the value are stored separately and not in a std::pair. Methods like insert or emplace take the key and the value separately instead of a std::pair. The insert method looks like std::pair insert(const CharT* key, const T& value) instead of std::pair insert(const std::pair& value) (see API for details).
• For iterators, operator*() and operator->() return a reference and a pointer to the value T instead of std::pair. For an access to the key string, the key() (which returns a const CharT*) or key_sv() (which returns a std::basic_string_view) method of the iterator must be called.
• No support for some bucket related methods (like bucket_size, bucket, ...).

These differences also apply between std::unordered_set and tsl::array_set.

Thread-safety and exception guarantees are similar to the STL containers.

Hash function

The default hash function used by the structure depends on the presence of std::string_view. If it is available, std::hash is used, otherwise a simple FNV-1a hash function is used to avoid any dependency.

If you can't use C++17 or later, we recommend to replace the hash function with something like CityHash, MurmurHash, FarmHash, ... for better performances. On the tests we did, CityHash64 offers a ~40% improvement on reads compared to FNV-1a.

#include 

struct str_hash {
    std::size_t operator()(const char* key, std::size_t key_size) const {
        return CityHash64(key, key_size);
    }
};

tsl::array_map map;

The std::hash can't be used efficiently as the structure doesn't store any std::string object. Any time a hash would be needed, a temporary std::string would have to be created.

Growth policy

The library supports multiple growth policies through the GrowthPolicy template parameter. Three policies are provided by the library but you can easily implement your own if needed.

• tsl::ah::power_of_two_growth_policy. Default policy used by tsl::array_map/set. This policy keeps the size of the bucket array of the hash table to a power of two. This constraint allows the policy to avoid the usage of the slow modulo operation to map a hash to a bucket, instead of hash % 2n, it uses hash & (2n - 1) (see fast modulo). Fast but this may cause a lot of collisions with a poor hash function as the modulo with a power of two only masks the most significant bits in the end.
• tsl::ah::prime_growth_policy. Default policy used by tsl::array_pg_map/set. The policy keeps the size of the bucket array of the hash table to a prime number. When mapping a hash to a bucket, using a prime number as modulo will result in a better distribution of the hash across the buckets even with a poor hash function. To allow the compiler to optimize the modulo operation, the policy use a lookup table with constant primes modulos (see API for details). Slower than tsl::ah::power_of_two_growth_policy but more secure.
• tsl::ah::mod_growth_policy. The policy grows the map by a customizable growth factor passed in parameter. It then just use the modulo operator to map a hash to a bucket. Slower but more flexible.

To implement your own policy, you have to implement the following interface.

…

Installation

To use the library, just add the include directory to your include path. It is a header-only library.

If you use CMake, you can also use the tsl::array_hash exported target from the CMakeLists.txt with target_link_libraries.

# Example where the array-hash project is stored in a third-party directory
add_subdirectory(third-party/array-hash)
target_link_libraries(your_target PRIVATE tsl::array_hash)  

If the project has been installed through make install, you can also use find_package(tsl-array-hash REQUIRED) instead of add_subdirectory.

The code should work with any C++11 standard-compliant compiler and has been tested with GCC 4.8.4, Clang 3.5.0 and Visual Studio 2015.

To run the tests you will need the Boost Test library and CMake.

git clone https://github.com/Tessil/array-hash.git
cd array-hash/tests
mkdir build
cd build
cmake ..
cmake --build .
./tsl_array_hash_tests

Usage

The API can be found here. If std::string_view is available, the API changes slightly and can be found here.

Example

…

struct deserializer {
    // Must support the following types for U: std::uint64_t, float and T if a map is used.
    template
    U operator()();
    void operator()(CharT* value_out, std::size_t value_size);
};

Note that the implementation leaves binary compatibility (endianness, float binary representation, size of int, ...) of the types it serializes/deserializes in the hands of the provided function objects if compatibility is required.

More details regarding the serialize and deserialize methods can be found in the API.

…

Serialization with Boost Serialization and compression with zlib

It's possible to use a serialization library to avoid some of the boilerplate if the types to serialize are more complex.

The following example uses Boost Serialization with the Boost zlib compression stream to reduce the size of the resulting serialized file.

…

License

The code is licensed under the MIT license, see the LICENSE file for details.