The Parallel Hashmap

Overview

This repository aims to provide an set of excellent hash map implementations, with the following characteristics:

• Header only: nothing to build, just copy the parallel_hashmap directory to your project and you are good to go.

• drop-in replacement for std::unordered_map and std::unordered_set

• Compiler with C++11 support required, C++14 and C++17 APIs are provided (such as try_emplace)

• Very efficient, significantly faster than your compiler's unordered map/set or Boost's, or than sparsepp

• Memory friendly: low memory usage, although a little higher than sparsepp

• Supports heterogeneous lookup

• Easy to forward declare: just include phmap_fwd_decl.h in your header files to forward declare Parallel Hashmap containers. This header is only 111 lines including comments.

• Tested on Windows (vs2015 & vs2017), linux (g++ 5, 6, 7, 8, clang++ 3.9, 4.0, 5.0) and MacOS (g++ and clang++) - click on travis and appveyor icons above for detailed test status.

Fast and memory friendly

Click here For a full writeup explaining the design and benefits of the Parallel Hashmap.

The hashmaps provided here are built upon those open sourced by Google in the Abseil library. They use closed hashing, where values are stored directly into a memory array, avoiding memory indirections. By using parallel SSE2 instructions, these hashmaps are able to look up items by checking 16 slots in parallel, allowing the implementation to remain fast even when the table is filled up to 87.5% capacity.

IMPORTANT: This repository borrows code from the abseil-cpp repository, with modifications, and may behave differently from the original. This repository is an independent work, with no guarantees implied or provided by the authors. Please visit abseil-cpp for the official Abseil libraries.

Installation

Copy the parallel_hashmap directory to your project. Update your include path. That's all.

A cmake configuration files (CMakeLists.txt) is provided for building the tests and examples. Command for building and running the tests is: mkdir build && cd build && cmake -DPHMAP_BUILD_TESTS=ON -DPHMAP_BUILD_EXAMPLES=ON .. && cmake --build . && make test

Example

#include <iostream>
#include <string>
#include <parallel_hashmap/phmap.h>

using phmap::flat_hash_map;
 
int main()
{
    // Create an unordered_map of three strings (that map to strings)
    flat_hash_map<std::string, std::string> email = 
    {
        { "tom",  "tom@gmail.com"},
        { "jeff", "jk@gmail.com"},
        { "jim",  "jimg@microsoft.com"}
    };
 
    // Iterate and print keys and values 
    for (const auto& n : email) 
        std::cout << n.first << "'s email is: " << n.second << "\n";
 
    // Add a new entry
    email["bill"] = "bg@whatever.com";
 
    // and print it
    std::cout << "bill's email is: " << email["bill"] << "\n";
 
    return 0;
}

Various hash maps and their pros and cons

The header parallel_hashmap/phmap.h provides the implementation for the following eight hash tables:

• phmap::flat_hash_set
• phmap::flat_hash_map
• phmap::node_hash_set
• phmap::node_hash_map
• phmap::parallel_flat_hash_set
• phmap::parallel_flat_hash_map
• phmap::parallel_node_hash_set
• phmap::parallel_node_hash_map

The full types with template parameters can be found in the parallel_hashmap/phmap_fwd_decl.h header, which is useful for forward declaring the Parallel Hashmaps when necessary.

Key decision points:

• The flat hash maps may move the keys and values in memory. So if you keep a pointer to something inside a flat hash map, this pointer may become invalid when the map is mutated. The node hash maps don't, and should be used instead if this is a problem.

• The flat hash maps will use less memory, and usually be faster than the node hash maps, so use them if you can. A possible exception is when the values inserted in the hash map are large (say more than 100 bytes [needs testing]).

• The parallel hash maps are preferred when you have a few hash maps that will store a very large number of values. The non-parallel hash maps are preferred if you have a large number of hash maps, each storing a relatively small number of values.

• The benefits of the parallel hash maps are:
  a. reduced peak memory usage (when resizing), and
  b. multithreading support (and inherent internal parallelism)

Changes to Abseil's hashmaps

• The default hash framework is std::hash, not absl::Hash. However, if you prefer the default to be the Abseil hash framework, include the Abseil headers before phmap.h and define the preprocessor macro PHMAP_USE_ABSL_HASHEQ.

• The erase(iterator) and erase(const_iterator) both return an iterator to the element following the removed element, as does the std::unordered_map. A non-standard void _erase(iterator) is provided in case the return value is not needed.

• No new types, such as absl::string_view, are provided. All types with a std::hash<> implementation are supported by phmap tables (including std::string_view of course if your compiler provides it).

• The Abseil hash tables internally randomize a hash seed, so that the table iteration order is non-deterministic. This can be useful to pervent Denial Of Service attacks when a hash table is used for a customer facing web service, but it can make debugging more difficult. The phmap hashmaps by default do not implement this randomization, but it can be enabled by adding #define PHMAP_NON_DETERMINISTIC 1 before including the header phmap.h (as is done in raw_hash_set_test.cc).

• Unlike the Abseil hash maps, we do an internal mixing of the hash value provided. This prevents serious degradation of the hash table performance when the hash function provided by the user has poor entropy distribution. The cost in performance is very minimal, and this helps provide reliable performance even with not so good hash functions.

Memory usage

type	memory usage	additional peak memory usage when resizing
flat tables
node tables
parallel flat tables
parallel node tables

• size() is the number of values in the container, as returned by the size() method
• load_factor() is the ratio: size() / bucket_count(). It varies between 0.4375 (just after the resize) to 0.875 (just before the resize). The size of the bucket array doubles at each resize.
• the value 9 comes from sizeof(void *) + 1, as the node hash maps store one pointer plus one byte of metadata for each entry in the bucket array.
• flat tables store the values, plus one byte of metadata per value), directly into the bucket array, hence the sizeof(C::value_type) + 1.
• the additional peak memory usage (when resizing) corresponds the the old bucket array (half the size of the new one, hence the 0.5), which contains the values to be copied to the new bucket array, and which is freed when the values have been copied.
• the parallel hashmaps, when created with a template parameter N=4, create 16 submaps. When the hash values are well distributed, and in single threaded mode, only one of these 16 submaps resizes at any given time, hence the factor 0.03 roughly equal to 0.5 / 16

Iterator invalidation

The rules are the same as for std::unordered_map, and are valid for all the phmap containers:

Operations	Invalidated
All read only operations, swap, std::swap	Never
clear, rehash, reserve, operator=	Always
insert, emplace, emplace_hint, operator[]	Only if rehash triggered
erase	Only to the element erased

Example 2 - providing a hash function for a user-defined class

In order to use a flat_hash_set or flat_hash_map, a hash function should be provided. Even though a the hash function can be provided via the HashFcn template parameter, we recommend injecting a specialization of std::hash for the class into the "std" namespace. We provide a convenient and small header phmap_utils.h which allows to easily add such specializations.

For example:

file "Person.h"

#include <parallel_hashmap/phmap_utils.h> // minimal header providing phmap::HashState()
#include <string>
using std::string;

struct Person
{
    bool operator==(const Person &o) const
    { 
        return _first == o._first && _last == o._last && _age == o._age; 
    }

    string _first;
    string _last;
    int    _age;
};

namespace std
{
    // inject specialization of std::hash for Person into namespace std
    // ----------------------------------------------------------------
    template<> struct hash<Person>
    {
        std::size_t operator()(Person const &p) const
        {
            return phmap::HashState().combine(0, p._first, p._last, p._age);
        }
    };
}

The std::hash specialization for Person combines the hash values for both first and last name and age, using the convenient phmap::HashState() function, and returns the combined hash value.

file "main.cpp"

#include "Person.h"   // defines Person  with std::hash specialization

#include <iostream>
#include <parallel_hashmap/phmap.h>

int main()
{
    // As we have defined a specialization of std::hash() for Person,
    // we can now create sparse_hash_set or sparse_hash_map of Persons
    // ----------------------------------------------------------------
    phmap::flat_hash_set<Person> persons = 
        { { "John", "Mitchell", 35 },
          { "Jane", "Smith",    32 },
          { "Jane", "Smith",    30 },
        };

    for (auto& p: persons)
        std::cout << p._first << ' ' << p._last << " (" << p._age << ")" << '\n';

}

Thread safety

Parallel Hashmap containers follow the thread safety rules of the Standard C++ library. In Particular:

• A single phmap hash table is thread safe for reading from multiple threads. For example, given a hash table A, it is safe to read A from thread 1 and from thread 2 simultaneously.

• If a single hash table is being written to by one thread, then all reads and writes to that hash table on the same or other threads must be protected. For example, given a hash table A, if thread 1 is writing to A, then thread 2 must be prevented from reading from or writing to A.

• It is safe to read and write to one instance of a type even if another thread is reading or writing to a different instance of the same type. For example, given hash tables A and B of the same type, it is safe if A is being written in thread 1 and B is being read in thread 2.

• The parallel tables can be made fully thread-safe, by providing a synchronization type (for example std::mutex) as the last template argument. Because locking is performed at the submap level, a high level of concurrency can still be achieved.

• Examples on how to use various mutex types, including boost::mutex, boost::shared_mutex and absl::Mutex can be found in examples/bench.cc

Using the Parallel Hashmap from languages other than C++

While C++ is the native language of the Parallel Hashmap, we welcome bindings making it available for other languages. One such implementation has been created for Python and is described below:

• GetPy - A Simple, Fast, and Small Hash Map for Python: GetPy is a thin and robust binding to The Parallel Hashmap (https://github.com/greg7mdp/parallel-hashmap.git) which is the current state of the art for minimal memory overhead and fast runtime speed. The binding layer is supported by PyBind11 (https://github.com/pybind/pybind11.git) which is fast to compile and simple to extend. Serialization is handled by Cereal (https://github.com/USCiLab/cereal.git) which supports streaming binary serialization, a critical feature for the large hash maps this package is designed to support.