HashMap.h

A hash table mostly compatible with the C++11 std::unordered_map interface, but with much higher performance for many workloads.

Implementation

This hash table uses open addressing with linear probing and backshift deletion. Open addressing and linear probing minimizes memory allocations and achieves high cache efficiency. Backshift deletion keeps performance high for delete heavy workloads by not clobbering the hash table with tombestones.

Usage

HashMap is mostly compatible with the C++11 container interface. The main differences are:

• A key value to represent the empty key is required.
• Key and T needs to be default constructible.
• Iterators are invalidated on all modifying operations.
• It's invalid to perform any operations with the empty key.
• Destructors are not called on erase.
• Extensions for lookups using related key types.

Member functions:

• HashMap(size_type bucket_count, key_type empty_key);

  Construct a HashMap with bucket_count buckets and empty_key as the empty key.

The rest of the member functions are implemented as for std::unordered_map.

Example

  using namespace rigtorp;

  // Hash for using std::string as lookup key
  struct Hash {
    size_t operator()(int v) { return v * 7; }
    size_t operator()(const std::string &v) { return std::stoi(v) * 7; }
  };

  // Equal comparison for using std::string as lookup key
  struct Equal {
    bool operator()(int lhs, int rhs) { return lhs == rhs; }
    bool operator()(int lhs, const std::string &rhs) {
      return lhs == std::stoi(rhs);
    }
  };

  // Create a HashMap with 16 buckets and 0 as the empty key
  HashMap<int, int, Hash, Equal> hm(16, 0);
  hm.emplace(1, 1);
  hm[2] = 2;

  // Iterate and print key-value pairs
  for (const auto &e : hm) {
    std::cout << e.first << " = " << e.second << "\n";
  }

  // Lookup using std::string
  std::cout << hm.at("1") << "\n";

  // Erase entry
  hm.erase(1);

Build and install

The header-only HashMap::HashMap CMake target requires C++14. Configure with CMake 3.20 or newer:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
ctest --test-dir build --output-on-failure
cmake --install build --prefix /path/to/install

Standalone builds enable tests and examples, plus the benchmark on Linux and macOS. These are disabled by default when included with add_subdirectory.

CMake option	Standalone default	Purpose
HASHMAP_BUILD_TESTS	ON	Build tests when BUILD_TESTING is also enabled
HASHMAP_BUILD_EXAMPLES	ON	Build the example
HASHMAP_BUILD_BENCHMARKS	ON on POSIX, OFF on Windows	Build the benchmark
HASHMAP_INSTALL	ON	Generate installation and package files
HASHMAP_WARNINGS_AS_ERRORS	OFF	Treat warnings in project executables as errors

To consume an installed package:

find_package(HashMap 1 CONFIG REQUIRED)
target_link_libraries(my_app PRIVATE HashMap::HashMap)

The same target is available after add_subdirectory(path/to/HashMap). Installation respects CMAKE_INSTALL_INCLUDEDIR and CMAKE_INSTALL_LIBDIR.

Benchmark

A benchmark src/HashMapBenchmark.cpp is included with the sources. The benchmark simulates a delete heavy workload where items are repeatedly inserted and deleted.

The benchmark uses std::chrono::steady_clock and requires no benchmark framework or vendored dependencies. CMake enables comparisons against installed Abseil and sparsehash when found. Set HASHMAP_BENCHMARK_WITH_ABSL=OFF or HASHMAP_BENCHMARK_WITH_SPARSEHASH=OFF to disable either comparison. Abseil must provide its CMake package; sparsehash must provide google/dense_hash_map. No dependencies are downloaded during configuration.

The benchmark is built on Linux and macOS; Windows builds the example and tests. All containers in a run use the same hash: hardware CRC32 on x86-64 and MurmurHash3's fmix64 integer mixer on ARM and other architectures. Hash choices differ across architectures, so compare containers within the same run.

I ran this benchmark on the following configuration:

• AMD Ryzen 9 3900X
• Linux 5.8.4-200.fc32.x86_64
• gcc (GCC) 10.2.1 20200723 (Red Hat 10.2.1-1)
• Isolated a core complex (CCX) using isolcpus for running the benchmark

When working set fits in L3 cache (HashMapBenchmark -c 100000 -i 100000000):

Implementation	mean ns/iter	max ns/iter
HashMap	24	1082
absl::flat_hash_map	24	2074
google::dense_hash_map	49	689846
std::unordered_map	67	10299

When working set is larger than L3 cache (HashMapBenchmark -c 10000000 -i 1000000000):

Implementation	mean ns/iter	max ns/iter
HashMap	75	19026
absl::flat_hash_map	101	19848
google::dense_hash_map	111	226083255
std::unordered_map	408	22422

Cited by

HashMap has been cited by the following papers:

• Koppl, Dominik. “Separate Chaining Meets Compact Hashing.” (2019). https://arxiv.org/abs/1905.00163

About

This project was created by Erik Rigtorp <erik@rigtorp.se>.