とても簡単なリングバッファ

command

clang++ -DNDEBUG -O2 -std=c++17 main.cc
g++ -DNDEBUG -O2 -std=c++17 main.cc

LICENSE.txt

Copyright 2023 Hiroki Kumazaki

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

main.cc

#include <assert.h>
#include <atomic>
#include <chrono>
#include <cstddef>
#include <cstdint>
#include <iomanip>
#include <iostream>
#include <mutex>
#include <queue>
#include <thread>
#define _XOPEN_SOURCE_EXTENDED 1
#include <sys/mman.h>
#include <vector>

class StdQueue {
public:
  explicit StdQueue() {}
  bool enqueue(int item) {
    buffer_.push(item);
    return true;
  }
  bool dequeue(int *dest) {
    if (buffer_.empty()) {
      return false;
    }
    *dest = buffer_.front();
    return true;
  }

private:
  std::queue<int> buffer_;
};

class RingBuffer0 {
public:
  explicit RingBuffer0(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is empty.
  bool enqueue(int item) {
    if (write_idx_ - read_idx_ == buffer_.size()) {
      return false;
    }
    buffer_[write_idx_ % buffer_.size()] = item;
    write_idx_++;
    return true;
  }

  // Returns true on success. Fails if the buffer is full.
  bool dequeue(int *dest) {
    if (write_idx_ == read_idx_) {
      return false;
    }
    *dest = buffer_[read_idx_ % buffer_.size()];
    read_idx_++;
    return true;
  }

private:
  std::vector<int> buffer_;
  uint64_t read_idx_{0};
  uint64_t write_idx_{0};
};

class RingBuffer1 {
public:
  explicit RingBuffer1(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is empty.
  bool enqueue(int item) {
    if (write_idx_ - read_idx_ == buffer_.size()) {
      return false;
    }
    buffer_[write_idx_ & (buffer_.size() - 1)] = item;
    write_idx_++;
    return true;
  }

  // Returns true on success. Fails if the buffer is full.
  bool dequeue(int *dest) {
    if (write_idx_ == read_idx_) {
      return false;
    }
    *dest = buffer_[read_idx_ & (buffer_.size() - 1)];
    read_idx_++;
    return true;
  }

private:
  std::vector<int> buffer_;
  uint64_t read_idx_{0};
  uint64_t write_idx_{0};
};

class RingBufferMutex {
public:
  explicit RingBufferMutex(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is empty.
  bool enqueue(int item) {
    std::scoped_lock lock(mutex_);
    if (write_idx_ - read_idx_ == buffer_.size()) {
      return false;
    }
    buffer_[write_idx_ & (buffer_.size() - 1)] = item;
    write_idx_++;
    return true;
  }

  // Returns true on success. Fails if the buffer is full.
  bool dequeue(int *dest) {
    std::scoped_lock lock(mutex_);
    if (write_idx_ == read_idx_) {
      return false;
    }
    *dest = buffer_[read_idx_ & (buffer_.size() - 1)];
    read_idx_++;
    return true;
  }

private:
  std::mutex mutex_;
  std::vector<int> buffer_;
  uint64_t read_idx_{0};
  uint64_t write_idx_{0};
};

class RingBuffer2 {
public:
  explicit RingBuffer2(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is empty.
  bool enqueue(int item) {
    uint64_t write_idx = write_idx_.load(std::memory_order_relaxed);
    uint64_t read_idx = read_idx_.load(std::memory_order_acquire);
    if (write_idx - read_idx == buffer_.size()) {
      return false;
    }
    buffer_[write_idx & (buffer_.size() - 1)] = item;
    write_idx_.store(write_idx + 1, std::memory_order_release);
    return true;
  }

  // Returns true on success. Fails if the buffer is full.
  bool dequeue(int *dest) {
    uint64_t read_idx = read_idx_.load(std::memory_order_relaxed);
    uint64_t write_idx = write_idx_.load(std::memory_order_acquire);
    if (write_idx == read_idx) {
      return false;
    }
    *dest = buffer_[read_idx & (buffer_.size() - 1)];
    read_idx_.store(read_idx + 1, std::memory_order_release);
    return true;
  }

private:
  std::vector<int> buffer_;
  alignas(64) std::atomic<uint64_t> read_idx_{0};
  alignas(64) std::atomic<uint64_t> write_idx_{0};
};

class RingBuffer3 {
public:
  explicit RingBuffer3(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is empty.
  bool enqueue(int item) {
    uint64_t write_idx = write_idx_.load(std::memory_order_relaxed);
    if (write_idx - cached_read_idx_ == buffer_.size()) {
      cached_read_idx_ = read_idx_.load(std::memory_order_acquire);
      assert(cached_read_idx_ <= write_idx);
      if (write_idx - cached_read_idx_ == buffer_.size()) {
        return false;
      }
    }
    buffer_[write_idx & (buffer_.size() - 1)] = item;
    write_idx_.store(write_idx + 1, std::memory_order_release);
    return true;
  }

  // Returns true on success. Fails if the buffer is full.
  bool dequeue(int *dest) {
    uint64_t read_idx = read_idx_.load(std::memory_order_relaxed);
    if (cached_write_idx_ == read_idx) {
      cached_write_idx_ = write_idx_.load(std::memory_order_acquire);
      assert(read_idx <= cached_write_idx_);
      if (cached_write_idx_ == read_idx) {
        return false;
      }
    }
    *dest = buffer_[read_idx & (buffer_.size() - 1)];
    read_idx_.store(read_idx + 1, std::memory_order_release);
    return true;
  }

private:
  std::vector<int> buffer_;
  alignas(64) std::atomic<uint64_t> read_idx_{0};
  alignas(64) uint64_t cached_read_idx_{0};
  alignas(64) std::atomic<uint64_t> write_idx_{0};
  alignas(64) uint64_t cached_write_idx_{0};
};

class RingBuffer4 {
public:
  explicit RingBuffer4(size_t size)
      : buffer_((int *)mmap(0, size * sizeof(int), PROT_READ | PROT_WRITE,
                            MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB, 0, 0)),
        size_(size) {}
  ~RingBuffer4() { munmap(buffer_, size_ * sizeof(int)); }

  // Returns true on success. Fails if the buffer is empty.
  bool enqueue(int item) {
    uint64_t write_idx = write_idx_.load(std::memory_order_relaxed);
    if (write_idx - cached_read_idx_ == size_) {
      cached_read_idx_ = read_idx_.load(std::memory_order_acquire);
      assert(cached_read_idx_ <= write_idx);
      if (write_idx - cached_read_idx_ == size_) {
        std::this_thread::sleep_for(std::chrono::milliseconds(1));
        return false;
      }
    }
    buffer_[write_idx & (size_ - 1)] = item;
    write_idx_.store(write_idx + 1, std::memory_order_release);
    return true;
  }

  // Returns true on success. Fails if the buffer is full.
  bool dequeue(int *dest) {
    uint64_t read_idx = read_idx_.load(std::memory_order_release);
    if (cached_write_idx_ == read_idx) {
      cached_write_idx_ = write_idx_.load(std::memory_order_acquire);
      assert(read_idx <= cached_write_idx_);
      if (cached_write_idx_ == read_idx) {
        std::this_thread::sleep_for(std::chrono::milliseconds(1));
        return false;
      }
    }
    *dest = buffer_[read_idx & (size_ - 1)];
    read_idx_.store(read_idx + 1, std::memory_order_release);
    return true;
  }

private:
  int *buffer_;
  size_t size_;
  alignas(64) std::atomic<uint64_t> read_idx_{0};
  alignas(64) uint64_t cached_read_idx_{0};
  alignas(64) std::atomic<uint64_t> write_idx_{0};
  alignas(64) uint64_t cached_write_idx_{0};
};

#define ASSERT_TRUE(x)                                                         \
  {                                                                            \
    if (!(x)) {                                                                \
      std::cerr << "Assertion fail: " #x << " at " << __LINE__ << "\n";        \
    }                                                                          \
  }

template <typename RingBufferType> void test() {
  RingBufferType rb(4);
  int result;
  ASSERT_TRUE(!rb.dequeue(&result));
  ASSERT_TRUE(rb.enqueue(1));
  ASSERT_TRUE(rb.enqueue(2));
  ASSERT_TRUE(rb.enqueue(3));
  ASSERT_TRUE(rb.enqueue(4));
  ASSERT_TRUE(!rb.enqueue(5));
  ASSERT_TRUE(rb.dequeue(&result));
  ASSERT_TRUE(result == 1);
  ASSERT_TRUE(rb.dequeue(&result));
  ASSERT_TRUE(result == 2);
  ASSERT_TRUE(rb.dequeue(&result));
  ASSERT_TRUE(result == 3);
  ASSERT_TRUE(rb.dequeue(&result));
  ASSERT_TRUE(result == 4);
  ASSERT_TRUE(!rb.dequeue(&result));
}

constexpr uint64_t kCount = 500000;

template <typename RingBufferType> double benchmark_single(RingBufferType &rb) {
  auto start = std::chrono::system_clock::now();
  int result;
  for (uint64_t i = 0; i < kCount; ++i) {
    for (int j = 0; j < 1000; ++j) {
      rb.enqueue(j);
    }
    for (int j = 0; j < 1000; ++j) {
      rb.dequeue(&result);
    }
  }
  auto end = std::chrono::system_clock::now();
  double duration =
      std::chrono::duration_cast<std::chrono::milliseconds>(end - start)
          .count();
  const int count = kCount * (1000 + 1000);
  std::cerr << count << " ops in " << duration << " ms \t";
  return count / duration;
}

template <typename RingBufferType> double benchmark(RingBufferType &rb) {
  auto start = std::chrono::system_clock::now();
  std::thread workers[2] = {
      std::thread([&]() {
        cpu_set_t cpuset;
        CPU_ZERO(&cpuset);
        CPU_SET(0, &cpuset);
        if (pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t),
                                   &cpuset) == -1) {
          perror("pthread_setaffinity_no");
          exit(1);
        }
        for (uint64_t i = 0; i < kCount; ++i) {
          int count = 1000;
          while (0 < count) {
            if (rb.enqueue(count)) {
              count--;
            }
          }
        }
      }),
      std::thread([&]() {
        cpu_set_t cpuset;
        CPU_ZERO(&cpuset);
        CPU_SET(1, &cpuset);
        if (pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t),
                                   &cpuset) == -1) {
          perror("pthread_setaffinity_no");
          exit(1);
        }
        int result;
        for (uint64_t i = 0; i < kCount; ++i) {
          int count = 1000;
          while (0 < count) {
            if (rb.dequeue(&result)) {
              count--;
            }
          }
        }
      })};
  for (auto &w : workers) {
    w.join();
  }
  auto end = std::chrono::system_clock::now();
  double duration =
      std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count();
  const int count = kCount * (1000 + 1000);
  std::cerr << count << " ops in " << duration << " ns \t";
  return 1000000.0 * kCount * (1000 + 1000) / duration;
}

int main() {
  test<RingBuffer1>();
  test<RingBuffer2>();
  test<RingBuffer3>();
  StdQueue queue;
  RingBuffer0 rb0(2 * 1024 * 1024);
  RingBuffer1 rb1(2 * 1024 * 1024);
  RingBufferMutex rbm(2 * 1024 * 1024);
  RingBuffer2 rb2(2 * 1024 * 1024);
  RingBuffer3 rb3(2 * 1024 * 1024);
  RingBuffer4 rb4(2 * 1024 * 1024);
  std::cout << std::setprecision(10);
  std::cout << "StdQueue_single: " << benchmark_single(queue) << " ops/ms\n";
  std::cout << "RingBufferMutex: " << benchmark_single(rbm) << " ops/ms\n";
  std::cout << "RingBuffer0_single: " << benchmark_single(rb0) << " ops/ms\n";
  std::cout << "RingBuffer1_single: " << benchmark_single(rb1) << " ops/ms\n";
  std::cout << "RingBuffer2_single: " << benchmark_single(rb2) << " ops/ms\n";
  std::cout << "RingBuffer3_single: " << benchmark_single(rb3) << " ops/ms\n";
  std::cout << "RingBufferMutex: " << benchmark(rbm) << " opms\n";
  std::cout << "RingBuffer2: " << benchmark(rb2) << " ops/ms\n";
  std::cout << "RingBuffer3: " << benchmark(rb3) << " ops/ms\n";
  std::cout << "RingBuffer4: " << benchmark(rb4) << " ops/ms\n";
}

ring_buffer_bench_mac.cc

#include <atomic>
#include <chrono>
#include <cstddef>
#include <cstdint>
#include <iomanip>
#include <iostream>
#include <mutex>
#include <queue>
#include <thread>
#include <vector>

class StdQueue {
public:
  explicit StdQueue() {}
  bool enqueue(int item) {
    buffer_.push(item);
    return true;
  }
  bool dequeue(int *dest) {
    if (buffer_.empty()) {
      return false;
    }
    *dest = buffer_.front();
    return true;
  }

private:
  std::queue<int> buffer_;
};

class RingBuffer0 {
public:
  explicit RingBuffer0(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is full.
  bool enqueue(int item) {
    if (write_idx_ - read_idx_ == buffer_.size()) {
      return false;
    }
    buffer_[write_idx_ % buffer_.size()] = item;
    write_idx_++;
    return true;
  }

  // Returns true on success. Fails if the buffer is empty.
  bool dequeue(int *dest) {
    if (write_idx_ == read_idx_) {
      return false;
    }
    *dest = buffer_[read_idx_ % buffer_.size()];
    read_idx_++;
    return true;
  }

private:
  std::vector<int> buffer_;
  uint64_t read_idx_{0};
  uint64_t write_idx_{0};
};

class RingBuffer1 {
public:
  explicit RingBuffer1(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is full.
  bool enqueue(int item) {
    if (write_idx_ - read_idx_ == buffer_.size()) {
      return false;
    }
    buffer_[write_idx_ & (buffer_.size() - 1)] = item;
    write_idx_++;
    return true;
  }

  // Returns true on success. Fails if the buffer is empty.
  bool dequeue(int *dest) {
    if (write_idx_ == read_idx_) {
      return false;
    }
    *dest = buffer_[read_idx_ & (buffer_.size() - 1)];
    read_idx_++;
    return true;
  }

private:
  std::vector<int> buffer_;
  uint64_t read_idx_{0};
  uint64_t write_idx_{0};
};

class RingBufferMutex {
public:
  explicit RingBufferMutex(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is full.
  bool enqueue(int item) {
    std::scoped_lock lock(mutex_);
    if (write_idx_ - read_idx_ == buffer_.size()) {
      return false;
    }
    buffer_[write_idx_ & (buffer_.size() - 1)] = item;
    write_idx_++;
    return true;
  }

  // Returns true on success. Fails if the buffer is empty.
  bool dequeue(int *dest) {
    std::scoped_lock lock(mutex_);
    if (write_idx_ == read_idx_) {
      return false;
    }
    *dest = buffer_[read_idx_ & (buffer_.size() - 1)];
    read_idx_++;
    return true;
  }

private:
  std::mutex mutex_;
  std::vector<int> buffer_;
  uint64_t read_idx_{0};
  uint64_t write_idx_{0};
};

class RingBuffer2 {
public:
  explicit RingBuffer2(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is full.
  bool enqueue(int item) {
    uint64_t write_idx = write_idx_.load(std::memory_order_relaxed);
    uint64_t read_idx = read_idx_.load(std::memory_order_acquire);
    if (write_idx - read_idx == buffer_.size()) {
      return false;
    }
    buffer_[write_idx & (buffer_.size() - 1)] = item;
    write_idx_.store(write_idx + 1, std::memory_order_release);
    return true;
  }

  // Returns true on success. Fails if the buffer is empty.
  bool dequeue(int *dest) {
    uint64_t read_idx = read_idx_.load(std::memory_order_relaxed);
    uint64_t write_idx = write_idx_.load(std::memory_order_acquire);
    if (write_idx == read_idx) {
      return false;
    }
    *dest = buffer_[read_idx & (buffer_.size() - 1)];
    read_idx_.store(read_idx + 1, std::memory_order_release);
    return true;
  }

private:
  std::vector<int> buffer_;
  alignas(64) std::atomic<uint64_t> read_idx_{0};
  alignas(64) std::atomic<uint64_t> write_idx_{0};
};

class RingBuffer3 {
public:
  explicit RingBuffer3(size_t size) : buffer_(size) {}

  // Returns true on success. Fails if the buffer is full.
  bool enqueue(int item) {
    uint64_t write_idx = write_idx_.load(std::memory_order_relaxed);
    if (write_idx - cached_read_idx_ == buffer_.size()) {
      cached_read_idx_ = read_idx_.load(std::memory_order_acquire);
      if (write_idx - cached_read_idx_ == buffer_.size()) {
        return false;
      }
    }
    buffer_[write_idx & (buffer_.size() - 1)] = item;
    write_idx_.store(write_idx + 1, std::memory_order_release);
    return true;
  }

  // Returns true on success. Fails if the buffer is empty.
  bool dequeue(int *dest) {
    uint64_t read_idx = read_idx_.load(std::memory_order_relaxed);
    if (cached_write_idx_ == read_idx) {
      cached_write_idx_ = write_idx_.load(std::memory_order_acquire);
      if (cached_write_idx_ == read_idx) {
        return false;
      }
    }
    *dest = buffer_[read_idx & (buffer_.size() - 1)];
    read_idx_.store(read_idx + 1, std::memory_order_release);
    return true;
  }

private:
  std::vector<int> buffer_;
  alignas(64) std::atomic<uint64_t> read_idx_{0};
  alignas(64) uint64_t cached_read_idx_{0};
  alignas(64) std::atomic<uint64_t> write_idx_{0};
  alignas(64) uint64_t cached_write_idx_{0};
};

constexpr uint64_t kCount = 500000;

template <typename RingBufferType> double benchmark_single(RingBufferType &rb) {
  auto start = std::chrono::system_clock::now();
  int result;
  for (uint64_t i = 0; i < kCount; ++i) {
    for (int j = 0; j < 1000; ++j) {
      rb.enqueue(j);
    }
    for (int j = 0; j < 1000; ++j) {
      rb.dequeue(&result);
    }
  }
  auto end = std::chrono::system_clock::now();
  double duration =
      std::chrono::duration_cast<std::chrono::milliseconds>(end - start)
          .count();
  const int count = kCount * (1000 + 1000);
  std::cerr << count << " ops in " << duration << " ms \t";
  return count / duration;
}

template <typename RingBufferType> double benchmark(RingBufferType &rb) {
  auto start = std::chrono::system_clock::now();
  std::thread workers[2] = {
      std::thread([&]() {
        for (uint64_t i = 0; i < kCount; ++i) {
          int count = 1000;
          while (0 < count) {
            if (rb.enqueue(count)) {
              count--;
            }
          }
        }
      }),
      std::thread([&]() {
        int result;
        for (uint64_t i = 0; i < kCount; ++i) {
          int count = 1000;
          while (0 < count) {
            if (rb.dequeue(&result)) {
              count--;
            }
          }
        }
      })};
  for (auto &w : workers) {
    w.join();
  }
  auto end = std::chrono::system_clock::now();
  double duration =
      std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count();
  const int count = kCount * (1000 + 1000);
  std::cerr << count << " ops in " << duration << " ns \t";
  return 1000000.0 * kCount * (1000 + 1000) / duration;
}

int main() {
  StdQueue queue;
  RingBuffer0 rb0(2 * 1024 * 1024);
  RingBuffer1 rb1(2 * 1024 * 1024);
  RingBufferMutex rbm(2 * 1024 * 1024);
  RingBuffer2 rb2(2 * 1024 * 1024);
  RingBuffer3 rb3(2 * 1024 * 1024);
  std::cout << std::setprecision(10);
  std::cout << "StdQueue_single: " << benchmark_single(queue) << " ops/ms\n";
  std::cout << "RingBufferMutex: " << benchmark_single(rbm) << " ops/ms\n";
  std::cout << "RingBuffer0_single: " << benchmark_single(rb0) << " ops/ms\n";
  std::cout << "RingBuffer1_single: " << benchmark_single(rb1) << " ops/ms\n";
  std::cout << "RingBuffer2_single: " << benchmark_single(rb2) << " ops/ms\n";
  std::cout << "RingBuffer3_single: " << benchmark_single(rb3) << " ops/ms\n";
  std::cout << "RingBufferMutex: " << benchmark(rbm) << " opms\n";
  std::cout << "RingBuffer2: " << benchmark(rb2) << " ops/ms\n";
  std::cout << "RingBuffer3: " << benchmark(rb3) << " ops/ms\n";
}