实现

统一封装

为了便于测试，这里对 AtomicSharedMutex 和 shared_mutex 进行统一封装：

class STDRWLock {
 public:
  STDRWLock() = default;
  ~STDRWLock() = default;
  void RLock() {
    mutex_.lock_shared();
  }
  void RUnlock() {
    mutex_.unlock_shared();
  }
  void WLock() {
    mutex_.lock();
  }
  void WUnlock() {
    mutex_.unlock();
  }
 private:
  std::shared_mutex mutex_;
};

class AtomicRWLock {
 public:
  AtomicRWLock() = default;
  ~AtomicRWLock() = default;
  void RLock() {
    mutex_.LockShared();
  }
  void RUnlock() {
    mutex_.UnlockShared();
  }
  void WLock() {
    mutex_.Lock();
  }
  void WUnlock() {
    mutex_.Unlock();
  }
 private:
  AtomicSharedMutex mutex_;
};

实现基础测试用例

实现一个模版类用于测试不同的锁：

template <class Lock>
class RWLockBenchMark {
 public:
  RWLockBenchMark() = default;
  ~RWLockBenchMark() = default;
  int GetValue() {
    int value = 0;
    lock_.RLock();
    value = num_;
    lock_.RUnlock();
    return value;
  }
  void ModifyValue(int& value) {
    lock_.WLock();
    num_ = value;
    lock_.WUnlock();
  }
 private:
  int num_;
  Lock lock_;
};

实现单元测试

实现基础的单元测试，同时针对标准库与原子实现对应的测试函数，这里的结果以元组的形式进行输出便于后续使用脚本语言进行分析：

/*
 * 测试线程数
 */
int g_thread_count        = 4;
/*
 * 循环次数
 */
int g_loop_count          = 10000;
/*
 * 测试迭代次数
 */
int g_epoch               = 10;

// ...

template <class Lock>
void test_impl(clock_t& reader_total, clock_t& writer_total) {
  RWLockBenchMark<Lock> ops;
  clock_t writer_spend_time{}, reader_spend_time[g_thread_count]{}, reader_time {};
  std::thread* reader_ths[g_thread_count]{};
  std::thread writer_th;
  int value;
  
  auto reader = [&](int index) {
    clock_t start = clock();
    for (int i = 0; i < g_loop_count; i++) {
      value = ops.GetValue();
    }
    clock_t end = clock();
    reader_spend_time[index] = end - start;
  };

  auto writer = [&]{
    clock_t start = clock();
    for (int i = 0; i < g_loop_count; i++) {
      ops.ModifyValue(i);
    }
    clock_t end = clock();
    writer_spend_time = end - start;
  };

  for (int i = 0; i < g_thread_count; i++) {
    reader_ths[i] = new std::thread(reader, i);
  }

  writer_th = std::thread(writer);

  for (int i = 0; i < g_thread_count; i++) {
    reader_ths[i]->join();
  }
  writer_th.join();

  for (int i = 0; i < g_thread_count; i++) {
    reader_time += reader_spend_time[i];
  }
  reader_time /= g_thread_count;

  reader_total = reader_time;
  writer_total = writer_spend_time;
}

void std_test() {
  clock_t writer_spend, reader_spend;
  clock_t writer_spend_sum{}, reader_spend_sum{};

  for (int i = 0; i < g_epoch; i++) {
    test_impl<STDRWLock>(reader_spend, writer_spend);
    writer_spend_sum += writer_spend;
    reader_spend_sum += reader_spend;
  }

  printf("(%d, %ld, %ld)\n", g_thread_count, reader_spend_sum / g_epoch, writer_spend_sum / g_epoch);
}

void atomic_test() {
  clock_t writer_spend, reader_spend;
  clock_t writer_spend_sum{}, reader_spend_sum{};

  for (int i = 0; i < g_epoch; i++) {
    test_impl<AtomicRWLock>(reader_spend, writer_spend);
    writer_spend_sum += writer_spend;
    reader_spend_sum += reader_spend;
  }

  printf("(%d, %ld, %ld)\n", g_thread_count, reader_spend_sum / g_epoch, writer_spend_sum / g_epoch);
}

测试

由于手边只有一种类型的 Arm 设备，虽然是 A58 但都是 ArmV7 的版本，这里暂时只能使用这两种进行对比

测试设备：

Intel(R) Core(TM) i5-7300HQ 4核
A58-ArmV7 4核

编译选项
两者均以 -O2 等级进行优化生成

读写比例
测试的线程读写线程比例为 $[2, 100] : 1$ ，随着读者线程数的上升统计读写线程进行完整操作消耗时间的平均值

图表颜色说明

蓝线为标准库实现
红线为原子实现

X86

X86 下的测试结果

---
config:
    xyChart:
        width: 900
        height: 600
    themeVariables:
        xyChart:
            plotColorPalette: "#28678D,#f34544"
---

xychart-beta
title "读者操作"
x-axis "读者线程数" 2 --> 100
y-axis "操作消耗平均时间（ms）"
line [1113, 2398, 2298, 1355, 1515, 1775, 1593, 1711, 585, 1179, 1684, 1778, 1779, 1777, 1635, 1880, 1562, 1563, 1386, 1648, 1639, 1762, 1759, 1682, 1810, 1572, 1535, 1020, 1804, 1469, 1710, 1595, 1857, 1713, 1778, 1840, 1602, 1712, 1678, 1623, 1899, 1409, 1624, 1631, 1822, 1763, 1648, 1734, 1535, 1474, 1418, 1676, 1759, 1853, 1679, 1554, 1853, 1446, 1394, 1630, 1697, 1401, 1646, 1633, 1907, 1437, 1859, 1642, 1339, 1685, 1644, 1759, 1513, 1491, 1597, 1732, 1745, 1785, 1612, 1558, 1662, 1909, 1825, 1556, 1769, 1688, 1707, 1633, 1880, 1763, 1644, 1805, 1760, 1641, 1616, 1599, 1872, 1664, 1863, 1843]
line [148, 223, 220, 295, 340, 374, 456, 424, 508, 642, 639, 537, 514, 767, 768, 902, 946, 743, 1139, 1066, 1115, 1199, 937, 828, 761, 880, 1274, 1328, 1694, 1984, 2060, 1663, 2150, 1857, 2057, 1927, 2562, 2247, 2706, 2146, 2549, 3159, 2722, 3591, 4072, 2789, 2988, 2738, 2978, 4366, 2737, 2227, 2350, 4169, 3659, 3911, 3272, 3503, 4808, 5675, 3733, 4616, 4769, 2860, 3903, 2554, 3733, 4283, 3347, 4527, 4993, 3682, 4607, 3590, 4797, 4264, 3591, 4173, 4433, 4130, 4328, 4294, 6093, 4001, 4038, 4253, 4219, 3455, 4166, 5482, 5105, 3708, 3998, 4370, 3843, 4254, 3909, 3932, 4755]

---
config:
    xyChart:
        width: 900
        height: 600
    themeVariables:
        xyChart:
            plotColorPalette: "#28678D,#f34544"
---

xychart-beta
title "写者操作"
x-axis "读者线程数" 2 --> 100
y-axis "操作消耗平均时间（ms）"
line [1290, 3892, 4222, 2694, 2595, 3900, 3116, 3012, 2448, 2580, 3334, 3658, 3082, 4245, 2419, 2291, 2019, 2060, 1424, 3642, 3095, 2644, 4063, 4677, 4144, 3035, 3078, 2299, 3988, 2584, 3461, 2501, 3945, 3364, 3777, 4627, 2947, 3468, 2752, 3683, 3719, 3869, 3839, 3670, 3025, 3881, 2755, 3984, 3156, 2906, 3739, 2512, 3832, 3884, 2724, 2544, 2981, 2167, 4049, 4412, 3262, 2129, 3312, 3298, 2783, 3364, 2648, 3441, 3623, 3561, 4197, 3513, 3104, 2788, 2645, 3689, 3041, 3292, 2598, 4888, 3058, 3247, 2879, 2881, 3627, 3246, 5128, 2906, 3508, 4363, 2441, 3251, 3862, 4574, 3496, 3500, 3551, 3889, 3178, 4238]
line [131, 156, 131, 166, 179, 235, 267, 233, 288, 451, 390, 368, 368, 503, 533, 579, 517, 532, 703, 589, 876, 795, 772, 529, 366, 722, 865, 1067, 900, 1310, 1642, 1025, 1650, 1352, 1012, 1829, 1920, 1772, 1848, 1536, 2147, 1660, 1557, 2892, 2928, 1843, 1845, 561, 2160, 1902, 1821, 848, 1271, 1161, 2022, 2273, 1965, 1789, 2518, 2755, 2199, 2337, 2168, 1082, 2050, 1037, 1830, 1897, 898, 2368, 1192, 1968, 1147, 1803, 1617, 1800, 1340, 1944, 1794, 1214, 2017, 1993, 2478, 1915, 1164, 1503, 1479, 1486, 1300, 2018, 2074, 1579, 1267, 1917, 1627, 1844, 2384, 1827, 2388]

A58 - ArmV7

---
config:
    xyChart:
        width: 900
        height: 600
    themeVariables:
        xyChart:
            plotColorPalette: "#28678D,#f34544"
---

xychart-beta
title "读者操作"
x-axis "读者线程数" 2 --> 100
y-axis "操作消耗平均时间（ms）"
line [2457, 2534, 3098, 3555, 2603, 2752, 2182, 2549, 2008, 2224, 2276, 1920, 1883, 1972, 1828, 1795, 1914, 2013, 1934, 2692, 2190, 2667, 2439, 2265, 2556, 2413, 2317, 2288, 2327, 2241, 2454, 2293, 2100, 2433, 2223, 2191, 2349, 2108, 1859, 2288, 2270, 2163, 2177, 2029, 2106, 2020, 1959, 2258, 2204, 2069, 2149, 1969, 1918, 2047, 2216, 2050, 1946, 1941, 1891, 1914, 1931, 2124, 1935, 1898, 1897, 1784, 2034, 2005, 1888, 1893, 1829, 1808, 1923, 1952, 1997, 1970, 1887, 1859, 1891, 1899, 1772, 1907, 1841, 1887, 1949, 1854, 1823, 1932, 1881, 1877, 1796, 1814, 1909, 1927, 1832, 1814, 1886, 1844, 1819]
line [201, 183, 209, 260, 301, 371, 361, 382, 527, 588, 554, 514, 566, 634, 652, 629, 576, 629, 659, 715, 677, 868, 720, 779, 830, 933, 935, 959, 890, 1217, 929, 967, 1002, 1021, 1084, 1064, 1367, 1169, 1184, 1428, 1223, 1317, 1304, 1385, 1373, 1481, 1419, 1418, 1439, 1417, 1442, 1423, 1403, 1504, 1431, 1470, 1540, 1548, 1547, 1604, 1574, 1566, 1568, 1618, 1824, 2018, 1733, 1707, 1697, 1697, 1760, 1753, 1828, 1729, 1817, 1800, 1856, 1910, 1904, 1821, 1825, 1844, 1920, 1916, 1872, 1988, 1950, 1982, 1951, 2006, 1917, 1983, 2003, 1928, 1916, 2062, 2012, 2053, 2143]

---
config:
    xyChart:
        width: 900
        height: 600
    themeVariables:
        xyChart:
            plotColorPalette: "#28678D,#f34544"
---

xychart-beta
title "写者操作"
x-axis "读者线程数" 2 --> 100
y-axis "操作消耗平均时间（ms）"
line [3789, 3695, 5413, 11001, 9174, 10076, 5790, 8715, 3593, 5966, 6616, 3933, 3498, 5221, 3953, 5069, 7497, 5400, 5790, 9477, 5003, 10011, 11999, 10363, 12755, 12189, 11128, 10552, 13025, 11198, 10668, 8748, 7696, 14514, 13213, 9938, 8251, 8354, 5258, 9767, 10209, 9887, 10354, 10827, 8651, 9789, 8060, 10968, 13114, 8849, 10323, 7687, 7386, 10777, 12034, 9359, 8319, 7868, 6982, 7863, 9990, 11183, 10215, 8532, 8367, 5561, 8930, 11887, 8338, 8045, 4703, 6089, 10610, 6788, 7118, 8658, 7816, 5522, 8628, 8273, 7855, 10658, 6850, 6248, 8997, 9190, 9267, 7038, 11003, 9391, 8637, 6109, 7737, 9697, 5912, 8634, 8606, 8555, 7449]
line [251, 182, 215, 259, 301, 354, 354, 386, 480, 539, 581, 560, 611, 738, 676, 714, 669, 703, 743, 716, 673, 900, 918, 841, 923, 1004, 1006, 1054, 923, 1318, 1032, 1095, 1001, 1151, 1191, 1182, 1512, 1160, 1245, 1279, 1345, 1339, 1279, 1362, 1116, 1647, 1440, 1445, 1323, 1564, 1175, 1463, 1533, 1489, 1585, 1443, 1648, 1823, 1730, 1697, 1722, 1821, 1672, 1745, 1981, 1666, 1670, 2159, 1765, 2021, 1662, 1841, 2027, 1890, 1936, 1945, 2045, 1962, 1924, 1978, 2029, 1983, 2056, 2403, 2088, 2114, 2196, 2343, 2279, 2235, 2022, 2231, 2113, 2029, 1814, 2382, 2193, 2182, 1997]

总结

综合来看，随着读者线程数的上升，写者操作在一定程度上会保持着一个良好的状态，这个在 Arm 上的表现尤为明显，但是读者操作在不同的架构下均会在到达一定数量时的平均耗时会超过内核调度

总结

不难看出，原子操作更加适用于快速响应的轻量线程任务管理中，对于高并发来说，采用内核进行管理是最佳的选择，毕竟内核中会有一个托管队列用于管理对应的互斥锁。同时，在不同的架构下，线程数对原子操作所产生的影响是不同的，这个需要根据具体的场景进行测试才能做出妥善的选择

ZYXeeker`s blog

原子读写锁 —— 测试

实现

统一封装

实现基础测试用例

实现单元测试

测试

X86

A58 - ArmV7

总结

总结

下一篇：原子读写锁 —— 实现