在高性能的伺服器程式當中,定時器是必不可少的部件,而且定時器的效率是直接影響到服務的性能。在衆多的開源項目中,定時器設計都有各有各的方法,例如ACE和libEvent都采用了最小堆的算法實作,還有其他的開源項目采用平衡二叉樹來做定時的器管理算法。不管是最小堆還是平衡二叉樹,其定時器掃描都是O(1),但定時器插入和删除都是O(logN)的複雜度。在定時事件少的情況下,這種算法是足夠的,如果超過上百萬的定時事件,效率會成為瓶頸。是以revolver在定時器的實作上并沒有使用通用的平衡二叉樹和最小堆,而是采用了輪轉HASH算法來做定時器管理。
什麼是輪轉HASH算法?輪轉HASH是通過4個時間輪的轉動來觸發定時事件,就像時鐘的秒針輪、分針輪、時針輪之間的關系一樣。如圖:
![](https://img.laitimes.com/img/_0nNw4CM6IyYiwiM6ICdiwiI0gTMx81dsQWZ4lmZf1GLlpXazVmcvwFciV2dsQXYtJ3bm9CX9s2RkBnVHFmb1clWvB3MaVnRtp1XlBXe0xCMy81dvRWYoNHLwEzX5xCMx8FesU2cfdGLwMzX0xiRGZkRGZ0Xy9GbvNGLpZTY1EmMZVDUSFTU4VFRR9Fd4VGdsYTMfVmepNHLrJXYtJXZ0F2dvwVZnFWbp1zczV2YvJHctM3cv1Ce-cGcq5SO1UjMycDZjZWN5ATZyMmMzYzXzMjM0cTMyIzLclDMyIDMy8CXn9Gbi9CXzV2Zh1WavwVbvNmLvR3YxUjLyM3Lc9CX6MHc0RHaiojIsJye.jpg)
每個輪的有256個刻度,4個輪剛好是一個uint32_t整型數。最小輪的一刻度表示一個reactor的event loop時間(5ms)
1.定時器的掃描
1)擷取目前的系統時間cur_ts,判斷與上次掃描時刻prev_ts之間差距超過最小刻度,計算需要轉動最小刻度的數量
scale = (cur_ts - prev_ts) / loop_delay;
2) 進行第一輪的輪轉,轉動scale個刻度,所有轉動的刻度中的定時事件全部進行觸發。如果轉動後沒有超過最大刻度256,輪轉結束。
3)如果超過最大刻度256,那麼從它的父輪(第二輪)上輪轉1刻度。将父輪上輪轉的刻度所有的定時事件全部按時間映射到子輪上。相當于重複2步。這是個疊代過程,如果父輪到了256,那麼繼續輪轉父輪的父輪。直到到達第四輪轉動為止。這和秒鐘走一圈就是分針走一個刻度的原理是一樣的。
代碼如下:
template<class HANDLER, class FUNCTOR, class LOCK>
uint32_t CTimerQueue_T<HANDLER, FUNCTOR, LOCK>::expire()
{
BASE_GUARD_RETURN(LOCK, cf_mon, mutex_, 0);
uint32_t ret = SELECT_DELAY; //預設20MS
CBaseTimeValue cur_timer = CBaseTimeValue::get_time_value();
if(cur_timer > prev_time_)
{
uint32_t scale = static_cast<uint32_t>((cur_timer.msec() - prev_time_.msec()) / SELECT_DELAY);
if(scale > 0)
{
ret = revolver(scale);
prev_time_ = cur_timer;
}
}
return ret;
}
template<class HANDLER, class FUNCTOR, class LOCK>
uint32_t CTimerQueue_T<HANDLER, FUNCTOR, LOCK>::revolver(uint32_t scale)
{
//std::cout << "pos, first = " << rings_[0].get_pos() << ", second = " << rings_[1].get_pos()
// << ", third = " << rings_[2].get_pos() << ", fourth = " << rings_[3].get_pos() <<std::endl;
uint32_t ret = SELECT_DELAY;
uint8_t index = 0;
uint32_t rewind_scale = scale;
while(rewind_scale > 0)
{
index = 0;
if(rings_[index].cycle(rewind_scale, this)) //掃描第一輪
{
index ++;
uint32_t sc = 1;
while(rings_[index].cycle(sc, this))//掃描下一輪,刻度隻往前推進1格
{
sc = 1;
index ++;
if(index >= RINGS_SIZE)
{
start_time_ = CBaseTimeValue::get_time_value();
break;
}
}
}
}
return ret;
}
2.定時事件的插入
1)首先計算需觸發的時刻與定時器輪目前時刻的距離
d = (cur_ts - start_ts + delay) / loop_delay;
2)分别計算在4個輪的位置
first = (uint8_t)(timeout_stamp_ / FIRST_ROUND);
second = (uint8_t)((timeout_stamp_ % FIRST_ROUND) / SECOND_ROUND);
third = (uint8_t)((timeout_stamp_ % SECOND_ROUND) / THIRD_ROUND);
fourth = (uint8_t) (timeout_stamp_ % THIRD_ROUND);
3)通過計算得到的為止,儲存到對應的輪刻度上,插入就完畢了。例子,如果計算的first = 0, second = 2, third = 30, fouth = 1,就會儲存到第3輪的第2刻度上。
代碼如下:
template<class HANDLER, class FUNCTOR, class LOCK>
uint32_t CTimerQueue_T<HANDLER, FUNCTOR, LOCK>::schedule(HANDLER handler, const void *act, uint32_t delay, uint32_t interval)
{
BASE_GUARD_RETURN(LOCK, cf_mon, mutex_, 0);
BaseTimerNode_T<HANDLER>* timer_obj = node_pool_.pop_obj();
if(timer_obj != NULL)
{
uint32_t timer_id = get_free_node();
CBaseTimeValue cur_timer = CBaseTimeValue::get_time_value();
//計算距離
uint64_t distance = delay / SELECT_DELAY; //直接以目前時間作為坐标,相差一個掃描間隔20MS
if(cur_timer > start_time_)
distance = (cur_timer.msec() - start_time_.msec() + delay) / SELECT_DELAY;
distance = distance % (UNINT32_MAX);
timer_obj->set(handler, act, (uint32_t)(core_max(distance, 1)), interval, timer_id);
heap_[timer_id] = timer_obj;
used_num_ ++;
//插入事件
insert_node(timer_obj);
upcall_functor().registration(timer_obj->get_handler(), timer_id);
return timer_id;
}
return 0;
}
template<class HANDLER, class FUNCTOR, class LOCK>
void CTimerQueue_T<HANDLER, FUNCTOR, LOCK>::insert_node(BaseTimerNode_T<HANDLER>* node)
{
uint32_t timer_id = node->get_timer_id();
uint8_t poss[RINGS_SIZE] = {0};
//擷取位置
node->get_revolver_pos(poss[RINGS_SIZE - 1], poss[RINGS_SIZE - 2], poss[RINGS_SIZE - 3], poss[RINGS_SIZE - 4]);
uint8_t index = RINGS_SIZE - 1;
//進行插入
while(!rings_[index].add_element(poss[index], timer_id))
{
if(index == 0)
break ;
index --;
}
}
3.定時事件的删除
删除和插入計算差不多步驟
1)首先計算需要删除定時事件觸發的時刻與定時器輪目前時刻的距離
d = (cur_ts - start_ts + delay) / loop_delay;
2)分别計算在4個輪的位置
first = (uint8_t)(timeout_stamp_ / FIRST_ROUND);
second = (uint8_t)((timeout_stamp_ % FIRST_ROUND) / SECOND_ROUND);
third = (uint8_t)((timeout_stamp_ % SECOND_ROUND) / THIRD_ROUND);
fourth = (uint8_t) (timeout_stamp_ % THIRD_ROUND);
3)根據位置坐标找到對應的輪位置删除輪上的定時事件
代碼:
template<class HANDLER, class FUNCTOR, class LOCK>
void CTimerQueue_T<HANDLER, FUNCTOR, LOCK>::cancel_timer(uint32_t timer_id, const void **act)
{
BASE_GUARD(LOCK, cf_mon, mutex_);
if(timer_id < heap_size_ && heap_[timer_id] != NULL)
{
//查找對應的定時事件内容
BaseTimerNode_T<HANDLER>* timer_obj = heap_[timer_id];
//删除輪上的定時事件
delete_node(timer_obj);
heap_[timer_id] = NULL;
if(used_num_ > 0)
used_num_ --;
freeTimers_.push_back(timer_id);
*act = timer_obj->get_act();
upcall_functor().cancel_timer(timer_obj->get_handler(), timer_id);
node_pool_.push_obj(timer_obj);
}
}
template<class HANDLER, class FUNCTOR, class LOCK>
void CTimerQueue_T<HANDLER, FUNCTOR, LOCK>::delete_node(BaseTimerNode_T<HANDLER>* node)
{
uint32_t timer_id = node->get_timer_id();
uint8_t poss[RINGS_SIZE] = {0};
node->get_revolver_pos(poss[RINGS_SIZE - 1], poss[RINGS_SIZE - 2], poss[RINGS_SIZE - 3], poss[RINGS_SIZE - 4]);
//删除掉對應的定時事件
for(uint8_t index = 0; index < RINGS_SIZE; index ++) //在每個輪上進行删除
{
rings_[index].delete_element(poss[index], timer_id);
}
}
4.測試
在revovler的test工程中的main()函數中将test_timer_queue注釋去掉,就可以進行測試。以下是test_timer_queue的代碼:
void test_timer_queue()
{
srand(time(NULL));
CTimerFunctor functor;
TIMEQUEUE timer_queue(&functor);
CTest_Event_Handler handler;
handler.tq_ = &timer_queue;
CBaseTimeValue begin_timer = CBaseTimeValue::get_time_value();
for(int i = 0; i < 1000000; i ++)
{
insert_timer(&handler, (rand() % 240) * 1000, timer_queue);
}
CBaseTimeValue stop_timer = CBaseTimeValue::get_time_value();
stop_timer = stop_timer - begin_timer;
std::cout << "insert 1000000 timer, delay = " << stop_timer.msec() << " MS" << std::endl;
g_ts = stop_timer.get_time_value().msec();
#if _DEBUG
//timer_queue.set_ring_id();
#endif
std::cout << "exprie ......" << std::endl;
while(1)
{
uint32_t ms = timer_queue.expire();
usleep((1000));
}
}
這個函數可以測試插入100萬個定時事件的耗時多少,在100個定時事件在定時器管理的時候,CPU和記憶體都可以進行相對應的監控和檢視。我在window 7下面的release版本的資訊如下:
從上圖看,插入100萬個定時事件耗時978MS,也就是說,當有幾十萬個定時事件在運作的時候,插入一個定時事件隻需要0.97微秒。
以下是100萬個定時事件在處理過程中的CPU和記憶體占用圖。
5.總結
定時器的實作從效率和功能上實作都達到了最初設想的效率,但是記憶體使用上稍微過高(120M),如果用C的代碼實作,對記憶體做嚴格的控制和記憶體管理,應該會好很多,以後優化的工作應該重點在這。如果需要完整的代碼,請到:
https://github.com/yuanrongxi/revolver 下載下傳。