[英]throttling an await delay for inbound messages to a number of messages per second
我試圖將循環(正在發送消息)限制為每秒特定數量的消息。 _throttle
是每秒的消息數。
我的初始算法如下所示,但是延遲並不平穩。
我可以做些什么改進來消除頗為顛簸的延遲和突發消息。
我已經玩過滴答聲和最大間隔,但是入站數量如此之大,難以彌補。 在我的實現中,關閉油門后可以達到的最大速率約為15000 /秒。 我正在以每秒300到1000之間的速率進行測試,因此我試圖將其速度放慢一些。
private class ThrottleCalculator
{
private readonly int _throttle;
private DateTime _lastCalculation = DateTime.Now;
private int _count = 0;
private int _interval = 0;
public ThrottleCalculator(int throttle)
{
this._throttle = throttle;
}
public async Task CalculateThrottle()
{
this._count += 1;
var elapsed = DateTime.Now.Subtract(this._lastCalculation).TotalMilliseconds;
var tick = 50;
if (elapsed > tick)
{
this._lastCalculation = DateTime.Now;
int projection = this._count * (1000 / tick);
var errorTerm = this._throttle - projection;
this._interval = this._interval - errorTerm;
if (this._interval < 0)
this._interval = 0;
// this is often several thousand, so I have to limit.
if (this._interval > 100)
this._interval = 100;
await Task.Delay(this._interval);
this._count = 0;
}
}
}
使用此代碼的代碼每次迭代都調用此代碼。
var throttle = new ThrottleCalculator(600); // 600/s
while (message = getMessage())
{
... // do stuff with message.
if (throttle != null)
await throttle.CalculateThrottle();
對於其他嘗試這樣做的人,正確的方法是PID控制器算法 。
Proportional / Integral / Derivative Controller
我以Wiki底部的算法為基礎。 我的kp / ki / kd
似乎可以很好地與此處的值配合使用,將它們保持成比例似乎可以產生穩定的消息流以及非常緊湊的延遲值。
private class ThrottleCalculator
{
private readonly int _throttle;
private DateTime _lastCalculationTime;
private double _measured = 0;
private double _totalError = 0;
private double _integral = 0;
private double _lastError = 0;
public ThrottleCalculator(int throttle)
{
this._throttle = throttle;
this._lastCalculationTime = DateTime.MinValue;
}
public async Task CalculateThrottle()
{
var kp = -.1d; // proportional gain
var ki = -.1d; // integral gain
var kd = -.1d; // derivative gain
var dt = 30d; // rate of change of time. calculcations every ms;
this._measured += 1;
if (this._lastCalculationTime == DateTime.MinValue)
this._lastCalculationTime = DateTime.Now;
var elapsed = (double)DateTime.Now.Subtract(this._lastCalculationTime)
.TotalMilliseconds;
if (elapsed > dt)
{
this._lastCalculationTime = DateTime.Now;
var error = ((double)this._throttle / (1000d / dt)) - this._measured;
this._totalError += error;
var integral = this._totalError;
var derivative = (error - this._lastError) / elapsed;
var actual = (kp * error) + (ki * integral) + (kd * derivative);
var output = actual;
if (output < 1)
output = 0;
// i don't like this, but it seems necessary
// so that wild wait values are never used.
if (output > dt * 4)
output = dt * 4;
if (output > 0)
await Task.Delay((int)output);
this._measured = 0;
this._lastError = error;
}
}
}
我的價值觀是這樣的:
Actual: 19.2000 Output: 19.2000 Integral: -209 Derivative: .0000 Error: 17
Actual: 17.5000 Output: 17.5000 Integral: -192 Derivative: .0000 Error: 17
Actual: 15.8000 Output: 15.8000 Integral: -175 Derivative: .0000 Error: 17
Actual: 33.8104 Output: 33.8104 Integral: -255 Derivative: -3.1040 Error: -80
Actual: 21.8931 Output: 21.8931 Integral: -238 Derivative: 2.0686 Error: 17
Actual: 20.4000 Output: 20.4000 Integral: -221 Derivative: .0000 Error: 17
Actual: 18.7000 Output: 18.7000 Integral: -204 Derivative: .0000 Error: 17
Actual: 17.0000 Output: 17.0000 Integral: -187 Derivative: .0000 Error: 17
Actual: 15.3000 Output: 15.3000 Integral: -170 Derivative: .0000 Error: 17
Actual: 31.0752 Output: 31.0752 Integral: -239 Derivative: -2.7520 Error: -69
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.