简体   繁体   English

使用大量“锁定”重构代码以获得更多无锁代码

[英]Refactoring code with a lot of “locks” to more lock-free code

Upd thanks to Matthew Watson for noticing and note that I plan to port my code to c++-linux so I prefer "platform-independent" code 更多感谢Matthew Watson注意到并注意到我打算将我的代码移植到c ++ - linux所以我更喜欢“平台无关”的代码

My trading application is almost lock-free. 我的交易应用程序几乎无锁。 The code below is the only place where I do use locks. 下面的代码是我唯一使用锁的地方。 Let me start with the code, it's pretty long but don't worry there are a lot of repeating parts so it's simple. 让我从代码开始,它很长但不要担心有很多重复部分,所以它很简单。 I prefer to add all "repeating" parts to better demonstrate how my things work: 我更喜欢添加所有“重复”部分,以更好地展示我的工作方式:

Task.Factory.StartNew(() =>
{
    while (true)
    {
        Iterate();
    }
}, TaskCreationOptions.LongRunning);

private void Iterate()
{
    bool marketDataUpdated = false;

    lock (ordersToRegisterLock)
    {
        if (ordersToRegister.Count > 0)
        {
            marketDataUpdated = true;
            while (ordersToRegister.Count > 0)
            {
                Order order = ordersToRegister.Dequeue();
                // Stage1, Process
            }
        }
    }

    lock (aggrUpdatesLock)
    {
        if (aggrUpdates.Count > 0)
        {
            marketDataUpdated = true;
            while (!aggrUpdates.IsNullOrEmpty())
            {
                var entry = aggrUpdates.Dequeue();
                // Stage1, Process
            }
        }
    }

    lock (commonUpdatesLock)
    {
        if (commonUpdates.Count > 0)
        {
            marketDataUpdated = true;
            while (!commonUpdates.IsNullOrEmpty())
            {
                var entry = commonUpdates.Dequeue();
                // Stage1, Process
            }
        }
    }

    lock (infoUpdatesLock)
    {
        if (infoUpdates.Count > 0)
        {
            marketDataUpdated = true;
            while (!infoUpdates.IsNullOrEmpty())
            {
                var entry = infoUpdates.Dequeue();
                // Stage1, Process
            }
        }
    }

    lock (tradeUpdatesLock)
    {
        if (tradeUpdates.Count > 0)
        {
            marketDataUpdated = true;
            while (!tradeUpdates.IsNullOrEmpty())
            {
                var entry = tradeUpdates.Dequeue();
                // Stage1, Process
            }    

        }
    }

    if (marketDataUpdated)
    {
        // Stage2 !
        // make a lot of work. expensive operation. recalculate strategies, place orders etc.
    }
}

private readonly Queue<Order> ordersToRegister = new Queue<Order>();
private readonly object ordersToRegisterLock = new object();

private readonly Queue<AggrEntry> aggrUpdates = new Queue<AggrEntry>();
private readonly object aggrUpdatesLock = new object();

private readonly Queue<CommonEntry> commonUpdates = new Queue<CommonEntry>();
private readonly object commonUpdatesLock = new object();

private readonly Queue<InfoEntry> infoUpdates = new Queue<InfoEntry>();
private readonly object infoUpdatesLock = new object();

private readonly Queue<TradeEntry> tradeUpdates = new Queue<TradeEntry>();
private readonly object tradeUpdatesLock = new object();


    public void RegistorOrder(object sender, Gate.RegisterOrderArgs e)
    {
        lock (ordersToRegisterLock)
        {
            ordersToRegister.Enqueue(e.order);
        }
    }

    public void TradeUpdated(object sender, Gate.TradeArgs e)
    {
        lock (tradeUpdatesLock)
        {
            foreach (var entry in e.entries)
            {
                tradeUpdates.Enqueue(entry);
            }
        }
    }

    public void InfoUpdated(object sender, Gate.InfoArgs e)
    {
        lock (infoUpdatesLock)
        {
            foreach (var entry in e.entries)
            {
                infoUpdates.Enqueue(entry);
            }
        }
    }

    public void CommonUpdated(object sender, Gate.CommonArgs e)
    {
        lock (commonUpdatesLock)
        {
            foreach (var entry in e.entries)
            {
                commonUpdates.Enqueue(entry);
            }
        }
    }

    public void AggrUpdated(object sender, Gate.AggrArgs e)
    {
        lock (aggrUpdatesLock)
        {
            foreach (var entry in e.entries)
            {
                aggrUpdates.Enqueue(entry);
            }
        }
    }

In my code I have two stages. 在我的代码中,我有两个阶段。 Stage1 is update stage and Stage2 is working stage. Stage1是更新阶段, Stage2处于工作阶段。 I need to switch between these two stages as fast as possible, like that: 我需要尽快在这两个阶段之间切换,如下所示:

  • any updates? 任何更新? no 没有
  • any updates? 任何更新? no 没有
  • any updated? 任何更新? yes, order updated! 是的,订单更新! apply update, do Stage2 应用更新,做Stage2
  • any updates? 任何更新? no 没有
  • any updates? 任何更新? yes, order need to be registered! 是的,订单需要注册! apply update, do Stage2 应用更新,做Stage2
  • any updates? 任何更新? yes, trade occured, apply update, do Stage2 交易发生,应用更新,做Stage2

In Stage2 I should not update, but should keep "collecting" updates so I can apply they later. Stage2我不应该更新,但应该保持“收集”更新,以便我可以在以后应用它们。

And important thing - this is very latency-critical code so I agree to "spent" one core for having minimal latency! 重要的是 - 这是非常关键的延迟代码,因此我同意“花费”一个核心以获得最小的延迟! So when any update occure I need to process it asap and perform Stage2 . 因此,当发生任何更新时,我需要尽快处理它并执行Stage2

So I hope now it's clear what I need to achieve and it's clear how I have implemented that. 所以我希望现在很清楚我需要实现什么,而且很清楚我是如何实现的。 Now it's time to discuss how good my code is. 现在是讨论我的代码有多好的时候了。 I do see several potential problems: 我确实看到了几个潜在的问题:

  • a lot of locks! 很多锁! can it be replaced with some "lock-free" code? 可以用一些“无锁”代码替换它吗? spinlock with CAS or something? 用CAS还是什么螺旋锁?
  • occupy 100% of CPU core, can I save some CPU resources without affecting latency ? 占用100%的CPU核心,可以节省一些CPU资源而不影响延迟吗?
  • can/should I tell .NET to use "dedicated" core (set task affinity?) to avoid extra "switch"? 可以/我应该告诉.NET使用“专用”核心(设置任务亲和力?)来避免额外的“切换”?
  • I add to Queues from one thread and I read Queues from another thread. 我从一个线程添加到队列,我从另一个线程读取队列。 Could it be a problem? 这可能是个问题吗? If adding and reading to a Queue is volatile? 如果添加和读取队列是不稳定的? Is it possible that my reading thread will not see update from a Queue because cache-update problem? 我的阅读线程是否可能无法从队列中看到更新,因为缓存更新问题?

Any suggestions how to improve what I wrote are welcome, thanks! 欢迎提出任何有关如何改进我写作的建议,谢谢!

upd partly solved - as I understand I better to replace queries to lock-free (likely ring-buffer based?) queries.. i think i will use c++ version of disruptor later. upd部分解决了 - 据我所知,我最好将查询替换为无锁(可能是基于环缓冲的?)查询。我想我稍后会使用c ++版本的disruptor。 Also I've used this article http://www.umbraworks.net/bl0g/rebuildall/2010/03/08/Running_NET_threads_on_selected_processor_cores and replaced Task with a Thread running on the "fixed" core, however i'm still using "busy-spin", probably I should use something smarter? 我也使用了这篇文章http://www.umbraworks.net/bl0g/rebuildall/2010/03/08/Running_NET_threads_on_selected_processor_cores并用“固定”核心上运行的线程替换了任务,但是我还在使用“忙” -spin“,可能我应该使用更聪明的东西?

With the code below, you are no longer locked during "stage 1" processing: 使用下面的代码,您在“阶段1”处理期间不再被锁定:

Task.Factory.StartNew(() =>
{
    while (true)
    {
        Iterate();
    }
}, TaskCreationOptions.LongRunning);


private void Iterate()
{
    bool marketDataUpdated = false;

    foreach (Order order in ordersToRegister)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    foreach (var entry in aggrUpdates)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    foreach (var entry in commonUpdates)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    foreach (var entry in infoUpdates)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    foreach (var entry in tradeUpdates)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    if (marketDataUpdated)
    {
        // Stage2 !
        // make a lot of work. expensive operation. recalculate strategies, place orders etc.
    }
}

private readonly ConcurrentQueue<Order> ordersToRegister = new ConcurrentQueue<Order>();

private readonly ConcurrentQueue<AggrEntry> aggrUpdates = new ConcurrentQueue<AggrEntry>();

private readonly ConcurrentQueue<CommonEntry> commonUpdates = new ConcurrentQueue<CommonEntry>();

private readonly ConcurrentQueue<InfoEntry> infoUpdates = new ConcurrentQueue<InfoEntry>();

private readonly ConcurrentQueue<TradeEntry> tradeUpdates = new ConcurrentQueue<TradeEntry>();

    public void RegistorOrder(object sender, Gate.RegisterOrderArgs e)
    {
        ordersToRegister.Enqueue(e.order);
    }

    public void TradeUpdated(object sender, Gate.TradeArgs e)
    {
        foreach (var entry in e.entries)
        {
            tradeUpdates.Enqueue(entry);
        }
    }

    public void InfoUpdated(object sender, Gate.InfoArgs e)
    {
        foreach (var entry in e.entries)
        {
            infoUpdates.Enqueue(entry);
        }
    }

    public void CommonUpdated(object sender, Gate.CommonArgs e)
    {
        foreach (var entry in e.entries)
        {
            commonUpdates.Enqueue(entry);
        }
    }

    public void AggrUpdated(object sender, Gate.AggrArgs e)
    {
        foreach (var entry in e.entries)
        {
            aggrUpdates.Enqueue(entry);
        }
    }

Here is an approach that might be more portable. 这是一种可能更便携的方法。 Hope it helps. 希望能帮助到你。

public class SafeQueue<T> : Queue<T>
{
    public T SafeDequeue()
    {
        lock (this)
        {
            return (Count > 0) ? Dequeue() : null;
        }
    }

    public void SafeEnqueue(T entry)
    {
        lock (this)
        {
            Enqueue(entry);
        }
    }
}

Task.Factory.StartNew(() =>
{
    while (true)
    {
        Iterate();
    }
}, TaskCreationOptions.LongRunning);


private void Iterate()
{
    bool marketDataUpdated = false;

    while ((Order order = ordersToRegister.SafeDequeue()) != null)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    while ((var entry = aggrUpdates.SafeDequeue()) != null)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    while ((var entry = commonUpdates.SafeDequeue()) != null)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    while ((var entry = infoUpdates.SafeDequeue()) != null)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    while ((var entry = tradeUpdates.SafeDequeue()) != null)
    {
        marketDataUpdated = true;
        // Stage1, Process
    }

    if (marketDataUpdated)
    {
        // Stage2 !
        // make a lot of work. expensive operation. recalculate strategies, place orders etc.
    }
}

private readonly SafeQueue<Order> ordersToRegister = new SafeQueue<Order>();

private readonly SafeQueue<AggrEntry> aggrUpdates = new SafeQueue<AggrEntry>();

private readonly SafeQueue<CommonEntry> commonUpdates = new SafeQueue<CommonEntry>();

private readonly SafeQueue<InfoEntry> infoUpdates = new SafeQueue<InfoEntry>();

private readonly SafeQueue<TradeEntry> tradeUpdates = new SafeQueue<TradeEntry>();

    public void RegistorOrder(object sender, Gate.RegisterOrderArgs e)
    {
        ordersToRegister.SafeEnqueue(e.order);
    }

    public void TradeUpdated(object sender, Gate.TradeArgs e)
    {
        foreach (var entry in e.entries)
        {
            tradeUpdates.SafeEnqueue(entry);
        }
    }

    public void InfoUpdated(object sender, Gate.InfoArgs e)
    {
        foreach (var entry in e.entries)
        {
            infoUpdates.SafeEnqueue(entry);
        }
    }

    public void CommonUpdated(object sender, Gate.CommonArgs e)
    {
        foreach (var entry in e.entries)
        {
            commonUpdates.SafeEnqueue(entry);
        }
    }

    public void AggrUpdated(object sender, Gate.AggrArgs e)
    {
        foreach (var entry in e.entries)
        {
            aggrUpdates.SafeEnqueue(entry);
        }
    }

声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.

 
粤ICP备18138465号  © 2020-2024 STACKOOM.COM