为什么在取消大量HTTP请求时取消会阻塞这么长时间？

Question

背景

我有一些代码使用来自一个特定主机的内容执行批量HTML页面处理.它尝试使用大量(~400)同时发出的HTTP请求HttpClient.我认为同时连接的最大数量受到限制ServicePointManager.DefaultConnectionLimit,所以我没有应用自己的并发限制.

在将所有请求异步发送到HttpClient使用后Task.WhenAll,可以使用CancellationTokenSource和取消整个批处理操作CancellationToken.可以通过用户界面查看操作的进度,并且可以单击按钮以执行取消.

问题

对CancellationTokenSource.Cancel()块的调用大约需要5到30秒.这会导致用户界面冻结.怀疑是因为该方法正在调用注册取消通知的代码.

我考虑过的

1. 限制同时HTTP请求任务的数量.我认为这是一种解决方法,因为HttpClient似乎已经排除了过多的请求本身.
2. CancellationTokenSource.Cancel()在非UI线程中执行方法调用.这不太好用; 在大多数其他人完成任务之前,任务实际上并未运行.我认为async该方法的一个版本可以很好地工作,但我找不到一个.另外,我的印象是它适合在UI线程中使用该方法.

示范

码

class Program
{
    private const int desiredNumberOfConnections = 418;

    static void Main(string[] args)
    {
        ManyHttpRequestsTest().Wait();

        Console.WriteLine("Finished.");
        Console.ReadKey();
    }

    private static async Task ManyHttpRequestsTest()
    {
        using (var client = new HttpClient())
        using (var cancellationTokenSource = new CancellationTokenSource())
        {
            var requestsCompleted = 0;

            using (var allRequestsStarted = new CountdownEvent(desiredNumberOfConnections))
            {
                Action reportRequestStarted = () => allRequestsStarted.Signal();
                Action reportRequestCompleted = () => Interlocked.Increment(ref requestsCompleted);
                Func<int, Task> getHttpResponse = index => GetHttpResponse(client, cancellationTokenSource.Token, reportRequestStarted, reportRequestCompleted);
                var httpRequestTasks = Enumerable.Range(0, desiredNumberOfConnections).Select(getHttpResponse);

                Console.WriteLine("HTTP requests batch being initiated");
                var httpRequestsTask = Task.WhenAll(httpRequestTasks);

                Console.WriteLine("Starting {0} requests (simultaneous connection limit of {1})", desiredNumberOfConnections, ServicePointManager.DefaultConnectionLimit);
                allRequestsStarted.Wait();

                Cancel(cancellationTokenSource);
                await WaitForRequestsToFinish(httpRequestsTask);
            }

            Console.WriteLine("{0} HTTP requests were completed", requestsCompleted);
        }
    }

    private static void Cancel(CancellationTokenSource cancellationTokenSource)
    {
        Console.Write("Cancelling...");

        var stopwatch = Stopwatch.StartNew();
        cancellationTokenSource.Cancel();
        stopwatch.Stop();

        Console.WriteLine("took {0} seconds", stopwatch.Elapsed.TotalSeconds);
    }

    private static async Task WaitForRequestsToFinish(Task httpRequestsTask)
    {
        Console.WriteLine("Waiting for HTTP requests to finish");

        try
        {
            await httpRequestsTask;
        }
        catch (OperationCanceledException)
        {
            Console.WriteLine("HTTP requests were cancelled");
        }
    }

    private static async Task GetHttpResponse(HttpClient client, CancellationToken cancellationToken, Action reportStarted, Action reportFinished)
    {
        var getResponse = client.GetAsync("http://www.google.com", cancellationToken);

        reportStarted();
        using (var response = await getResponse)
            response.EnsureSuccessStatusCode();
        reportFinished();
    }
}

产量

为什么取消阻止这么久？还有,我做错了什么或者做得更好？

Answer 1

在非UI线程中执行CancellationTokenSource.Cancel()方法调用.这不太好用; 在大多数其他人完成任务之前,任务实际上并未运行.

这告诉我的是,你可能正在遭受'线程池耗尽',这是你的线程池队列中有很多项目(来自HTTP请求完成),需要一段时间来完成它们.取消可能阻止某些线程池工作项执行,它不能跳到队列的头部.

这表明您确实需要从您的考虑清单中选择选项1.限制自己的工作,以便线程池队列保持相对较短.无论如何,这对于整体应用响应非常有用.

我最喜欢的限制异步工作的方法是使用Dataflow.像这样的东西:

var block = new ActionBlock<Uri>(
    async uri => {
        var httpClient = new HttpClient(); // HttpClient isn't thread-safe, so protect against concurrency by using a dedicated instance for each request.
        var result = await httpClient.GetAsync(uri);
        // do more stuff with result.
    },
    new ExecutionDataflowBlockOptions { MaxDegreeOfParallelism = 20, CancellationToken = cancellationToken });
for (int i = 0; i < 1000; i++)
    block.Post(new Uri("http://www.server.com/req" + i));
block.Complete();
await block.Completion; // waits until everything is done or canceled.

作为替代方案,您可以使用在TaskCreationOptions.LongRunning中传递的Task.Factory.StartNew,以便您的任务获得一个新线程(不隶属于线程池),这将允许它立即启动并从那里调用Cancel.但你应该解决线程池耗尽问题.