Develop Reference

C# Abortable Asynchronous Fifo Queue - leaking massive amounts of memory

I need to process data from a producer in FIFO fashion with the ability to abort processing if the same producer produces a new bit of data.
So I implemented an abortable FIFO queue based on Stephen Cleary's AsyncCollection (called AsyncCollectionAbortableFifoQueuein my sample) and one on TPL's BufferBlock (BufferBlockAbortableAsyncFifoQueue in my sample). Here's the implementation based on AsyncCollection
public class AsyncCollectionAbortableFifoQueue<T> : IExecutableAsyncFifoQueue<T>
{
private AsyncCollection<AsyncWorkItem<T>> taskQueue = new AsyncCollection<AsyncWorkItem<T>>();
private readonly CancellationToken stopProcessingToken;
public AsyncCollectionAbortableFifoQueue(CancellationToken cancelToken)
{
stopProcessingToken = cancelToken;
_ = processQueuedItems();
}
public Task<T> EnqueueTask(Func<Task<T>> action, CancellationToken? cancelToken)
{
var tcs = new TaskCompletionSource<T>();
var item = new AsyncWorkItem<T>(tcs, action, cancelToken);
taskQueue.Add(item);
return tcs.Task;
}
protected virtual async Task processQueuedItems()
{
while (!stopProcessingToken.IsCancellationRequested)
{
try
{
var item = await taskQueue.TakeAsync(stopProcessingToken).ConfigureAwait(false);
if (item.CancelToken.HasValue && item.CancelToken.Value.IsCancellationRequested)
item.TaskSource.SetCanceled();
else
{
try
{
T result = await item.Action().ConfigureAwait(false);
item.TaskSource.SetResult(result); // Indicate completion
}
catch (Exception ex)
{
if (ex is OperationCanceledException && ((OperationCanceledException)ex).CancellationToken == item.CancelToken)
item.TaskSource.SetCanceled();
item.TaskSource.SetException(ex);
}
}
}
catch (Exception) { }
}
}
}
public interface IExecutableAsyncFifoQueue<T>
{
Task<T> EnqueueTask(Func<Task<T>> action, CancellationToken? cancelToken);
}
processQueuedItems is the task that dequeues AsyncWorkItem's from the queue, and executes them unless cancellation has been requested.
The asynchronous action to execute gets wrapped into an AsyncWorkItem which looks like this
internal class AsyncWorkItem<T>
{
public readonly TaskCompletionSource<T> TaskSource;
public readonly Func<Task<T>> Action;
public readonly CancellationToken? CancelToken;
public AsyncWorkItem(TaskCompletionSource<T> taskSource, Func<Task<T>> action, CancellationToken? cancelToken)
{
TaskSource = taskSource;
Action = action;
CancelToken = cancelToken;
}
}
Then there's a task looking and dequeueing items for processing and either processing them, or aborting if the CancellationToken has been triggered.
That all works just fine - data gets processed, and if a new piece of data is received, processing of the old is aborted. My problem now stems from these Queues leaking massive amounts of memory if I crank up the usage (producer producing a lot more than the consumer processes). Given it's abortable, the data that is not processed, should be discarded and eventually disappear from memory.
So let's look at how I'm using these queues. I have a 1:1 match of producer and consumer. Every consumer handles data of a single producer. Whenever I get a new data item, and it doesn't match the previous one, I catch the queue for the given producer (User.UserId) or create a new one (the 'executor' in the code snippet). Then I have a ConcurrentDictionary that holds a CancellationTokenSource per producer/consumer combo. If there's a previous CancellationTokenSource, I call Cancel on it and Dispose it 20 seconds later (immediate disposal would cause exceptions in the queue). I then enqueue processing of the new data. The queue returns me a task that I can await so I know when processing of the data is complete, and I then return the result.
Here's that in code
internal class SimpleLeakyConsumer
{
private ConcurrentDictionary<string, IExecutableAsyncFifoQueue<bool>> groupStateChangeExecutors = new ConcurrentDictionary<string, IExecutableAsyncFifoQueue<bool>>();
private readonly ConcurrentDictionary<string, CancellationTokenSource> userStateChangeAborters = new ConcurrentDictionary<string, CancellationTokenSource>();
protected CancellationTokenSource serverShutDownSource;
private readonly int operationDuration = 1000;
internal SimpleLeakyConsumer(CancellationTokenSource serverShutDownSource, int operationDuration)
{
this.serverShutDownSource = serverShutDownSource;
this.operationDuration = operationDuration * 1000; // convert from seconds to milliseconds
}
internal async Task<bool> ProcessStateChange(string userId)
{
var executor = groupStateChangeExecutors.GetOrAdd(userId, new AsyncCollectionAbortableFifoQueue<bool>(serverShutDownSource.Token));
CancellationTokenSource oldSource = null;
using (var cancelSource = userStateChangeAborters.AddOrUpdate(userId, new CancellationTokenSource(), (key, existingValue) =>
{
oldSource = existingValue;
return new CancellationTokenSource();
}))
{
if (oldSource != null && !oldSource.IsCancellationRequested)
{
oldSource.Cancel();
_ = delayedDispose(oldSource);
}
try
{
var executionTask = executor.EnqueueTask(async () => { await Task.Delay(operationDuration, cancelSource.Token).ConfigureAwait(false); return true; }, cancelSource.Token);
var result = await executionTask.ConfigureAwait(false);
userStateChangeAborters.TryRemove(userId, out var aborter);
return result;
}
catch (Exception e)
{
if (e is TaskCanceledException || e is OperationCanceledException)
return true;
else
{
userStateChangeAborters.TryRemove(userId, out var aborter);
return false;
}
}
}
}
private async Task delayedDispose(CancellationTokenSource src)
{
try
{
await Task.Delay(20 * 1000).ConfigureAwait(false);
}
finally
{
try
{
src.Dispose();
}
catch (ObjectDisposedException) { }
}
}
}
In this sample implementation, all that is being done is wait, then return true.
To test this mechanism, I wrote the following Data producer class:
internal class SimpleProducer
{
//variables defining the test
readonly int nbOfusers = 10;
readonly int minimumDelayBetweenTest = 1; // seconds
readonly int maximumDelayBetweenTests = 6; // seconds
readonly int operationDuration = 3; // number of seconds an operation takes in the tester
private readonly Random rand;
private List<User> users;
private readonly SimpleLeakyConsumer consumer;
protected CancellationTokenSource serverShutDownSource, testAbortSource;
private CancellationToken internalToken = CancellationToken.None;
internal SimpleProducer()
{
rand = new Random();
testAbortSource = new CancellationTokenSource();
serverShutDownSource = new CancellationTokenSource();
generateTestObjects(nbOfusers, 0, false);
consumer = new SimpleLeakyConsumer(serverShutDownSource, operationDuration);
}
internal void StartTests()
{
if (internalToken == CancellationToken.None || internalToken.IsCancellationRequested)
{
internalToken = testAbortSource.Token;
foreach (var user in users)
_ = setNewUserPresence(internalToken, user);
}
}
internal void StopTests()
{
testAbortSource.Cancel();
try
{
testAbortSource.Dispose();
}
catch (ObjectDisposedException) { }
testAbortSource = new CancellationTokenSource();
}
internal void Shutdown()
{
serverShutDownSource.Cancel();
}
private async Task setNewUserPresence(CancellationToken token, User user)
{
while (!token.IsCancellationRequested)
{
var nextInterval = rand.Next(minimumDelayBetweenTest, maximumDelayBetweenTests);
try
{
await Task.Delay(nextInterval * 1000, testAbortSource.Token).ConfigureAwait(false);
}
catch (TaskCanceledException)
{
break;
}
//now randomly generate a new state and submit it to the tester class
UserState? status;
var nbStates = Enum.GetValues(typeof(UserState)).Length;
if (user.CurrentStatus == null)
{
var newInt = rand.Next(nbStates);
status = (UserState)newInt;
}
else
{
do
{
var newInt = rand.Next(nbStates);
status = (UserState)newInt;
}
while (status == user.CurrentStatus);
}
_ = sendUserStatus(user, status.Value);
}
}
private async Task sendUserStatus(User user, UserState status)
{
await consumer.ProcessStateChange(user.UserId).ConfigureAwait(false);
}
private void generateTestObjects(int nbUsers, int nbTeams, bool addAllUsersToTeams = false)
{
users = new List<User>();
for (int i = 0; i < nbUsers; i++)
{
var usr = new User
{
UserId = $"User_{i}",
Groups = new List<Team>()
};
users.Add(usr);
}
}
}
It uses the variables at the beginning of the class to control the test. You can define the number of users (nbOfusers - every user is a producer that produces new data), the minimum (minimumDelayBetweenTest) and maximum (maximumDelayBetweenTests) delay between a user producing the next data and how long it takes the consumer to process the data (operationDuration).
StartTests starts the actual test, and StopTests stops the tests again.
I'm calling these as follows
static void Main(string[] args)
{
var tester = new SimpleProducer();
Console.WriteLine("Test successfully started, type exit to stop");
string str;
do
{
str = Console.ReadLine();
if (str == "start")
tester.StartTests();
else if (str == "stop")
tester.StopTests();
}
while (str != "exit");
tester.Shutdown();
}
So, if I run my tester and type 'start', the Producer class starts producing states that are consumed by Consumer. And memory usage starts to grow and grow and grow. The sample is configured to the extreme, the real-life scenario I'm dealing with is less intensive, but one action of the producer could trigger multiple actions on the consumer side which also have to be executed in the same asynchronous abortable fifo fashion - so worst case, one set of data produced triggers an action for ~10 consumers (that last part I stripped out for brevity).
When I'm having a 100 producers, and each producer produces a new data item every 1-6 seconds (randomly, also the data produces is random). Consuming the data takes 3 seconds.. so there's plenty of cases where there's a new set of data before the old one has been properly processed.
Looking at two consecutive memory dumps, it's obvious where the memory usage is coming from.. it's all fragments that have to do with the queue. Given that I'm disposing every TaskCancellationSource and not keeping any references to the produced data (and the AsyncWorkItem they're put into), I'm at a loss to explain why this keeps eating up my memory and I'm hoping somebody else can show me the errors of my way. You can also abort testing by typing 'stop'.. you'll see that no longer is memory being eaten, but even if you pause and trigger GC, memory is not being freed either.
The source code of the project in runnable form is on Github. After starting it, you have to type start (plus enter) in the console to tell the producer to start producing data. And you can stop producing data by typing stop (plus enter)

Your code has so many issues making it impossible to find a leak through debugging. But here are several things that already are an issue and should be fixed first:
Looks like getQueue creates a new queue for the same user each time processUseStateUpdateAsync gets called and does not reuse existing queues:
var executor = groupStateChangeExecutors.GetOrAdd(user.UserId, getQueue());
CancellationTokenSource is leaking on each call of the code below, as new value created each time the method AddOrUpdate is called, it should not be passed there that way:
userStateChangeAborters.AddOrUpdate(user.UserId, new CancellationTokenSource(), (key, existingValue
Also code below should use the same cts as you pass as new cts, if dictionary has no value for specific user.UserId:
return new CancellationTokenSource();
Also there is a potential leak of cancelSource variable as it gets bound to a delegate which can live for a time longer than you want, it's better to pass concrete CancellationToken there:
executor.EnqueueTask(() => processUserStateUpdateAsync(user, state, previousState,
cancelSource.Token));
By some reason you do not dispose aborter here and in one more place:
userStateChangeAborters.TryRemove(user.UserId, out var aborter);
Creation of Channel can have potential leaks:
taskQueue = Channel.CreateBounded<AsyncWorkItem<T>>(new BoundedChannelOptions(1)
You picked option FullMode = BoundedChannelFullMode.DropOldest which should remove oldest values if there are any, so I assume that that stops queued items from processing as they would not be read. It's a hypotheses, but I assume that if an old item is removed without being handled, then processUserStateUpdateAsync won't get called and all resources won't be freed.
You can start with these found issues and it should be easier to find the real cause after that.

Non-blocking (lock-free) one-time initialization

Original question:
I need to initialize something only once in a multi-threaded application (when first thread enters the block). Subsequent threads should skip the initialization without waiting for it to complete.
I've found this blog entry Lock-free Thread Safe Initialisation in C# but it does not do exactly what I want, as it makes the other threads wait for the initialization to complete (if I understand it correctly).
Here's an example that presents the issue, although it does not work due to lack of synchronization:
using System;
using System.Collections.Concurrent;
using System.Threading.Tasks;
namespace LockFreeInitialization
{
public class Program
{
private readonly ConcurrentQueue<int> _jobsQueue = new ConcurrentQueue<int>();
private volatile bool _initialized;
private async Task EnqueueAndProcessJobsAsync(int taskId, int jobId)
{
Enqueue(taskId, jobId);
/* "Critical section"? Only the first thread to arrive should
* execute OneTimeInitAsync. Subsequent threads should always
* skip this part. This is where things go wrong as all the
* tasks execute this section due to lack of synchronization. */
if (!_initialized)
{
await OneTimeInitAsync(taskId);
}
/* Before and during initialization, all threads should skip
* the ProcessQueueAsync. After initialization is completed,
* it does not matter which thread will execute it (since the
* _jobsQueue is thread-safe). */
if (_initialized)
{
await ProcessQueueAsync(taskId);
}
Console.WriteLine($"Task {taskId} completed.");
}
private void Enqueue(int taskId, int jobId)
{
Console.WriteLine($"Task {taskId} enqueues job {jobId}.");
_jobsQueue.Enqueue(jobId);
}
private async Task OneTimeInitAsync(int taskId)
{
Console.WriteLine($"Task {taskId} is performing initialization");
/* Do some lengthy initialization */
await Task.Delay(TimeSpan.FromSeconds(3));
_initialized = true;
Console.WriteLine($"Task {taskId} completed initialization");
}
private async Task ProcessQueueAsync(int taskId)
{
while (_jobsQueue.TryDequeue(out int jobId))
{
/* Do something lengthy with the jobId */
await Task.Delay(TimeSpan.FromSeconds(1));
Console.WriteLine($"Task {taskId} completed job {jobId}.");
}
}
private static void Main(string[] args)
{
var p = new Program();
var rand = new Random();
/* Start 4 tasks in parallel */
for (var threadId = 1; threadId < 5; threadId++)
{
p.EnqueueAndProcessJobsAsync(threadId, rand.Next(10));
}
/* Give tasks chance to finish */
Console.ReadLine();
}
}
}
Both OneTimeInitAsync and ProcessQueueAsync are lengthy operations that in real-life scenario would communicate with some remote services. Using lock would block the other threads, while I want them to just pile up their work onto _jobsQueue and go their way. I tried using ManualResetEvent to no avail.
Does anyone know how would I make this work? Thanks in advance.
UPDATE (solution)
Based on the discussion below I understood that the scenario presented was not complete to describe my problem. However thanks to the answers and comments I thought about redesigning the solution a bit, to work as I wanted it to.
So imagine two remote services ServiceA (job processor) and ServiceB (job repository) that the client app has to communicate with. We need to establish a connection to ServiceA, meanwhile we get data for multiple jobs from ServiceB. As the jobs data become available, we process jobs (in batches) using ServiceA (real-life example involves Signal-R connection to ServiceA and some job IDs from ServiceB that need to be sent to ServiceA). Here's the code example:
public class StackOverflowSolution
{
private readonly ConcurrentQueue<int> _jobsQueue = new ConcurrentQueue<int>();
/* Just to randomize waiting times */
private readonly Random _random = new Random();
/* Instance-scoped one-time initialization of a remote ServiceA connection */
private async Task<string> InitializeConnectionAsync()
{
Console.WriteLine($"{nameof(InitializeConnectionAsync)} started");
await Task.Delay(TimeSpan.FromSeconds(_random.Next(5) + 1));
Console.WriteLine($"{nameof(InitializeConnectionAsync)} completed");
return "Connection";
}
/* Preparation of a job (assume it requires communication with remote ServiceB) */
private async Task InitializeJobAsync(int id)
{
Console.WriteLine($"{nameof(InitializeJobAsync)}({id}) started");
await Task.Delay(TimeSpan.FromSeconds(_random.Next(10) + 1));
_jobsQueue.Enqueue(id);
Console.WriteLine($"{nameof(InitializeJobAsync)}({id}) completed");
}
/* Does something to the ready jobs in the _jobsQueue using connection to
* ServiceA */
private async Task ProcessQueueAsync(string connection)
{
var sb = new StringBuilder("Processed ");
bool any = false;
while (_jobsQueue.TryDequeue(out int idResult))
{
any = true;
sb.Append($"{idResult}, ");
}
if (any)
{
await Task.Delay(TimeSpan.FromMilliseconds(_random.Next(500)));
Console.WriteLine(sb.ToString());
}
}
/* Orchestrates the processing */
public async Task RunAsync()
{
/* Start initializing the conection to ServiceA */
Task<string> connectionTask = InitializeConnectionAsync();
/* Start initializing jobs */
var jobTasks = new List<Task>();
foreach (int id in new[] {1, 2, 3, 4})
{
jobTasks.Add(InitializeJobAsync(id));
}
/* Wait for initialization to complete */
string connection = await connectionTask;
/* Trigger processing of jobs as they become ready */
var queueProcessingTasks = new List<Task>();
while (jobTasks.Any())
{
jobTasks.Remove(await Task.WhenAny(jobTasks));
queueProcessingTasks.Add(ProcessQueueAsync(connection));
}
await Task.WhenAll(queueProcessingTasks);
}
public static void Main()
{
new StackOverflowSolution().RunAsync().Wait();
}
}
Output example:
InitializeConnectionAsync started
InitializeJobAsync(1) started
InitializeJobAsync(2) started
InitializeJobAsync(3) started
InitializeJobAsync(4) started
InitializeJobAsync(5) started
InitializeJobAsync(3) completed
InitializeJobAsync(2) completed
InitializeConnectionAsync completed
Processed 3, 2,
InitializeJobAsync(1) completed
Processed 1,
InitializeJobAsync(5) completed
Processed 5,
InitializeJobAsync(4) completed
Processed 4,
Thanks for all the feedback!

Honestly the semantics of EnqueueAndProcessJobsAsync for your code is simply not a good idea, given your description of what you're actually doing and what you actually need.
Currently the Task returned from EnqueueAndProcessJobsAsync waits for initialization if initialization wasn't started by someone else, then it completes whenever the queue is empty, or as soon as this logical call context happened to process an item that errored. That...just doesn't make sense.
What you clearly want is for that Task to be completed whenever the job passed in completed (which will of course require initialization to complete), or to be errored if that job errors, and to be unaffected by any other job's errors. Fortunately in addition to being much more useful, it's also far easier to do.
As far as the actual initialization goes, you can just use a Lazy<Task> to ensure proper synchronization of the asynchronous initialization, and to expose the Task to any future calls which can tell them when initialization has finished.
public class MyAsyncQueueRequireingInitialization
{
private readonly Lazy<Task> whenInitialized;
public MyAsyncQueueRequireingInitialization()
{
whenInitialized = new Lazy<Task>(OneTimeInitAsync);
}
//as noted in comments, the taskID isn't actually needed for initialization
private async Task OneTimeInitAsync()
{
Console.WriteLine($"Performing initialization");
/* Do some lengthy initialization */
await Task.Delay(TimeSpan.FromSeconds(3));
Console.WriteLine($"Completed initialization");
}
public async Task ProcessJobAsync(int taskID, int jobId)
{
await whenInitialized.Value;
/* Do something lengthy with the jobId */
await Task.Delay(TimeSpan.FromSeconds(1));
Console.WriteLine($"Completed job {jobId}.");
}
}

As noted in the comments to the OP, a better solution might be to do the initialization in single-threaded mode, and then launch the threads that do the actual work.
If that doesn't work for you, you'll need a lock of some sort - but you can use that lock just for scheduling, to make it less blocking. I would implement something like this:
private bool _initializationIsScheduled = false;
private object _initializationIsScheduledLock = new object();
private bool _isInitialized = false;
private object _isInitializedLock = new object();
private async Task EnqueueAndProcessJobs(int taskId, int jobId)
{
var shouldDoHeavyWork = false;
lock(_initializationIsScheduledLock)
{
if (!_initializationIsScheduled)
{
shouldDoHeavyWork = true;
_initializationIsScheduled= true;
}
}
if (shouldDoHeavyWork)
{
await OneTimeInitAsync(taskId);
lock (_isInitializedLock)
{
_isInitialized = true;
}
}
lock (_isInitializedLock)
{
if (_isInitialized)
{
shouldDoHeavyWork = true;
}
}
if (shouldDoHeavyWork)
{
await ProcessQueueAsync(taskId);
}
Console.WriteLine($"Task {taskId} completed.");
}
Note how the only time the thread locks up the other threads, is when it's about to check or set one of the flags that controls its work. In other words, threads don't have to wait for each other when actually doing heavy work, just when scheduling it (i.e. for a couple of CPU cycles when setting the boolean flags).
The code isn't exactly beautiful, but you should be able to refactor the above sample into something that's reasonably legible... :)

Maybe you can try something like this:
static bool IsInitializing = false;
static int FirstThreadId = -1;
// check if initialised
// return if initialised
// somewhere in init mehtod
lock (lockObject)
{
// first method start initializing
IsInitializing = true;
// set some id to the thread that start initializtion
FirstThreadId = THIS_THREAD_OR_TASK_ID;
}
if (IsInitializing && FristThread != THIS_THREAD_OR_TASK_ID)
return; // skip initializing
lock must work quick, yeap

Do multiple awaits to the same Task from a single thread resume in FIFO order?

Supposing a Task is created and awaited multiple times from a single thread. Is the resume order FIFO?
Simplistic example: Is the Debug.Assert() really an invariant?
Task _longRunningTask;
async void ButtonStartSomething_Click()
{
// Wait for any previous runs to complete before starting the next
if (_longRunningTask != null) await _longRunningTask;
// Check our invariant
Debug.Assert(_longRunningTask == null, "This assumes awaits resume in FIFO order");
// Initialize
_longRunningTask = Task.Delay(10000);
// Yield and wait for completion
await _longRunningTask;
// Clean up
_longRunningTask = null;
}
Initialize and Clean up are kept to a bare minimum for the sake of simplicity, but the general idea is that the previous Clean up MUST be complete before the next Initialize runs.

The short answer is: no, it's not guaranteed.
Furthermore, you should not use ContinueWith; among other problems, it has a confusing default scheduler (more details on my blog). You should use await instead:
private async void ButtonStartSomething_Click()
{
// Wait for any previous runs to complete before starting the next
if (_longRunningTask != null) await _longRunningTask;
_longRunningTask = LongRunningTaskAsync();
await _longRunningTask;
}
private async Task LongRunningTaskAsync()
{
// Initialize
await Task.Delay(10000);
// Clean up
_longRunningTask = null;
}
Note that this could still have "interesting" semantics if the button can be clicked many times while the tasks are still running.
The standard way to prevent the multiple-execution problem for UI applications is to disable the button:
private async void ButtonStartSomething_Click()
{
ButtonStartSomething.Enabled = false;
await LongRunningTaskAsync();
ButtonStartSomething.Enabled = true;
}
private async Task LongRunningTaskAsync()
{
// Initialize
await Task.Delay(10000);
// Clean up
}
This forces your users into a one-operation-at-a-time queue.

The order of execution is pre-defined, however there is potential race condition on _longRunningTask variable if ButtonStartSomething_Click() is called concurrently from more than one thread (not likely the case).
Alternatively, you can explicitly schedule tasks using a queue. As a bonus a work can be scheduled from non-async methods, too:
void ButtonStartSomething_Click()
{
_scheduler.Add(async() =>
{
// Do something
await Task.Delay(10000);
// Do something else
});
}
Scheduler _scheduler;
class Scheduler
{
public Scheduler()
{
_queue = new ConcurrentQueue<Func<Task>>();
_state = STATE_IDLE;
}
public void Add(Func<Task> func)
{
_queue.Enqueue(func);
ScheduleIfNeeded();
}
public Task Completion
{
get
{
var t = _messageLoopTask;
if (t != null)
{
return t;
}
else
{
return Task.FromResult<bool>(true);
}
}
}
void ScheduleIfNeeded()
{
if (_queue.IsEmpty)
{
return;
}
if (Interlocked.CompareExchange(ref _state, STATE_RUNNING, STATE_IDLE) == STATE_IDLE)
{
_messageLoopTask = Task.Run(new Func<Task>(RunMessageLoop));
}
}
async Task RunMessageLoop()
{
Func<Task> item;
while (_queue.TryDequeue(out item))
{
await item();
}
var oldState = Interlocked.Exchange(ref _state, STATE_IDLE);
System.Diagnostics.Debug.Assert(oldState == STATE_RUNNING);
if (!_queue.IsEmpty)
{
ScheduleIfNeeded();
}
}
volatile Task _messageLoopTask;
ConcurrentQueue<Func<Task>> _queue;
static int _state;
const int STATE_IDLE = 0;
const int STATE_RUNNING = 1;
}

Found the answer under Task.ContinueWith(). It appear to be: no
Presuming await is just Task.ContinueWith() under the hood, there's documentation for TaskContinuationOptions.PreferFairness that reads:
A hint to a TaskScheduler to schedule task in the order in which they were scheduled, so that tasks scheduled sooner are more likely to run sooner, and tasks scheduled later are more likely to run later.
(bold-facing added)
This suggests there's no guarantee of any sorts, inherent or otherwise.
Correct ways to do this
For the sake of someone like me (OP), here's a look at the more correct ways to do this.
Based on Stephen Cleary's answer:
private async void ButtonStartSomething_Click()
{
// Wait for any previous runs to complete before starting the next
if (_longRunningTask != null) await _longRunningTask;
// Initialize
_longRunningTask = ((Func<Task>)(async () =>
{
await Task.Delay(10);
// Clean up
_longRunningTask = null;
}))();
// Yield and wait for completion
await _longRunningTask;
}
Suggested by Raymond Chen's comment:
private async void ButtonStartSomething_Click()
{
// Wait for any previous runs to complete before starting the next
if (_longRunningTask != null) await _longRunningTask;
// Initialize
_longRunningTask = Task.Delay(10000)
.ContinueWith(task =>
{
// Clean up
_longRunningTask = null;
}, TaskContinuationOptions.OnlyOnRanToCompletion);
// Yield and wait for completion
await _longRunningTask;
}
Suggested by Kirill Shlenskiy's comment:
readonly SemaphoreSlim _taskSemaphore = new SemaphoreSlim(1);
async void ButtonStartSomething_Click()
{
// Wait for any previous runs to complete before starting the next
await _taskSemaphore.WaitAsync();
try
{
// Do some initialization here
// Yield and wait for completion
await Task.Delay(10000);
// Do any clean up here
}
finally
{
_taskSemaphore.Release();
}
}
(Please -1 or comment if I've messed something up in either.)
Handling exceptions
Using continuations made me realize one thing: awaiting at multiple places gets complicated really quickly if _longRunningTask can throw exceptions.
If I'm going to use continuations, it looks like I need to top it off by handling all exceptions within the continuation as well.
i.e.
_longRunningTask = Task.Delay(10000)
.ContinueWith(task =>
{
// Clean up
_longRunningTask = null;
}, TaskContinuationOptions.OnlyOnRanToCompletion);
.ContinueWith(task =>
{
// Consume or handle exceptions here
}, TaskContinuationOptions.OnlyOnFaulted);
// Yield and wait for completion
await _longRunningTask;
If I use a SemaphoreSlim, I can do the same thing in the try-catch, and have the added option of bubbling exceptions directly out of ButtonStartSomething_Click.

Why is this TAP async/await code slower than the TPL version?

I had to write a console application that called Microsoft Dynamics CRM web service to perform an action on over eight thousand CRM objects. The details of the web service call are irrelevant and not shown here but I needed a multi-threaded client so that I could make calls in parallel. I wanted to be able to control the number of threads used from a config setting and also for the application to cancel the whole operation if the number of service errors reached a config-defined threshold.
I wrote it using Task Parallel Library Task.Run and ContinueWith, keeping track of how many calls (threads) were in progress, how many errors we'd received, and whether the user had cancelled from the keyboard. Everything worked fine and I had extensive logging to assure myself that threads were finishing cleanly and that everything was tidy at the end of the run. I could see that the program was using the maximum number of threads in parallel and, if our maximum limit was reached, waiting until a running task completed before starting another one.
During my code review, my colleague suggested that it would be better to do it with async/await instead of tasks and continuations, so I created a branch and rewrote it that way. The results were interesting - the async/await version was almost twice as slow, and it never reached the maximum number of allowed parallel operations/threads. The TPL one always got to 10 threads in parallel whereas the async/await version never got beyond 5.
My question is: have I made a mistake in the way I have written the async/await code (or the TPL code even)? If I have not coded it wrong, can you explain why the async/await is less efficient, and does that mean it is better to carry on using TPL for multi-threaded code.
Note that the code I tested with did not actually call CRM - the CrmClient class simply thread-sleeps for a duration specified in the config (five seconds) and then throws an exception. This meant that there were no external variables that could affect the performance.
For the purposes of this question I created a stripped down program that combines both versions; which one is called is determined by a config setting. Each of them starts with a bootstrap runner that sets up the environment, creates the queue class, then uses a TaskCompletionSource to wait for completion. A CancellationTokenSource is used to signal a cancellation from the user. The list of ids to process is read from an embedded file and pushed onto a ConcurrentQueue. They both start off calling StartCrmRequest as many times as max-threads; subsequently, every time a result is processed, the ProcessResult method calls StartCrmRequest again, keeping going until all of our ids are processed.
You can clone/download the complete program from here: https://bitbucket.org/kentrob/pmgfixso/
Here is the relevant configuration:
<appSettings>
<add key="TellUserAfterNCalls" value="5"/>
<add key="CrmErrorsBeforeQuitting" value="20"/>
<add key="MaxThreads" value="10"/>
<add key="CallIntervalMsecs" value="5000"/>
<add key="UseAsyncAwait" value="True" />
</appSettings>
Starting with the TPL version, here is the bootstrap runner that kicks off the queue manager:
public static class TplRunner
{
private static readonly CancellationTokenSource CancellationTokenSource = new CancellationTokenSource();
public static void StartQueue(RuntimeParameters parameters, IEnumerable<string> idList)
{
Console.CancelKeyPress += (s, args) =>
{
CancelCrmClient();
args.Cancel = true;
};
var start = DateTime.Now;
Program.TellUser("Start: " + start);
var taskCompletionSource = new TplQueue(parameters)
.Start(CancellationTokenSource.Token, idList);
while (!taskCompletionSource.Task.IsCompleted)
{
if (Console.KeyAvailable)
{
if (Console.ReadKey().Key != ConsoleKey.Q) continue;
Console.WriteLine("When all threads are complete, press any key to continue.");
CancelCrmClient();
}
}
var end = DateTime.Now;
Program.TellUser("End: {0}. Elapsed = {1} secs.", end, (end - start).TotalSeconds);
}
private static void CancelCrmClient()
{
CancellationTokenSource.Cancel();
Console.WriteLine("Cancelling Crm client. Web service calls in operation will have to run to completion.");
}
}
Here is the TPL queue manager itself:
public class TplQueue
{
private readonly RuntimeParameters parameters;
private readonly object locker = new object();
private ConcurrentQueue<string> idQueue = new ConcurrentQueue<string>();
private readonly CrmClient crmClient;
private readonly TaskCompletionSource<bool> taskCompletionSource = new TaskCompletionSource<bool>();
private int threadCount;
private int crmErrorCount;
private int processedCount;
private CancellationToken cancelToken;
public TplQueue(RuntimeParameters parameters)
{
this.parameters = parameters;
crmClient = new CrmClient();
}
public TaskCompletionSource<bool> Start(CancellationToken cancellationToken, IEnumerable<string> ids)
{
cancelToken = cancellationToken;
foreach (var id in ids)
{
idQueue.Enqueue(id);
}
threadCount = 0;
// Prime our thread pump with max threads.
for (var i = 0; i < parameters.MaxThreads; i++)
{
Task.Run((Action) StartCrmRequest, cancellationToken);
}
return taskCompletionSource;
}
private void StartCrmRequest()
{
if (taskCompletionSource.Task.IsCompleted)
{
return;
}
if (cancelToken.IsCancellationRequested)
{
Program.TellUser("Crm client cancelling...");
ClearQueue();
return;
}
var count = GetThreadCount();
if (count >= parameters.MaxThreads)
{
return;
}
string id;
if (!idQueue.TryDequeue(out id)) return;
IncrementThreadCount();
crmClient.CompleteActivityAsync(new Guid(id), parameters.CallIntervalMsecs).ContinueWith(ProcessResult);
processedCount += 1;
if (parameters.TellUserAfterNCalls > 0 && processedCount%parameters.TellUserAfterNCalls == 0)
{
ShowProgress(processedCount);
}
}
private void ProcessResult(Task<CrmResultMessage> response)
{
if (response.Result.CrmResult == CrmResult.Failed && ++crmErrorCount == parameters.CrmErrorsBeforeQuitting)
{
Program.TellUser(
"Quitting because CRM error count is equal to {0}. Already queued web service calls will have to run to completion.",
crmErrorCount);
ClearQueue();
}
var count = DecrementThreadCount();
if (idQueue.Count == 0 && count == 0)
{
taskCompletionSource.SetResult(true);
}
else
{
StartCrmRequest();
}
}
private int GetThreadCount()
{
lock (locker)
{
return threadCount;
}
}
private void IncrementThreadCount()
{
lock (locker)
{
threadCount = threadCount + 1;
}
}
private int DecrementThreadCount()
{
lock (locker)
{
threadCount = threadCount - 1;
return threadCount;
}
}
private void ClearQueue()
{
idQueue = new ConcurrentQueue<string>();
}
private static void ShowProgress(int processedCount)
{
Program.TellUser("{0} activities processed.", processedCount);
}
}
Note that I am aware that a couple of the counters are not thread safe but they are not critical; the threadCount variable is the only critical one.
Here is the dummy CRM client method:
public Task<CrmResultMessage> CompleteActivityAsync(Guid activityId, int callIntervalMsecs)
{
// Here we would normally call a CRM web service.
return Task.Run(() =>
{
try
{
if (callIntervalMsecs > 0)
{
Thread.Sleep(callIntervalMsecs);
}
throw new ApplicationException("Crm web service not available at the moment.");
}
catch
{
return new CrmResultMessage(activityId, CrmResult.Failed);
}
});
}
And here are the same async/await classes (with common methods removed for the sake of brevity):
public static class AsyncRunner
{
private static readonly CancellationTokenSource CancellationTokenSource = new CancellationTokenSource();
public static void StartQueue(RuntimeParameters parameters, IEnumerable<string> idList)
{
var start = DateTime.Now;
Program.TellUser("Start: " + start);
var taskCompletionSource = new AsyncQueue(parameters)
.StartAsync(CancellationTokenSource.Token, idList).Result;
while (!taskCompletionSource.Task.IsCompleted)
{
...
}
var end = DateTime.Now;
Program.TellUser("End: {0}. Elapsed = {1} secs.", end, (end - start).TotalSeconds);
}
}
The async/await queue manager:
public class AsyncQueue
{
private readonly RuntimeParameters parameters;
private readonly object locker = new object();
private readonly CrmClient crmClient;
private readonly TaskCompletionSource<bool> taskCompletionSource = new TaskCompletionSource<bool>();
private CancellationToken cancelToken;
private ConcurrentQueue<string> idQueue = new ConcurrentQueue<string>();
private int threadCount;
private int crmErrorCount;
private int processedCount;
public AsyncQueue(RuntimeParameters parameters)
{
this.parameters = parameters;
crmClient = new CrmClient();
}
public async Task<TaskCompletionSource<bool>> StartAsync(CancellationToken cancellationToken,
IEnumerable<string> ids)
{
cancelToken = cancellationToken;
foreach (var id in ids)
{
idQueue.Enqueue(id);
}
threadCount = 0;
// Prime our thread pump with max threads.
for (var i = 0; i < parameters.MaxThreads; i++)
{
await StartCrmRequest();
}
return taskCompletionSource;
}
private async Task StartCrmRequest()
{
if (taskCompletionSource.Task.IsCompleted)
{
return;
}
if (cancelToken.IsCancellationRequested)
{
...
return;
}
var count = GetThreadCount();
if (count >= parameters.MaxThreads)
{
return;
}
string id;
if (!idQueue.TryDequeue(out id)) return;
IncrementThreadCount();
var crmMessage = await crmClient.CompleteActivityAsync(new Guid(id), parameters.CallIntervalMsecs);
ProcessResult(crmMessage);
processedCount += 1;
if (parameters.TellUserAfterNCalls > 0 && processedCount%parameters.TellUserAfterNCalls == 0)
{
ShowProgress(processedCount);
}
}
private async void ProcessResult(CrmResultMessage response)
{
if (response.CrmResult == CrmResult.Failed && ++crmErrorCount == parameters.CrmErrorsBeforeQuitting)
{
Program.TellUser(
"Quitting because CRM error count is equal to {0}. Already queued web service calls will have to run to completion.",
crmErrorCount);
ClearQueue();
}
var count = DecrementThreadCount();
if (idQueue.Count == 0 && count == 0)
{
taskCompletionSource.SetResult(true);
}
else
{
await StartCrmRequest();
}
}
}
So, setting MaxThreads to 10 and CrmErrorsBeforeQuitting to 20, the TPL version on my machine completes in 19 seconds and the async/await version takes 35 seconds. Given that I have over 8000 calls to make this is a significant difference. Any ideas?

I think I'm seeing the problem here, or at least a part of it. Look closely at the two bits of code below; they are not equivalent.
// Prime our thread pump with max threads.
for (var i = 0; i < parameters.MaxThreads; i++)
{
Task.Run((Action) StartCrmRequest, cancellationToken);
}
And:
// Prime our thread pump with max threads.
for (var i = 0; i < parameters.MaxThreads; i++)
{
await StartCrmRequest();
}
In the original code (I am taking it as a given that it is functionally sound) there is a single call to ContinueWith. That is exactly how many await statements I would expect to see in a trivial rewrite if it is meant to preserve the original behaviour.
Not a hard and fast rule and only applicable in simple cases, but nevertheless a good thing to keep an eye out for.

I think you over complicated your solution and ended up not getting where you wanted in either implementation.
First of all, connections to any HTTP host are limited by the service point manager. The default limit for client environments is 2, but you can increase it yourself.
No matter how much threads you spawn, there won't be more active requests than those allwed.
Then, as someone pointed out, await logically blocks the execution flow.
And finally, you spent your time creating an AsyncQueue when you should have used TPL data flows.

When implemented with async/await, I would expect the I/O bound algorithm to run on a single thread. Unlike #KirillShlenskiy, I believe that the bit responsible for "bringing back" to caller's context is not responsible for the slow-down. I think you overrun the thread pool by trying to use it for I/O-bound operations. It's designed primarily for compute-bound ops.
Have a look at ForEachAsync. I feel that's what you're looking for (Stephen Toub's discussion, you'll find Wischik's videos meaningful too):
http://blogs.msdn.com/b/pfxteam/archive/2012/03/05/10278165.aspx
(Use degree of concurrency to reduce memory footprint)
http://vimeo.com/43808831
http://vimeo.com/43808833

How to limit number of HttpWebRequest per second towards a webserver?

I need to implement a throttling mechanism (requests per second) when using HttpWebRequest for making parallel requests towards one application server. My C# app must issue no more than 80 requests per second to a remote server. The limit is imposed by the remote service admins not as a hard limit but as "SLA" between my platform and theirs.
How can I control the number of requests per second when using HttpWebRequest?

I had the same problem and couldn't find a ready solution so I made one, and here it is. The idea is to use a BlockingCollection<T> to add items that need processing and use Reactive Extensions to subscribe with a rate-limited processor.
Throttle class is the renamed version of this rate limiter
public static class BlockingCollectionExtensions
{
// TODO: devise a way to avoid problems if collection gets too big (produced faster than consumed)
public static IObservable<T> AsRateLimitedObservable<T>(this BlockingCollection<T> sequence, int items, TimeSpan timePeriod, CancellationToken producerToken)
{
Subject<T> subject = new Subject<T>();
// this is a dummyToken just so we can recreate the TokenSource
// which we will pass the proxy class so it can cancel the task
// on disposal
CancellationToken dummyToken = new CancellationToken();
CancellationTokenSource tokenSource = CancellationTokenSource.CreateLinkedTokenSource(producerToken, dummyToken);
var consumingTask = new Task(() =>
{
using (var throttle = new Throttle(items, timePeriod))
{
while (!sequence.IsCompleted)
{
try
{
T item = sequence.Take(producerToken);
throttle.WaitToProceed();
try
{
subject.OnNext(item);
}
catch (Exception ex)
{
subject.OnError(ex);
}
}
catch (OperationCanceledException)
{
break;
}
}
subject.OnCompleted();
}
}, TaskCreationOptions.LongRunning);
return new TaskAwareObservable<T>(subject, consumingTask, tokenSource);
}
private class TaskAwareObservable<T> : IObservable<T>, IDisposable
{
private readonly Task task;
private readonly Subject<T> subject;
private readonly CancellationTokenSource taskCancellationTokenSource;
public TaskAwareObservable(Subject<T> subject, Task task, CancellationTokenSource tokenSource)
{
this.task = task;
this.subject = subject;
this.taskCancellationTokenSource = tokenSource;
}
public IDisposable Subscribe(IObserver<T> observer)
{
var disposable = subject.Subscribe(observer);
if (task.Status == TaskStatus.Created)
task.Start();
return disposable;
}
public void Dispose()
{
// cancel consumption and wait task to finish
taskCancellationTokenSource.Cancel();
task.Wait();
// dispose tokenSource and task
taskCancellationTokenSource.Dispose();
task.Dispose();
// dispose subject
subject.Dispose();
}
}
}
Unit test:
class BlockCollectionExtensionsTest
{
[Fact]
public void AsRateLimitedObservable()
{
const int maxItems = 1; // fix this to 1 to ease testing
TimeSpan during = TimeSpan.FromSeconds(1);
// populate collection
int[] items = new[] { 1, 2, 3, 4 };
BlockingCollection<int> collection = new BlockingCollection<int>();
foreach (var i in items) collection.Add(i);
collection.CompleteAdding();
IObservable<int> observable = collection.AsRateLimitedObservable(maxItems, during, CancellationToken.None);
BlockingCollection<int> processedItems = new BlockingCollection<int>();
ManualResetEvent completed = new ManualResetEvent(false);
DateTime last = DateTime.UtcNow;
observable
// this is so we'll receive exceptions
.ObserveOn(new SynchronizationContext())
.Subscribe(item =>
{
if (item == 1)
last = DateTime.UtcNow;
else
{
TimeSpan diff = (DateTime.UtcNow - last);
last = DateTime.UtcNow;
Assert.InRange(diff.TotalMilliseconds,
during.TotalMilliseconds - 30,
during.TotalMilliseconds + 30);
}
processedItems.Add(item);
},
() => completed.Set()
);
completed.WaitOne();
Assert.Equal(items, processedItems, new CollectionEqualityComparer<int>());
}
}

The Throttle() and Sample() extension methods (On Observable) allow you to regulate a fast sequence of events into a "slower" sequence.
Here is a blog post with an example of Sample(Timespan) that ensures a maxium rate.

My original post discussed how to add a throttling mechanism to WCF via client behavior extensions, but then was pointed out that I misread the question (doh!).
Overall the approach can be to check with a class that determines if we are violating the rate limit or not. There's already been a lot of discussion around how to check for rate violations.
Throttling method calls to M requests in N seconds
If you are violating the rate limit, then sleep for a fix interval and check again. If not, then go ahead and make the HttpWebRequest call.