Allocation free ‘async’ Methods

Task Asynchronous Programming (TAP) model will go down as one of the landmark of C# language revolution. The typical method signature with return type Task/Task<T> has since then made significant appearances in our programming life.

But despite all its glorious functionalities, it needs to be noted that it comes at a certain cost – performance bottleneck, especially when you are working in tight loops in a memory constrained environment. Memory allocations of Task/Task<T> (note both are references types) impacts the performances adversely in critical paths. What developers longed for in such scenarios is a comparatively light-weight value type, which could be returned instead of the references types in the async methods. This was a limitation till C# 7.x happened.

Starting from C# 7, an async method could return any type that has an accessible ‘GetAwaiter’ method, which returns an object that implements INotifyCompletion and ICriticalNotifyCompletion Interfaces. The ValueTask<T> is a "goto" under such circumstances. Being a value type, it doesn’t require additional memory allocations. This is extremely useful when you already have the ‘cached result’ and the operation can be completed synchronously. But like Task, it comes at a cost, which we will discuss later in the article.

To begin with, let’s write some sample code to demonstrate the need of ValueTask<T> or rather, need for replacing Task<T>. Consider the following code.

public interface IWeatherService
{
  Task<double> GetWeather();
}

public class MockWeatherService : IWeatherService 
{
  public async Task<double> GetWeather() => await Task.FromResult(30);
}

The above code demonstrate a Mock Weather Service. In real world scenario, this might be accessing a Web API to fetch current weather details. For the sake of example, only the temperature is fetched, which is hard coded. Consider GetWeather to be invoked quite frequently (say in a loop) by the Client. There is a chance this might turn disastrous if this was executed in a memory constraint environment.

Ideally, you wouldn’t want to fetch from the Web API always. You would prefer to cache the result, and only fetch from the Web API only after specific interval (from the time the last value was cached). Despite doing the operation (fetch from cache) synchronously, you are allocating memory for the reference type Task<T>. One would have wished to avoid this scenario and yet use the async syntax.

This is possible by making use of the ValueTask<T> DataType. Let’s change our syntax a bit.

public interface ICachedWeatherService
{
  ValueTask<double> GetWeather();
}
public class MockCachedWeatherService : ICachedWeatherService
{
  private DateTime _lastAccessedTime = DateTime.MinValue;
  private double _lastAccessedValue = 30;
  public async ValueTask<double> GetWeather()
  {
    if(DateTime.Now - _lastAccessedTime < TimeSpan.FromSeconds(10))
    {
      return _lastAccessedValue;
    }
    _lastAccessedTime = DateTime.Now;
    return await Task.FromResult(30);
   }
}

In the above code, we have changed the signature of the method (return type) to ValueTask<T>. Since it implements all the necessary requirements that is needed for async method (as discussed earlier), we could still use the same signature in client to access the method, just like in Task based syntax.

Using Task Return Type

var _weatherService = new MockWeatherService();
           
for(int i = 0; i < 10; i++)
{
  await Task.Delay(5000);
  var currentWeather = await _weatherService.GetWeather();
  Console.WriteLine($"Weather @ {DateTime.Now.ToString("hh:mm:ss")}={currentWeather}");
}

Using ValueTask Return Type

var _cachedWeatherService = new MockCachedWeatherService();
for (int i = 0; i < 10; i++)
{
  await Task.Delay(5000);
  var currentWeather = await _cachedWeatherService.GetWeather();
  Console.WriteLine($"Weather @ {DateTime.Now.ToString("hh:mm:ss")}={currentWeather}");
}

Check Complete Code Sample

Difference is, when the method needs to return from the cache, it returns ValueTask avoiding the pain of creating the Task Reference Type. However, when it has to go through the Web API, it can still return Task<T>.

Even though ValueTask seem highly useful, you need to maintain caution and should rather use it only if there is considerable performance issues that were measured. The reason for this is that, despite the ValueTask helps you avoid unnecessary allocation, it comes with its own cost. The ValueTask is a two field structure compared to Task<T>, which despite being reference type has single field. More fields would result in more data to copy. It fact Microsoft warns us in the API documentation itself.

Quoting from Documentation

There are tradeoffs to using a ValueTask<TResult> instead of a Task<TResult>. For example, while a ValueTask<TResult> can help avoid an allocation in the case where the successful result is available synchronously, it also contains two fields whereas a Task<TResult> as a reference type is a single field. This means that a method call ends up returning two fields worth of data instead of one, which is more data to copy. It also means that if a method that returns one of these is awaited within an async method, the state machine for that async method will be larger due to needing to store the struct that’s two fields instead of a single reference.

Hence, stick onto Task/Task<T> for almost all your use, until performance analysis pushes you to use ValueTask<T>. Complete code samples described in the example is available in my Github

Updated Span API : NonPortableCast and AsReadOnlySpan to MemoryMarshal.Cast and AsSpan

One of the interesting changes (not sure why Microsoft changed this) was moving/changing the Span<T> API’s a bit. For example, the Span<T>.NonPortableCast Method was moved as an extension method under MemoryMarshal.

Previous API

Span<byte> byteSpan = BitConverter.GetBytes(originalPrimitiveType);
Span<int> intSpan = byteSpan.NonPortableCast<byte, int>();

Current API

Span<byte> byteSpan = BitConverter.GetBytes(originalPrimitiveType);
Span<int> intSpan = MemoryMarshal.Cast<byte, int>(byteSpan);

Similarly, the string.AsReadonlySpan method has been renamed. The new API is as follows.

Previous API

ReadOnlySpan<char> keyString = "SomeText".AsReadOnlySpan();

Current API

ReadOnlySpan<char> keyString = "SomeText".AsSpan();

The MemoryMarshal extensions can be found under the System.Runtime.InteropServices namespace. Code discussed in my earlier post about Benchmarking Span has been updated in my Github.

The new String Interpolation and Ternary Operator

A quick tip for the new C# string interpolation. The new string interpolation has made things more simpler and readable. One might be tempted to include ternary operators within string interpolation.

However if you attempt something like following, the compiler would throw an error

Console.WriteLine($"{1==1?"Msg1":"Msg2"}");

The fix is as simple as it can get. Just include a brace within your ternary statement.

Console.WriteLine($"{(1==1?"Msg1":"Msg2")}");

That’s it. Brace up for a week with lot of fun and coding.

Deconstruct and Extension Methods

In an earlier post, we explored the Deconstruct feature for Non-Tuples. It was a pretty useful feature that came along in the 7.x series . If one does have access to source, adding deconstruct is quite easy. But what about classes we do not have (source code) access to ? For example, the .Net framework classes ? Can we add a deconstruct to it ? It turns out, it is quite simple as well. Extension Methods allows us to add deconstruct to existing classes without breaking the Open Closed Principle.

For demonstration purpose, let us write a deconstruct for Point Class under the System.Drawing namespace.

public static class ExtensionMethods
{
    public static void Deconstruct(this Point source,out double x,out double y)
    {
        x = source.X;
        y = source.Y;
    }
}

void Main()
{
       Point point = new Point(100,500);
       var (x,y) = point;
       Console.WriteLine($"Point.X={x},Point.Y={y}");
}

As observed, it is quite easy to add deconstruct to your existing classes using the Extension Methods.

Caliburn.Micro #007 : Bootstrapper with Unity

Unity is one of the most commonly used IOC containers used by developers, and it would be unfair if I don’t include it as a part of the Caliburn.Micro WPF Tutorials (Creating Bootstrappers).  This particular post is dedicated to building bootstrapper for Caliburn.Micro application using Unity as the IoC Container.


The key for creating Bootstrappers for Caliburn.Micro application lies in overriding 4 methods from the BootstrapperBase class. The 4 methods are
  1. Configure
  2. BuildUp
  3. GetInstance
  4. GetAllInstance
The implementation using SimpleContainer and MEF is described in the earlier parts of the Caliburn.Micro Tutorials. Let’s go ahead and implement the Configure method using Unity.
private IUnityContainer _unityContainer;
protected override void Configure()
{
    _unityContainer = new UnityContainer();
    _unityContainer.RegisterInstance<IWindowManager>(new WindowManager());
    _unityContainer.RegisterInstance<IEventAggregator>(new EventAggregator(), new ContainerControlledLifetimeManager());
    //  View Models
    _unityContainer.RegisterInstance<IShellViewModel>(new ShellViewModel());
}
Container using Unity is initialized using the UnityContainer class. RegisterInstance methods are used to register your Classes. For cases, where you need Singletons, you can make use of the ContainerControlledLifeTimeManager as seen in the above example. This place will also hold your ViewModel registrations as well.


Rest of the 3 methods are pretty simple and straightforward to understand.
protected override void BuildUp(object instance)
{
  _unityContainer.BuildUp(instance);
  base.BuildUp(instance);
}
protected override object GetInstance(Type service, string key)
{
  return string.IsNullOrEmpty(key) ? _unityContainer.Resolve(service, key) : _unityContainer.Resolve(service);
}
protected override IEnumerable<object> GetAllInstances(Type service)
{
  return _unityContainer.ResolveAll(service);
}
Finally, similar to earlier Bootstrappers, we use the OnStartup Method to invoke the ShellViewModel.

protected override void OnStartup(object sender, StartupEventArgs e)
{
   DisplayRootViewFor<IShellViewModel>();
}

The complete code sample can be accessed at my Github. The bootstrapper template is also included as a part of Caliburn.Micro Template Pack for WPF.

Lock for Async Methods

Despite the intended nature, we seldom might come across situation where we would want to place a lock around an async statement and make it behave differently, especially when accessed from different threads. Let’s try to mock the scenario first to demonstrate the issue.

The lock-async Problem

Consider the following code snippet.

void Main()
{
     Parallel.For(0,5,async (x,y)=>{await TestMethod();});
}

async Task<bool> TestMethod()
{
Console.WriteLine("In");
// Do some tasks here. For demonstration purpose we will use Task.Delay to mock a long running task.
await Task.Delay(2000);
Console.WriteLine("Out");
return await Task.FromResult(true);
}

The output of the above code would be as following, understandably allowing multiple threads to access the ‘intended long running process code‘.

In
In
In
In
In
Out
Out
Out
Out
Out

The most common workaround is placing a lock statement around the code. Try placing a lock around the await statement and you would surprised to know you are not allowed to do so.

“Cannot await in the body of a lock statement”

This is where the trouble starts.

The Solution

The workaround for the issue with Semaphore. Instead of lock, you could use a semaphore to ensure only one/limited threads access the async method at a time.

Let us rewrite the code with semaphore now.

void Main()
{
       Parallel.For(0,5,async (x,y)=>{await TestMethod();});
}
static SemaphoreSlim semaphoreSlim = new SemaphoreSlim(1,1);
async Task<bool> TestMethod()
{
await semaphoreSlim.WaitAsync();
try
{
Console.WriteLine("In");
await Task.Delay(2000);
Console.WriteLine("Out");
}
finally
{
semaphoreSlim.Release();
}
return await Task.FromResult(true);
}

As you can observe, the compile would accept this solution as desired. The output of following can seen below.

In
Out
In
Out
In
Out
In
Out
In
Out

Enjoy coding