tabs ↹ over ␣ ␣ ␣ spaces

by Jiří {x2} Činčura

Surprising cost of synchronous fast-paths in asynchronous methods

5 Aug 2021 7 mins .NET, Async, C#, Performance, Programming in general

I’m doing some performance related work in async space, and I was surprised by the slowness of synchronous code-paths in the asynchronous methods. Let’s look at some code.

The best-case scenario for fast-path is that you don’t have any await in the call chain. Something like this.

public Task<FooBar> Foo() => Bar();
public Task<FooBar> Bar() => Baz();
public Task<FooBar> Baz() => Task.FromResult(new FooBar());

Worse scenario is when you need to include some await (i.e., because of using).

public async Task<FooBar> Foo() => await Bar();
public async Task<FooBar> Bar() => await Baz();
public Task<FooBar> Baz() => Task.FromResult(new FooBar());

You can now pause and guess what the cost of the first example and second example is going to be compared to “normal” call. The first one should be more or less the same, right? There’s nothing specific to async/await. Just method calls. And what about ValueTask? And does the length of chain matter?

To get the answers I designed this benchmark.

using System.Runtime.CompilerServices;
using System.Threading.Tasks;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;
//using TestDataType = System.Boolean;
//using TestDataType = System.Decimal;
using TestDataType = System.String;

[MemoryDiagnoser, DisassemblyDiagnoser]
public class Test
{
	//public const TestDataType Result = true;
	//public const TestDataType Result = 1234.567m;
	public const TestDataType Result = "foobar";

	A _a;
	B _b;
	C _c;
	D _d;
	E _e;

	[GlobalSetup]
	public void GlobalSetup()
	{
		_a = new();
		_b = new();
		_c = new();
		_d = new();
		_e = new();
	}

	[Params(true, false)]
	public bool Deep { get; set; }

	[Benchmark(Baseline = true)]
	public TestDataType DirectCall()
	{
		return Deep
			? _a.Do(_b, _c, _d, _e)
			: _e.Do();
	}

	[Benchmark]
	public TestDataType DirectPathWithTask()
	{
		return Deep
			? _a.DoTask(_b, _c, _d, _e).GetAwaiter().GetResult()
			: _e.DoTask().GetAwaiter().GetResult();
	}

	[Benchmark]
	public async Task<TestDataType> AwaitPathWithTask()
	{
		return Deep
			? await _a.DoAwaitTask(_b, _c, _d, _e)
			: await _e.DoAwaitTask();
	}

	[Benchmark]
	public TestDataType DirectPathWithValueTask()
	{
		return Deep
			? _a.DoValueTask(_b, _c, _d, _e).GetAwaiter().GetResult()
			: _e.DoValueTask().GetAwaiter().GetResult();
	}

	[Benchmark]
	public async ValueTask<TestDataType> AwaitPathWithValueTask()
	{
		return Deep
			? await _a.DoAwaitValueTask(_b, _c, _d, _e)
			: await _e.DoAwaitValueTask();
	}
}

class A
{
	[MethodImpl(MethodImplOptions.NoInlining)]
	public TestDataType Do(B b, C c, D d, E e) => b.Do(c, d, e);
	public Task<TestDataType> DoTask(B b, C c, D d, E e) => b.DoTask(c, d, e);
	public async Task<TestDataType> DoAwaitTask(B b, C c, D d, E e) => await b.DoAwaitTask(c, d, e);
	public ValueTask<TestDataType> DoValueTask(B b, C c, D d, E e) => b.DoValueTask(c, d, e);
	public async ValueTask<TestDataType> DoAwaitValueTask(B b, C c, D d, E e) => await b.DoAwaitValueTask(c, d, e);
}
class B
{
	[MethodImpl(MethodImplOptions.NoInlining)]
	public TestDataType Do(C c, D d, E e) => c.Do(d, e);
	public Task<TestDataType> DoTask(C c, D d, E e) => c.DoTask(d, e);
	public async Task<TestDataType> DoAwaitTask(C c, D d, E e) => await c.DoAwaitTask(d, e);
	public ValueTask<TestDataType> DoValueTask(C c, D d, E e) => c.DoValueTask(d, e);
	public async ValueTask<TestDataType> DoAwaitValueTask(C c, D d, E e) => await c.DoAwaitValueTask(d, e);
}
class C
{
	[MethodImpl(MethodImplOptions.NoInlining)]
	public TestDataType Do(D d, E e) => d.Do(e);
	public Task<TestDataType> DoTask(D d, E e) => d.DoTask(e);
	public async Task<TestDataType> DoAwaitTask(D d, E e) => await d.DoAwaitTask(e);
	public ValueTask<TestDataType> DoValueTask(D d, E e) => d.DoValueTask(e);
	public async ValueTask<TestDataType> DoAwaitValueTask(D d, E e) => await d.DoAwaitValueTask(e);
}
class D
{
	[MethodImpl(MethodImplOptions.NoInlining)]
	public TestDataType Do(E e) => e.Do();
	public Task<TestDataType> DoTask(E e) => e.DoTask();
	public async Task<TestDataType> DoAwaitTask(E e) => await e.DoAwaitTask();
	public ValueTask<TestDataType> DoValueTask(E e) => e.DoValueTask();
	public async ValueTask<TestDataType> DoAwaitValueTask(E e) => await e.DoAwaitValueTask();
}
class E
{
	[MethodImpl(MethodImplOptions.NoInlining)]
	public TestDataType Do() => Test.Result;
	public Task<TestDataType> DoTask() => Task.FromResult(Test.Result);
	public async Task<TestDataType> DoAwaitTask() => await Task.FromResult(Test.Result);
	public ValueTask<TestDataType> DoValueTask() => ValueTask.FromResult(Test.Result);
	public async ValueTask<TestDataType> DoAwaitValueTask() => await ValueTask.FromResult(Test.Result);
}

First thing to mention is the usage of TestDataType “alias” to test with different datatypes – bool as something that has known set of values, decimal for a value type and string for a reference type. And although the numbers very slightly, the overall result is always the same.

The Deep parameter controls the length of the chain – either calling the terminal methods on E directly or going through A, B, C, D respectively.

The Do method is using the NoInlining flag to level the playing field a bit, because usually in real world the method will not be that simple and hence probably not worth inlining.

The rest is just the benchmark testing Task and ValueTask in scenarios I outlined above.

Here’s the results.

Method Deep Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated Code Size
DirectCall False 0.4399 ns 0.0064 ns 0.0060 ns 1.00 0.00 - - - - 87 B
DirectPathWithTask False 6.4300 ns 0.2099 ns 0.5781 ns 13.94 1.34 0.0043 - - 72 B 219 B
AwaitPathWithTask False 28.8549 ns 0.1742 ns 0.1630 ns 65.60 1.06 0.0129 - - 216 B 376 B
DirectPathWithValueTask False 8.8030 ns 0.0031 ns 0.0028 ns 19.98 0.25 - - - - 511 B
AwaitPathWithValueTask False 50.6709 ns 0.1732 ns 0.1535 ns 115.00 1.47 - - - - 900 B
DirectCall True 6.2216 ns 0.0131 ns 0.0122 ns 1.00 0.00 - - - - 118 B
DirectPathWithTask True 6.3206 ns 0.1992 ns 0.5872 ns 1.07 0.10 0.0043 - - 72 B 219 B
AwaitPathWithTask True 94.6696 ns 0.2804 ns 0.2486 ns 15.22 0.05 0.0300 - - 504 B 376 B
DirectPathWithValueTask True 9.8505 ns 0.2449 ns 0.3433 ns 1.58 0.06 - - - - 511 B
AwaitPathWithValueTask True 138.4892 ns 0.2050 ns 0.1601 ns 22.26 0.05 - - - - 900 B

Run with this setup.

BenchmarkDotNet=v0.13.0, OS=Windows 10.0.19043.1110 (21H1/May2021Update)
AMD Ryzen 7 5800X, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET 5.0.8 (5.0.821.31504), X64 RyuJIT
  DefaultJob : .NET 5.0.8 (5.0.821.31504), X64 RyuJIT

For a short call chain, the direct call beats both direct call with Task as well as direct call with ValueTask. Once the chain is longer the difference between direct calls and Task smooths out. The ValueTask is still losing a bit (more about that later).

When you put await (and hence state machines) into the mix, the difference is enormous, even for short call chains. That means you can’t simply do this as a fast-path.

public async ValueTask<FooBar> Foo()
{
	if (FastPath)
		return Bar();
	return await BarAsync();
}

Finally, you might be wondering why the ValueTask versions are consistently slower than Task versions. Shouldn’t it be faster, especially in fast-paths? Yes. But not in this simple sequential benchmark. The Task is quickly allocated and then collected from Gen 0, which is fairly cheap. Also, the code for ValueTask is bigger (Code Size column) with more stuff happening – bigger structure, more copying, runtime checks around null, IValueTaskSource<T> and even Task<T>. But once you put bit more pressure onto GC, for example in highly concurrent scenarios (you can try that with simple Parallel.For in DirectPathWithTask and DirectPathWithValueTask), the numbers flip in favor of ValueTask.

Summary

Doing synchronous fast-paths is not as straightforward as it may look. When considering fast-path any await in call chains must be avoided as much as possible. And even then, it’s not blindly equal to “normal” method call with int, decimal, or … Then one needs to consider usage of ValueTask over Task regarding concurrency and GC pressure.

Profile Picture Jiří Činčura is .NET, C# and Firebird expert. He focuses on data and business layers, language constructs, parallelism, databases and performance. For almost two decades he contributes to open-source, i.e. FirebirdClient. He works as a senior software engineer for Microsoft. Frequent speaker and blogger at www.tabsoverspaces.com.