Unused async/await pairs are removed to optimize and reduce compiled code #104

R0ma-D · 2017-04-03T20:47:21Z

If some method can be written without await, then it should be written it without await, and remove the async keyword from the method. A non-async method returning Task is more efficient than an async method returning a value/void. Check here:
http://blog.stephencleary.com/2012/02/async-and-await.html

Simple async method:
public async Task UndirectTaskMethod()
{
return await Task.FromResult(0);
}
generate complex method based on state machine like this:
[DebuggerStepThrough, AsyncStateMachine(typeof(AsyncClassObject.d__1))]
public Task UndirectTaskMethod()
{
AsyncClassObject.d__1 d__ = new AsyncClassObject.d__1();
d__.<>4__this = this;
d__.<>t__builder = AsyncTaskMethodBuilder.Create();
d__.<>1__state = -1;
AsyncTaskMethodBuilder <>t__builder = d__.<>t__builder;
<>t__builder.Start<AsyncClassObject.d__1>(ref d__);
return d__.<>t__builder.Task;
}

…code

aensidhe · 2017-05-10T09:07:24Z

Yes, but your change has some implications:

public class A
{
    private IDisposable d;
    public Task F()
    {
        return d.DoAsync();      
    }

    public void Dispose()
    {
         d.Dispose();
    }
}

If you call F, then in some other thread user call Dispose, the caller of F will receive very strange behaviour. If we set async/await in F, there will be either ODI, or he will get result.

Also, I would like to see some benchmarks with and without async. Without them we will not accept this PR.

R0ma-D · 2017-05-10T19:20:35Z

I've attach "benchmark" class. With async keyword it takes 493 ms to create 10000 tasks, and only 4 ms without it.

`class Program
{
private const int DelayInMilliseconds = 1;
private const int Size = 10000;

    static void Main(string[] args)
    {
        Measure(() => Task.FromResult(false));

        GC.Collect();
        Console.Write("Measure async ... ");    //493 ms
        var timeAsync = Measure(() => ExecuteWithAsync());
        Console.WriteLine(timeAsync + " ms");

        GC.Collect();
        Console.Write("Measure not async ... "); //4 ms
        var timeNotAsync = Measure(() => ExecuteWithoutAsync());
        Console.WriteLine(timeNotAsync + " ms");

        Console.ReadKey();
    }

    public static long Measure(Func<Task> creator)
    {
        Stopwatch stopwatch = new Stopwatch();

        stopwatch.Start();

        Task[] all = new Task[Size];
        for (int i = 0; i < Size; i++)
            all[i] = creator();
        stopwatch.Stop();

        Task.WaitAll(all);

        return stopwatch.ElapsedMilliseconds;
    }

    public static async Task ExecuteWithAsync()
    {
        await Task.Delay(DelayInMilliseconds);
    }

    public static Task ExecuteWithoutAsync()
    {
        return Task.Delay(DelayInMilliseconds);
    }
}`

aensidhe · 2017-05-12T09:40:58Z

We are not just "creating" tasks in air. We are making network calls with latencies large enough to be considered. Also, I suggest to use BenchmarkDotNet

I'll try to write a benchmark next week to test it.

How should we alleviate problem with disposable objects?

R0ma-D · 2017-05-12T18:51:32Z

I do not quite understand the problem. What will be unexpected for the end user? May be CancellationToken as input parameter decide problem? Or could you provide an example of d.DoAsync() method?

R0ma-D · 2017-05-12T19:32:49Z

I've attach new benchmark result:

Code with BenchmarkDotNet:
`public class Program
{
private const int DelayInMilliseconds = 1;
private const int Size = 10000;

    static void Main(string[] args)
    {
        BenchmarkRunner.Run(typeof(Program), new MethodInfo[] 
        {
            typeof(Program).GetMethod(nameof(ExecuteWithoutAsyncBenchmark)),
            typeof(Program).GetMethod(nameof(ExecuteWithAsyncBenchmark)),
        });
        Console.ReadKey();
    }

    [Benchmark]
    public void ExecuteWithAsyncBenchmark()
    {
        Measure(() => ExecuteWithAsync());
    }

    [Benchmark]
    public void ExecuteWithoutAsyncBenchmark()
    {
        Measure(() => ExecuteWithoutAsync());
    }
    
    public static void Measure(Func<Task> creator)
    {
        Task[] all = new Task[Size];
        for (int i = 0; i < Size; i++)
            all[i] = creator();
    }

    [Benchmark]
    public static async Task ExecuteWithAsync()
    {
        await Task.Delay(DelayInMilliseconds);
    }

    [Benchmark]
    public static Task ExecuteWithoutAsync()
    {
        return Task.Delay(DelayInMilliseconds);
    }
}`

aensidhe · 2017-06-07T15:58:45Z

So, I finally did the tests. Results shows no improvement on excluding async-await. Sorry, I had a lot of work to do on main job.

Code for non-async-await is at #106. I look forward where can we improve code more, but I think network exchange will remove any difference here.

Async-await

BenchmarkDotNet=v0.10.7, OS=Windows 10 Redstone 2 (10.0.15063)
Processor=Intel Core i5-4590 CPU 3.30GHz (Haswell), ProcessorCount=4
Frequency=3215209 Hz, Resolution=311.0218 ns, Timer=TSC
dotnet cli version=1.0.0
  [Host] : .NET Core 4.6.25211.01, 64bit RyuJIT
  Core   : .NET Core 4.6.25211.01, 64bit RyuJIT

Job=Core  Runtime=Core

Method	Mean	Error	StdDev
Call	438.6 us	13.10 us	37.59 us

Without async-await

BenchmarkDotNet=v0.10.7, OS=Windows 10 Redstone 2 (10.0.15063)
Processor=Intel Core i5-4590 CPU 3.30GHz (Haswell), ProcessorCount=4
Frequency=3215209 Hz, Resolution=311.0218 ns, Timer=TSC
dotnet cli version=1.0.0
  [Host] : .NET Core 4.6.25211.01, 64bit RyuJIT
  Core   : .NET Core 4.6.25211.01, 64bit RyuJIT

Job=Core  Runtime=Core

Method	Mean	Error	StdDev
Call	448.1 us	16.68 us	47.86 us

R0ma-D · 2017-06-08T21:32:09Z

It looks like you are right and all possible performance improvement spends by network communication. Thank you for your efforts.

aensidhe · 2017-07-12T11:23:08Z

Since we have improvements on our network stack in 0.8.0, I'll redo the tests this week, I hope.

aensidhe · 2017-11-08T11:34:36Z

Test for one call. 2200 RPS.

BenchmarkDotNet=v0.10.9, OS=Windows 10.0.16299
Processor=Intel Core i5-4590 CPU 3.30GHz (Haswell), ProcessorCount=4
Frequency=3215210 Hz, Resolution=311.0217 ns, Timer=TSC
.NET Core SDK=2.0.0
  [Host]     : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT
  netcore1.1 : .NET Core 1.1.2 (Framework 4.6.25211.01), 64bit RyuJIT
  netcore2.0 : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT

Jit=RyuJit  Platform=X64  Runtime=Core  
Toolchain=CoreCsProj

Method	Job	Mean	Error	StdDev	Median	P95	Scaled	ScaledSD	Gen 0	Allocated
Redis	netcore1.1	349.0 us	6.956 us	9.045 us	347.0 us	364.4 us	1.00	0.00	-	472 B
Tarantool	netcore1.1	322.2 us	4.919 us	4.601 us	320.8 us	328.9 us	0.92	0.03	1.4648	3256 B
Redis	netcore2.0	347.3 us	22.424 us	65.055 us	313.4 us	447.1 us	1.00	0.00	-	472 B
Tarantool	netcore2.0	429.8 us	11.462 us	13.644 us	427.1 us	453.5 us	1.28	0.22	1.0231	3256 B

Batch=100. 61k RPS.

BenchmarkDotNet=v0.10.9, OS=Windows 10.0.16299
Processor=Intel Core i5-4590 CPU 3.30GHz (Haswell), ProcessorCount=4
Frequency=3215210 Hz, Resolution=311.0217 ns, Timer=TSC
.NET Core SDK=2.0.0
  [Host]     : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT
  netcore1.1 : .NET Core 1.1.2 (Framework 4.6.25211.01), 64bit RyuJIT
  netcore2.0 : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT

Jit=RyuJit  Platform=X64  Runtime=Core  
Toolchain=CoreCsProj

Method	Job	Mean	Error	StdDev	P95	Scaled	ScaledSD	Gen 0	Gen 1	Allocated
Redis	netcore1.1	646.7 us	14.65 us	42.51 us	715.3 us	1.00	0.00	6.8359	-	21.5 KB
Tarantool	netcore1.1	1,801.1 us	71.58 us	209.92 us	2,255.3 us	2.80	0.37	125.0586	0.3516	291.98 KB
Redis	netcore2.0	677.9 us	13.36 us	30.15 us	724.2 us	1.00	0.00	6.8359	-	21.51 KB
Tarantool	netcore2.0	1,597.3 us	29.87 us	31.96 us	1,641.7 us	2.36	0.12	120.2539	1.2305	291.98 KB

Batch=1000, 90815 RPS

BenchmarkDotNet=v0.10.9, OS=Windows 10.0.16299
Processor=Intel Core i5-4590 CPU 3.30GHz (Haswell), ProcessorCount=4
Frequency=3215210 Hz, Resolution=311.0217 ns, Timer=TSC
.NET Core SDK=2.0.0
  [Host]     : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT
  netcore1.1 : .NET Core 1.1.2 (Framework 4.6.25211.01), 64bit RyuJIT
  netcore2.0 : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT

Jit=RyuJit  Platform=X64  Runtime=Core  
Toolchain=CoreCsProj

Method	Job	Mean	Error	StdDev	P95	Scaled	ScaledSD	Gen 0	Gen 1	Allocated
Redis	netcore1.1	622.8 us	12.43 us	33.40 us	670.6 us	1.00	0.00	6.8359	-	21.51 KB
Tarantool	netcore1.1	2,129.7 us	72.78 us	210.00 us	2,499.1 us	3.43	0.38	125.0000	1.2277	291.98 KB
Redis	netcore2.0	3,977.7 us	104.78 us	304.00 us	4,526.4 us	1.00	0.00	70.0605	11.3407	230.84 KB
Tarantool	netcore2.0	10,615.1 us	210.71 us	321.77 us	11,011.3 us	2.68	0.21	958.6397	426.0110	2914.63 KB

aensidhe · 2017-11-08T12:12:43Z

BenchmarkDotNet=v0.10.9, OS=Windows 10.0.16299
Processor=Intel Core i5-4590 CPU 3.30GHz (Haswell), ProcessorCount=4
Frequency=3215210 Hz, Resolution=311.0217 ns, Timer=TSC
.NET Core SDK=2.0.0
  [Host]     : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT
  netcore1.1 : .NET Core 1.1.2 (Framework 4.6.25211.01), 64bit RyuJIT
  netcore2.0 : .NET Core 2.0.0 (Framework 4.6.00001.0), 64bit RyuJIT

Jit=RyuJit  Platform=X64  Runtime=Core  
Toolchain=CoreCsProj

Method	Job	Mean	Error	StdDev	P95	Scaled	ScaledSD	Gen 0	Gen 1	Gen 2	Allocated
Redis	netcore1.1	3.899 ms	0.0759 ms	0.1159 ms	4.081 ms	1.00	0.00	76.6369	7.4405	-	236254 B
Call	netcore1.1	11.544 ms	0.2244 ms	0.2304 ms	11.875 ms	2.96	0.10	986.3782	482.7724	-	2984584 B
Select	netcore1.1	NA	NA	NA	NA	?	?	N/A	N/A	N/A	N/A
Redis	netcore2.0	3.671 ms	0.0732 ms	0.0685 ms	3.791 ms	1.00	0.00	72.9167	14.0625	-	236213 B
Call	netcore2.0	10.655 ms	0.1861 ms	0.1741 ms	10.952 ms	2.90	0.07	952.0833	429.1667	-	2984584 B
Select	netcore2.0	11.069 ms	0.1976 ms	0.1848 ms	11.349 ms	3.02	0.07	1532.7381	733.2589	89.2857	4480584 B

Benchmarks with issues:
IncrementBatchBenchmark.Select: netcore1.1(Jit=RyuJit, Platform=X64, Runtime=Core, Toolchain=CoreCsProj)

Unused async/await pairs are removed to optimize and reduce compiled …

9abc30c

…code

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Unused async/await pairs are removed to optimize and reduce compiled code #104

Unused async/await pairs are removed to optimize and reduce compiled code #104

R0ma-D commented Apr 3, 2017

aensidhe commented May 10, 2017

R0ma-D commented May 10, 2017

aensidhe commented May 12, 2017 •

edited

Loading

R0ma-D commented May 12, 2017

R0ma-D commented May 12, 2017

aensidhe commented Jun 7, 2017 •

edited

Loading

R0ma-D commented Jun 8, 2017

aensidhe commented Jul 12, 2017

aensidhe commented Nov 8, 2017

aensidhe commented Nov 8, 2017

Unused async/await pairs are removed to optimize and reduce compiled code #104

Are you sure you want to change the base?

Unused async/await pairs are removed to optimize and reduce compiled code #104

Conversation

R0ma-D commented Apr 3, 2017

aensidhe commented May 10, 2017

R0ma-D commented May 10, 2017

aensidhe commented May 12, 2017 • edited Loading

R0ma-D commented May 12, 2017

R0ma-D commented May 12, 2017

aensidhe commented Jun 7, 2017 • edited Loading

R0ma-D commented Jun 8, 2017

aensidhe commented Jul 12, 2017

aensidhe commented Nov 8, 2017

aensidhe commented Nov 8, 2017

aensidhe commented May 12, 2017 •

edited

Loading

aensidhe commented Jun 7, 2017 •

edited

Loading