Just code - Tamir Khason : Performance, PLINQ

New year – new blog or how to migrate Community Server to any other engine, supports XML-RPC

Tamir Khason — Fri, 02 Jan 2009 02:33:04 GMT

Please update your bookmarks, because the new url of this blog is http://khason.net/ (you have not update RSS feeds, it will be done automatically). Why I did it? Why I decided to go to “stand-alone”… Well. there are some reasons. Generally, I do not want to explain all those here, but trust me, there are some. The main reason is, that there is no responsible person in charge for this blog platform in Microsoft Israel. This why, if your blog is popular and you have a respect to your blog visitors, you cannot host it here… Take a look into new comments notifications in my inbox. Would you answer your readers with such “small amount” of SPAM and capcha, that cannot be fixed already for three years in this platform? This how my inbox looks like for last three years. So now, you should not ask me, why I not answered your email or comments. Aren’t you? :)

Currently, all comments in this blog are disabled, so if you want to comment, please use new url of posts (this will appear shortly in the beginning of each post). Also, this post will not be syndicated in RSS.

So, this post is the last. I loved this platform, and loved people started it. But, unfortunately, it seemed, that bloggers community is not important enough for new platform managers.

Also, If you want to learn how to use C# and XML-RCP to migrate from CS2007, used in this platform, visit my new home :)

Have a great year and, as always, be good people. This post marked with all possible tags automatically. Sorry.

Just released: Parallel Extensions – June CTP

Tamir Khason — Tue, 03 Jun 2008 00:32:02 GMT

[This blog was migrated. You will not be able to comment here.
The new URL of this post is http://khason.net/blog/just-released-parallel-extensions-%e2%80%93-june-ctp/]

Today, Parallel Programming team released 2nd CTP for Parallel Extensions to .NET Framework 3.5. Major features in this CTP are declarative parallelist (LINQ-to-Object, LINQ-to-XML), parallel tasks and imperative data and, the most important (for me) – user mode work stealing scheduler. It makes very efficient use of parallel hardware. Additional information regarding this release can be found in team blog. Also, please, report bugs to Connect in order to help those guys to produce better product.

Download June CTP of Parallel Extensions to .NET 3.5 >>

Parallel programming? Well, it’s all about CPU affinity or how to set processor affinity in WPF

Tamir Khason — Mon, 14 Apr 2008 19:57:08 GMT

[This blog was migrated. You will not be able to comment here.
The new URL of this post is http://khason.net/blog/parallel-programming-well-it%e2%80%99s-all-about-cpu-affinity-or-how-to-set-processor-affinity-in-wpf/]

Parallel computing is very cool technology, that makes you able to leverage tasks between processors in your system. Today it’s already impossible to buy single processor computer – when you’ll buy new PC, you’ll probably get dual core CPU at least. Today we’ll speak about how to manage affinities of tasks between CPUs in your system. For this purpose we’ll create a small game, named CoreWars. So, let’s the show begin.

What we want to do? Simple game, that includes some complicated math, executed large number of times. Here the code of one Dice roll

void SetDice()
        {
            List<Dice> dices = new List<Dice>();
            for (int i = 0; i < MaxTries; i++)
            {
                double x = rnd.NextDouble();
                double y = rnd.NextDouble();
                double g = (Math.Pow(x, 2) + Math.Pow(y, 2) - rnd.Next()) / (Math.Sqrt(2) * y + Math.Sqrt(2) * ((x - Math.Sqrt(2)) - 1));
                dices.Add((Dice)((int)g & 0x3));
            }
            int mP = dices.Count(dice => dice == Dice.Paper);
            int mR = dices.Count(dice => dice == Dice.Rock);
            int mS = dices.Count(dice => dice == Dice.Scissors);

            int m = (mP > mR) ? mP : mR;
            m = (m > mS) ? m : mS;

            CurrentDice = (Dice)((m & 0x3) % 0x3);
        }

Don’t even try to understand what’s going on here. The result is one dice state Paper, Rock or Scissors (the old kinds game). Now we want to run this number of times and affine each thread to one of our system’s core. How to do it?

Simple way – ask what thread are you running, then use ProcessorAffinity property of the ProcessThread to your CPU id.

ProcessThread t = Process.GetCurrentProcess().Threads.OfType<ProcessThread>().Single(pt => pt.Id == AppDomain.GetCurrentThreadId());
t.ProcessorAffinity = (IntPtr)(int)cpuID;

Very good, but what’s this warning about “System.AppDomain.GetCurrentThreadId()' is obsolete: 'AppDomain.GetCurrentThreadId has been deprecated because it does not provide a stable Id when managed threads are running on fibers (aka lightweight threads). To get a stable identifier for a managed thread, use the ManagedThreadId property on Thread. http://go.microsoft.com/fwlink/?linkid=14202 ” ? Why this happens to me? The simple answer is Thread Processor Affinity never was most reliable thing in Windows (as well as ThreadID). But we still want to make it. What to do?

Let’s go a bit unmanaged. We’ll use SetThreadAffinityMask and GetCurrentThread WinAPI methods to achieve what we want to.

[DllImport("kernel32.dll")]
static extern IntPtr GetCurrentThread();
[DllImport("kernel32.dll")]
static extern IntPtr SetThreadAffinityMask(IntPtr hThread, IntPtr dwThreadAffinityMask);
…
SetThreadAffinityMask(GetCurrentThread(), new IntPtr(1 << (int)cpuID));
…
SetThreadAffinityMask(GetCurrentThread(), new IntPtr(0));

So far, so good. Now threads. This step is very trivial

ParameterizedThreadStart pts = new ParameterizedThreadStart(dr.ThrowDices);
Thread t = new Thread(pts);
t.SetApartmentState(ApartmentState.MTA);
t.IsBackground = true;
t.Start(i);

We need also context switch (you remember, WPF)

ThreadPool.QueueUserWorkItem(delegate
            {
                int w = WaitHandle.WaitAny(wcs);
                Dispatcher.Invoke(DispatcherPriority.Background, (SendOrPostCallback)delegate(object o)
                {
                    tb.Text = string.Format("The winner is CPU #{0}", o);
                }, w);
            });

Now, what to do with binding? To Bind or not to Bind – that is the question! My answer is to bind! Remember, you need also a small piece of CPU time to process UI events and push frames in renderring thread. Let’s roll dices.

SetThreadAffinityMask(GetCurrentThread(), new IntPtr(1 << (int)cpuID));

            App.Current.Dispatcher.Invoke(DispatcherPriority.Background, (SendOrPostCallback)delegate
            {
                base.Clear();

                if (CollectionChanged != null)
                    CollectionChanged(this, new NotifyCollectionChangedEventArgs(NotifyCollectionChangedAction.Reset, null));
            }, null);

            for (int i = 0; i < MaxIteractions; i++)
            {
                DiceRoller dr = new DiceRoller();
                App.Current.Dispatcher.BeginInvoke(DispatcherPriority.Background, (SendOrPostCallback)delegate
                   {
                       base.Add(dr);

                       if (CollectionChanged != null)
                           CollectionChanged(this, new NotifyCollectionChangedEventArgs(NotifyCollectionChangedAction.Add, dr));
                   }, null);
            }
            SetThreadAffinityMask(GetCurrentThread(), new IntPtr(0));
            ResetEvent.Set();

Also do not forget to implement INotifyCollectionChanged on result collection and INotifyPropertyChanged on dice object.

After all preparations done, we can start to build basic user interface for the presentation layer. Main window

<StackPanel Name="Root">
        <StackPanel Name="LayoutRoot" Orientation="Horizontal"/>
        <Button Content="Start battle" Click="Button_Click"/>
        <TextBlock Name="tb" TextAlignment="Center" FontSize="15" Visibility="Collapsed"/>
    </StackPanel>

Templates

<DataTemplate DataType="{x:Type l:DiceRoller}" x:Key="drt">
            <Rectangle Width="5" Height="5">
                <Rectangle.Style>
                    <Style>
                        <Style.Triggers>
                            <DataTrigger Binding="{Binding Path=CurrentDice}" Value="Paper">
                                <Setter Property="Rectangle.Fill" Value="Red"/>
                            </DataTrigger>
                            <DataTrigger Binding="{Binding Path=CurrentDice}" Value="Rock">
                                <Setter Property="Rectangle.Fill" Value="Black"/>
                            </DataTrigger>
                            <DataTrigger Binding="{Binding Path=CurrentDice}" Value="Scissors">
                                <Setter Property="Rectangle.Fill" Value="Blue"/>
                            </DataTrigger>
                        </Style.Triggers>
                    </Style>
                </Rectangle.Style>
            </Rectangle>
        </DataTemplate>
        <ItemsPanelTemplate x:Key="wpt">
            <WrapPanel/>
        </ItemsPanelTemplate>

We done. the only task to perform is to detect the number of CPUs in the system and put appropriate number of content controls to show the race result. Do not forget binding too.

pc = Environment.ProcessorCount;

            for (int i = 0; i < pc; i++)
            {
                HeaderedContentControl hcc = new HeaderedContentControl();
                hcc.Width = this.Width / pc;
                hcc.HorizontalAlignment = HorizontalAlignment.Stretch;
                hcc.Header = string.Format("CPU {0}",i);

                ItemsControl ic = new ItemsControl();
                ic.ItemsPanel = Resources["wpt"] as ItemsPanelTemplate;
                ic.ItemTemplate = Resources["drt"] as DataTemplate;
                hcc.Content = ic;
                LayoutRoot.Children.Add(hcc);

                DiceRollers dr = new DiceRollers();
                this.Resources.Add(string.Format("roller{0}", i),dr);

                Binding b = new Binding();
                b.Source = dr;
                b.IsAsync = true;
                b.BindsDirectlyToSource = true;
                b.Mode = BindingMode.OneTime;
                ic.SetBinding(ItemsControl.ItemsSourceProperty, b);
            }

That’s all. Now we can race our dices and see what CPU works better. For real life scenario, this sample looks absolutely stupid, but just thing about processor affinity in Parallel Extension CTP – why not to make us able to use it? The answer is – K.I.S.S (keep it simple, stupid). And as for me, I do not like to be stupid and I want to manage affinities! I need high performance for my applications and only me can decide to which processor dispatch which task and how.

Have a nice day and be good people.

Source code for this article.

CoreWars.zip [61.4 Kb]

How to: High performance graphics in WPF

Tamir Khason — Mon, 03 Mar 2008 00:38:00 GMT

[This blog was migrated. You will not be able to comment here.
The new URL of this post is http://khason.net/blog/how-to-high-performance-graphics-in-wpf/]

Microsoft DPE: "WPF is ever best super performance oriented technology for creating hyper multi point graphs, using parallelism and huge in-memory persistence vector scrounged math distributed calculations... And ever more with branded new Microsoft Windows Vista 7.

Client: Errr, well.... Let's try to apply it for our VB program...

DPE: You can easily do it yourself, but it'd be better to call someone from Microsoft Consulting Services.

Client: Let it be...

MCS: Well. It's too huge for WPF to scale... WPF uses a retained rendering system. It saves every little pixel and make you able scale and repaint very often without the composition system blocking on callbacks to your code. However, 1,000,000 retained pixels is too huge to scale...

Client: I want it scale. They promised... They told, it'll scale. Make it to do what I want it to do!!!

MCS: Errr, well. Let it be!

This is very common dialog between DPE, MCS and clients. Sales men want it to do, what he need it to do. Client want it to do what sales men promised to do and Services men should make it to do what they both want it to do. Today we'll speak about retained, lazy and loose models to produce large scale graphics.

First problem: multithreading

Even before we start to work, you should know, that we cannot create dependency objects in thread other, then current UI thread. We can use locks, mutexes, semaphores, however we still can not create Dependency Objects in other thread. In order to get rid of it, we'll have to use INofityPropertyChanged implementation, instead of Dependency Objects. This means, no Dependency Properties.

So, we'll start with following code (I'll reuse nice code written by Luis Diego Fallas to create Mandelbrot Fractal set)

class FractsCollection : INotifyPropertyChanged
{

Second problem: rendering thread

Well, the problem is knows. There is only one UI thread. We wont it be only one, so we'll use our own HostVisual by Dwayne Need to enhance the performance.

Third problem: Retained objects

Actually, this is not problem. This is feature. And it can be extremely useful if you want to retain layout. Let's create a simple example: Dragon curve fractal. It has limited number of points, generated by well known final algorithm. So, we'll create our own geometry, derived from Shape class. The fasted geometry is StreamGeometry, so let's use it. First of all let's create the class and save the array of points.

public class DragonShape:Shape
    {
StreamGeometry dragonGeometry;
        double _angle;
        List<Point> Points;

Then we'll generate the pattern

void GeneratePattern()
        {
            ThreadPool.QueueUserWorkItem(delegate
            {
Move(5);
Turn(GetNextPoint * System.Math.PI / 180.0);

Then, by overriding DefiningGeometry property, create the fractal

protected override System.Windows.Media.Geometry DefiningGeometry
        {
            get {
                using (StreamGeometryContext context = dragonGeometry.Open())
                {
                    context.BeginFigure(Points[0], false, false);
                    context.PolyLineTo(Points, true, false);
                }
                return (Geometry)dragonGeometry.GetAsFrozen();}
        }

Don't forget to tell the shape, that geometry was changed

this.Dispatcher.BeginInvoke(System.Windows.Threading.DispatcherPriority.Background, (SendOrPostCallback)delegate
                    {
                        this.InvalidateVisual();
                    }, null);

Now, run it and we'll have very nice vector fractal generated that can be easily resized and scaled. Here the result.

This method will work fine for 1,000, even 10,000 points. But after a while you'll experience performance degradation. What to do? The client wants 10,000,000 (!) points (and in Winforms GDI+ it works for him)

Let's try to understand why. Because it is not retain. It's image! So, let's use image to make the play fair.

The fastest BitmapSource is InteropBitmap. It has an ability to update itself from the memory section. That's exactly what we'll use

format = PixelFormats.Bgr32;
            max = format.BitsPerPixel;
            uint count = (uint)(sWidth * sHeight * format.BitsPerPixel / 8);
            section = CreateFileMapping(new IntPtr(-1), IntPtr.Zero, 0x04, 0, count, null);
            map = MapViewOfFile(section, 0xF001F, 0, 0, count);
            pixels = new byte[count];
            Marshal.Copy(pixels, 0, map, (int)count);
            source = System.Windows.Interop.Imaging.CreateBitmapSourceFromMemorySection(section, (int)sWidth, (int)sHeight, format, (int)(sWidth * format.BitsPerPixel / 8), 0) as InteropBitmap;
            ThreadPool.QueueUserWorkItem(delegate
            {
                while (true)
                {
                    Generate();
                }
            });

To get the source to bind to we'll get frozen image. Call Invalidate first to reread the source.

InteropBitmap source;
        public BitmapSource Source
        {
            get
            {
                source.Invalidate();
                return (BitmapSource)source.GetAsFrozen();
            }
        }

Now, when we ready to display we can just put pixels simultaneously (by using Parallel extension and PLINQ) and tell the WPF that our count and ImageSource property updated upon each pixel.

unsafe
            {
                uint* pBuffer = (uint*)map;
                Parallel.For(0, (int)sHeight, delegate(int yi)
                {
                    foreach (var p in from xi in Enumerable.Range(0, (int)sWidth).AsParallel()
                                      let mappedX = xF(xi)
                                      let mappedY = yF(yi)
                                      let p0 = new TranslatePoint(xF(xi), yF(yi))
                                      let function = constructor(p0)
                                      select new
                                      {
                                          x = xi,
                                          y = yi,
                                          xD = mappedX,
                                          yD = mappedY,
                                          i = apply(function, p0)
                                                  .TakeWhile(
                                                     (x, j) => j < max && x.NormSquared() < 4.0)
                                                  .Count()
                                      })
                    {
                        pBuffer[(int)(p.x + p.y * sWidth)] = (uint)(uint)((uint)0xFF << 24) |
                             (uint)(p.i << 16) |
                             (uint)(5*p.i << 8) |
                             (uint)(15*p.i); ;
                        count++;
                        FireUpdate();
                    }
                });
            }

void FireUpdate()
        {
                if (PropertyChanged != null)
                {
                    PropertyChanged(this, new PropertyChangedEventArgs("PixelCount"));
                    PropertyChanged(this, new PropertyChangedEventArgs("Source"));
                }
        }

We done. Now let's see how fast it can be to generate and display 10,000,000 live pixels (it's about 16,000x16,00x32bit image) in screen. The maximum, I was able to get with my Dell Latitude D820 was 1,200,000 x 1,200,000 pixels indexed 4 bit image (it's about 100,000,000 points) and my memory is over :)

Not bad, ah? So WPF scales and DPE are right? Not quite right, but let them to do their work and we'll be behind to come all client's dreams (and DPE's promises) true.

Have a nice day and be good people. Now you can use WPF for drawing big number of points.