Twelve Ways to Improve WPF Performance

There is no shortage of information out there on how to speed up the performance of WPF applications, but too often the focus is on the weird stuff instead of the simpler issues. I’m not going to talk about things like writing to WritableBitmaps to optimize drawing—it’s a topic covered to death elsewhere. Instead, this is meant to be a slightly more practical guide to squeezing performance out of WPF in ways that are probably more likely affecting you.

Some general notes

ItemsControl and its subclasses ListBox and ListView exacerbate performance problems because these controls are highly dynamic (resolution happens “late”), involve WPF collections (which are slow), and have difficult and unpredictable lifetimes for their child controls. Scrollbar performance is often a big problem in larger WPF apps because of problems that seem trivial for small collections, but suddenly blow up with larger data sets.

Also, it can be difficult in WPF to know exactly when the system is finished with an object. For views, you get the FrameworkElement.Unloaded event, but it gets raised at times you might not expect (such as system theme changes) and not at times when you might (application shutdown). On viewmodels associated with views, you’ll never get a WPF notification that a viewmodel is about to go unused by a view. Blend-style behaviors also have their own set of lifetime problems.

Then there are some problems (like this and this) where WPF leaks for you too.

Finally, there are things (this, this, this, this, this, and this) that simply perform worse than you likely expect.

Finally, there are old UI/WinForms problems (this, this, and this) that never really went away—they’re just less likely to happen.

  1. Fix Binding Errors
  2. Hard-code widths and heights where possible
  3. Avoid CollectionView.Grouping
  4. Optimize bindings to collections that change
  5. Avoid DynamicResources
  6. Avoid ResourceDictionary
  7. Simplify your visual tree
  8. Be wary of System.Windows.Interactivity.Behavior<T>.OnDetaching
  9. Do not use DependencyPropertyDescriptor for any reason…ever
  10. Be careful of viewmodel events
  11. Batch up Dispatcher.BeginInvoke
  12. In general, beware of memory leaks

I. Fix Binding Errors and Exceptions

Every time a binding error occurs, your app hangs for just a split second as it writes out errors to the trace log. If you have a lot of binding errors, then those split seconds start to add up. Make sure to go through your bindings, especially those on ItemsControls (ListViews, custom grids, etc.) and verify that there are no binding errors.

Open up your app in the debugger and play around, especially where there is slowness. Make sure all bindings resolve without errors.

RelativeSource in DataTemplates may also result in bindings that break initially, but then later resolve properly. Be wary of them, and try to use inherited attached properties instead of relying on RelativeSource in DataTemplates.

Viewmodel bindings

  1. Make sure that your views and view models are in sync. Use ReSharper 6 to help you find broken bindings.
  2. If you’re binding to a collection of objects with mixed types, add different DataTemplates so that none of them refer to non-existent properties.
  3. Make sure that your converters aren’t throwing exceptions. These have a cost too.

View-based RelativeSource bindings

  1. When using ListBoxes and ListViews, it’s a common problem to have this problem. Avoid RelativeSource.FindAncestor expressions at all cost here, because the deferred behavior of templates cause the object and its bindings to be created (and resolved) before the ListBoxItem/ListViewItem is added to the visual tree.
  2. An alternative is to define an attached dependency property on the ListBoxItem/ListViewItem, and use property inheritance to give your child items the necessary property values. This essentially pushes property values down the visual tree instead of searching up.

II. Hard-code sizes where possible

This may not always be a practical or desirable solution, but layout passes perform faster when widths and heights do not have to be recalculated. They may also help stop a layout pass from rippling through an entire visual tree.

And always set specific widths on columns in a grid (be it a ListView + GridView or any third-party control), because these tend to be very expensive, especially with larger data sets.

III. Avoid CollectionView.Grouping

Grouping in WPF doesn’t perform terribly well, especially with ListViews and GridViews. Create a collection with mixed viewmodel types–your original collection, and one that represents the “group”. Use DataTemplates to change the appearance of your “group” objects.

For example, if you have a PersonViewModel class with a property that you want to group by (let’s say Region), it is faster to create a mixed collection of MyGroupViewModel and PersonViewModel objects, ordered correctly by group, with different DataTemplates, than it is to bind to a grouped collection. Unfortunately, it’s a lot more work.

IV. Optimize bindings to collections that change

Repeatedly calling ObservableCollection<T>.Add when the collection is data-bound can be a prohibitively expensive operation, especially with thousands of rows. Unfortunately, the framework provides no easy, satisfactory fix.

Fix 1: Use ObservableCollection as-is, but break bindings

  1. Break the binding to the collection.
  2. Update the collection while not data-bound.
  3. Re-bind.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
// some methods removed for brevity
public partial class MyViewModel : INotifyPropertyChanged
{
    private ObservableCollection<T> _people;
 
    public IList People
    {
        get { return _people; }
        private set
        {
            if (_people != value)
            {
                _people = value;
                OnPropertyChanged("People");
            }
        }
    }
 
    void BatchAddPeople(IEnumerable<Person> newPeople)
    {
        var currentPeople = _people;
 
        // stop WPF from listening to the changes that we're about
        // to perform
        this.People = null;
 
        // change
        foreach (var person in newPeople)
        {
            currentPeople.Add(person);
        }
 
        // cause WPF to rebind--but only once instead of once for
        // each person
        this.People = currentPeople;
    }
}

Fix 2: Use the older .NET 2.0-era collections instead

  1. Use System.ComponentModel.BindingList<T> (from the old days) instead; it has an API for suppressing change notifications.

Fix 3: Reimplement ObservableCollection.

  1. Create your own collection that implements INotifyCollectionChanged.
    • Raise INotifyCollectionChanged as sparingly as you can.
    • Raise the event with a NotifyCollectionChangedAction.Reset event for anything more trivial than a simple single-item add, remove, change, or move. Do not take advantage of the NotifyCollectionChangedEventArgs constructors that take collections of items; you will find support for it spotty at best.
  2. Implement System.Collections.IList on your collection. WPF does not use the generic System.Collections.Generic.IList<T> interface; it is completely ignored. If you don’t implement the interface, WPF will perform almost all operations (including accessing rows by number!) strictly by the IEnumerable implementation, and it won’t be very optimal or fast about it.
  3. (Probably) implement System.Collections.Generic.IList<T> as well. WPF doesn’t use it, but you probably will (through LINQ, Rx, etc.)

V. Avoid DynamicResources

Even in .NET 4.0, DynamicResource access is still slower than StaticResource access. And worse, once you start nesting DynamicResources (for example, a ListView whose Style contains a ControlTemplate that references objects through DynamicResources), you’re likely to run into situations where you leak controls.

VI. Avoid ResourceDictionary

This advice is practically impossible to follow, but do your best. There is a huge cost in constructing ResourceDictionaries, and depending on where you place them, you are probably constructing many more objects than you realize.

A common, sensible, and logical pattern is to keep usages of elements as close to where you use them as possible. Many people place resources in UserControl.Resources, or break up their themes into multiple ResourceDictionaries for clarity and separation. Although this is arguably good programming practice, it also tends to be insanely slow. If your windows/controls or your ListBoxItem/ListViewItems in a ListBox/ListView are coming up more slowly than you would like, it’s probably a combination of too much ResourceDictionary construction and/or DynamicResources. (Yes, even in .NET 4.0.) Collapse ResourceDictionaries as much as you can, and try to ensure that the contents of these dictionaries is only loaded once. The easiest and surest way is to include it in the resources of your System.Windows.Application object, almost always difficult or infeasible for composite applications.

I have also frequently taken to creating static classes that contain nothing but highly reusable resources (think static framework classes like the Brushes class) because it’s easier to guarantee that objects are only being created once, and hopefully at app startup instead of triggered by the user interacting with the application and forcing a lazy load at an undesirable time. Not necessary the healthiest design, but the performance is quite a bit better.

Using implicit ControlTemplate/DataTemplate styles will also help keep your code and XAML organized without the need for either StaticResource or DynamicResource.

VII. Simplify your visual tree

Shallow visual trees are better than deeper visual trees. Again, ItemsControls will usually exacerbate performance problems with deep visual trees because if they’re not being virtualized, they’re being destroyed and recreated; if they are being virtualized, changing DataContext in a deeper tree takes more time than changing DataContext in a shallower one.

VIII. Be wary of System.Windows.Interactivity.Behavior<T>.OnDetaching

Sadly, System.Windows.Interactivity.Behavior<T>.OnDetaching will generally not get called. Put a breakpoint and see for yourself.

OnAttached signals the addition of a behavior to a control (generally at instantiation of your XAML); OnDetaching signals the removal of a behavior from a control (generally never, as behaviors don’t get removed from controls). Don’t put sensitive disposing behavior in OnDetaching. The Unloaded event is a better place for that, but be aware that it will get raised every time the control is removed from the visual tree.

IX. Do not use DependencyPropertyDescriptor for any reason…ever

DependencyPropertyDescriptor.AddValueChanged classes cause the WPF framework to take a strong reference to the source of the event that isn’t removed until you call DependencyPropertyDescriptor.RemoveValueChanged. This class is frequently used in conjunction with Behaviors, so if you have RemoveValueChanged in OnDetaching, you’re likely leaking memory. Because of WPF’s references to your objects, it is not just enough to drop references to your view and view model.

A better alternative is to rely on data binding where you can; create a DependencyProperty for the sole purpose of listening to changes on your target property, and use the change notifications in DependencyProperty in order to listen to changes on the target property. It keeps the observed object (generally, your view model) from accidentally holding a strong reference to your view.

X. Be careful of view model events

If your views or behaviors rely on events being raised from a viewmodel (as innocuous as INotifyPropertyChanged.PropertyChanged or INotifyCollectionChanged.CollectionChanged), subscribe to them weakly. Viewmodels tend to have a longer lifetime than views (consider a virtualized ItemsControl), so it’s possible that your view model will inadvertently gather references to views. Use classes like PropertyChangedEventManager or CollectionChangedEventManager, or (painfully) use the WeakEventManager to create your own event manager for your custom events. It’s painful, but usually necessary in order to prevent view models from taking references to views.

XI. Batch up Dispatcher.BeginInvoke

If your application displays data from a network, you’re probably using background threads to accomplish the task (which is good). However, you’re probably not consciously counting your Dispatcher.BeginInvoke calls (which is not as good). The more you’re able to coalesce multiple Dispatcher.BeginInvokes into a single call, the more likely WPF will be able to help you out by making single subsequent layout and render passes, and the lower your overall CPU usage.

To be fair, WPF is much better at trying to help you here than VB6/WinForms—you won’t often see apps that dance and flicker incessantly any more—but there is still a non-zero cost to updating the screen, and if you have a particularly large application, this could be a problem for you.

XII. In general, beware of memory leaks

This is a bit of a generalization of the last few points, but memory leaks make apps behave worse over time. Fixing memory leaks goes a long way in fixing the performance of an application. And since most developers are constantly restarting WPF apps as they work on them, they often go undetected until the software is delivered.

—DKT

Collections, CollectionViews, and a WPF Binding Memory Leak

A colleague of mine recently came across some code that should work, but doesn’t:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
private ListBox SomeListBox;
private ObservableCollection<string> SomeStrings = new ObservableCollection<string>();
 
public void SomeMethod()
{
    SomeListBox.ItemsSource = null;
    ThreadPool.QueueUserWorkItem(o => AddToStrings());
}
 
private void AddToStrings()
{
    // this line will actually crash the app!
    SomeStrings.Add("blah");
 
    Dispatcher.BeginInvoke(new Action(() => SomeListBox.ItemsSource = SomeStrings);
}

So basically, detaching an ObservableCollection<T> from a view, adding an element on a background thread, and then reattaching it back on the UI thread again. Generally, view-bound objects can only be modified from the UI thread, but shouldn’t detaching the collection first should make that a non-issue?

When you actually run the code, you get one of those obnoxious “InvalidOperationException: The calling thread cannot access this object because a different thread owns it” messages, and then your app will probably die. These usually happen because you’re trying to modify a control or an object bound to a control, neither of which applies here. It also happens when you’re trying to do anything with a subclass of DispatcherObject or a DependencyObject on the wrong thread, but that doesn’t apply here either, right? ObservableCollection<T> does not have any inherent thread affinity, and it is definitely safe to pass between threads so long as access to it is synchronized.

So we took the control out of the picture by writing SomeListBox.ItemsSource = null, and ObservableCollection<T> can be modified from any thread. What’s the problem? Those are the only two objects in the picture, right?

ICollectionView

(If you already know what ICollectionView is, feel free to skip to the end of the story.)

Actually, when any IEnumerable is bound to an ItemsControl, an implementation of ICollectionView is created to wrap the collection, and it provides a means for sorting, grouping, and filtering without affecting the original collection. So when you bind a ListBox, or ListView, or the WPF Toolkit DataGrid, or some other third-party ItemsControl to a grid, a CollectionView is created specifically to maintain intermediate state on behalf of the control without having to touch the original collection.

Consider some List<string>:

Index Value
0 Cheese
1 Apple
2 Tomato
3 Salad

We bind it to a DataGrid, and then we tell the DataGrid to sort itself. The list should look like this:

Index Value
0 Apple
1 Cheese
2 Salad
3 Tomato

Note how the objects now have different indices. It’s an important detail—WPF manages to accomplish sorting, grouping, and filtering all without modifying the original collection because of the use of a CollectionView (or one of its subclasses ListCollectionView, BindingListCollectionView, or the internal CollectionViewProxy class). It also pulls off some shiny tricks occasionally, batching changes where appropriate and postponing actual CollectionView reordering (which can obviously get expensive) until subsequent pumps of the current Dispatcher. And that’s why CollectionView inherits from DispatcherObject.

And of course, if our collection changes, the CollectionView listens for those notifications and responds appropriately, causing hilarious consequences when the change happens on a background thread.

The Fix

The implementation of CollectionView is somewhat broken when collections implement INotifyCollectionChanged:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
internal CollectionView(IEnumerable collection, int moveToFirst)
{
    /* some code snipped for brevity */
    INotifyCollectionChanged changed = collection as INotifyCollectionChanged;
    if (changed != null)
    {
        // this is the problem
        changed.CollectionChanged += this.OnCollectionChanged;
 
        /* some more code */
    }
 
    /* still more code */
}

Nowhere anywhere is this event handler unwired. So collections essentially retain strong references of their corresponding CollectionViews. The data binding system that is responsible for creating the CollectionView maintains a weak reference, but as long as the original collection is alive, the data binding system will be hooked into the CollectionView, and therefore the original collection as well.

The fix actually isn’t so bad:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
using System.Collections;
using System.Collections.ObjectModel;
using System.Collections.Specialized;
using System.Windows;
using System.Windows.Data;
 
public class BetterObservableCollection<T> :
    ObservableCollection<T>, ICollectionViewFactory
{
    public ICollectionView CreateView()
    {
        return new BetterListCollectionView(this);
    }
}
 
public class BetterListCollectionView :
    ListCollectionView, IWeakEventListener
{
    public BetterListCollectionView(IList list) : base(list)
    {
        INotifyCollectionChanged changed = list as INotifyCollectionChanged;
        if (changed != null)
        {
            // this fixes the problem
            changed.CollectionChanged -= this.OnCollectionChanged;
            CollectionChangedEventManager.AddListener(list, this);
        }
    }
}

As soon as the collection is detached from the view, the view’s reference to the CollectionView is broken, and the CollectionView is allowed to disappear because nothing (neither the original collection nor the binding system) is holding on to a strong reference any more.

What “The Fix” Breaks

The CollectionView is responsible for managing sorting, grouping, and filtering. If you disconnect the collection from the view, the CollectionView will get garbage-collected, and the sorting, grouping, and/or filtering you had on the control will be lost. That may or may not matter for your particular use case, but it’s definitely something to think about. If that is important, it may make more sense to bind to a System.ComponentModel.BindingList<T>, which tracks its own grouping, sorting, and filtering inside the list itself—the tradeoff is then you lose the ability to have multiple views on the same collection. —DKT