onsdag 12. januar 2011

Move from .Net Remoting to WCF

When .Net 3.5 and Windows Communication Foundation was released I was very - very suprised that Microsoft simply dropped the binary remoting support completely in favour for interoperability. I had imagined that there was a space for applications which use the same types on both server and client side in WCF as well. There is nothing in the full name of WCF that indicates that this is for service orientation only. In fact, I still think that WCF should embrace different types of means for serializing messages between two or more endpoints. Hopefully such support will be added later as I guess .Net Remoting will not removed for a few more years. Regarding migrating from .Net Remoting to WCF then there is a good article here which I read a few years ago:
http://msdn.microsoft.com/en-us/library/aa730857(VS.80).aspx

Some weeks ago I decided to put together a small sample which serializes the same types through .Net Remoting and WCF. I wanted to see how much worse WCF performes compared to .Net Remoting, at least that was my working hypothesis. First it was only a shallow User object and then I made the object graph being serialized a bit more complex by adding Groups and Permissions and so forth to at least get a realistic graph from an OO perspective.

///


/// TODO: Document me!
///

/// Steinar Dragsnes
[Serializable]
public class User : IEntity
{
private string firstName;
private string lastName;
private string userName;
private string password;
private int id;

public object Payload { get; set; }

List groups = new List();
.....


///
/// TODO: Document me!
///

/// Steinar Dragsnes
[Serializable]
public class Group
{
private int id;
private string name;

private User administrator;
private List members;
private List permissions;
.....

///
/// TODO: Document me!
///

/// Steinar Dragsnes
[Serializable]
public class Permission
{
private int id;
private string name;
.....


In terms of size, this graph is still very small, but again it was able to illustrate some short comings of the DataContractSerializer's default settings: shared, bi-directional, many to many or circular references are not by default supported when creating the XML to be serialized. But this is not a show stopper, some googling turned up this page: http://social.msdn.microsoft.com/Forums/en-US/wcf/thread/f30ecd17-cac0-4cdc-8142-90b5f411936b/
which demonstrates both how to preserve references and stick in the NetDataContractSerializer. I moved the code snippets into the sample solution and now it works with WCF, but of course it cannot compare with .Net Remoting in terms of performance (also demonstrated by the sample).

So based on my little research on WCF, here are some problems with the current migration process from .Net Remoting to WCF if standard conventions are being used:
1. No serializer that skips building XML before serializing the object graph, creating the XML consumes time.
2. When the XML finally has been streamed, then in most cases where one has a real OO model, the stream size is bigger then that of .Net Remoting.
3. The lack of a LegacyRemotingSerializer that would take care of satisfying all features supported by .Net Remoting out of the box. Interoperability would of course not be possible.
4. References are not automatically preserved, work arounds exist though.
5. Open generics are not supported, not a big issue, but should be supported by a LegacyRemotingSerializer.
6. There should be a minimal impact of change in existing code when migrating from .Net Remoting to WCF.
7. Attribute usage should be avoided, decorating the service interfaces are ok, Spring.Net has exporters that does this for you so that your service interfaces can still stay poco.
8. In general a painful process, should be painless!
9. The end result is degraded performance.


Here is the output from the sample:
....now both channels have had their first serializations and everything is completely initialized....

Time used during serialization: 2.0002
Time used during wcf serialization: 7.0007
Time used during serialization: 1.0001
Time used during wcf serialization: 68.0068
Time used during serialization: 1.0001
Time used during wcf serialization: 3.0003
Time used during serialization: 1.0001
Time used during wcf serialization: 4.0004
Time used during serialization: 1.0001
Time used during wcf serialization: 27.0027
Time used during serialization: 1.0001
Time used during wcf serialization: 4.0004
Time used during serialization: 2.0002
Time used during wcf serialization: 5.0005
Time used during serialization: 1.0001
Time used during wcf serialization: 5.0005
Time used during serialization: 1.0001
Time used during wcf serialization: 5.0005
Time used during serialization: 1.0001
Time used during wcf serialization: 4.0004
Time used during serialization: 1.0001
Time used during wcf serialization: 5.0005
Time used during serialization: 1.0001
Time used during wcf serialization: 6.0006
Client Remoting and WCF tests is complete...



While reading more about the NetDataContractSerializer I realized that there is a bug in the configuration of the WCF infrastructure and that additional custom code must be written to remove all references to the DataContractSerializer as the strategy chosen for building the xml to be binary serialized.

///
/// TODO: Document me!
///

/// Steinar Dragsnes
public class WcfServiceEndpointConfigurer
{
private WcfSerializationMode mode;
private WcfDataContractType contractType;

///
/// Default ctor setting up the instance as default to produce WCF services as .Net Remoting services.
///

public WcfServiceEndpointConfigurer() : this(WcfSerializationMode.AsNetRemoting, WcfDataContractType.NetDataContractSerializer) { }

///
/// General ctor where the invoker may configure how the services should be set up.
///

/// Serialization mode or protocol; net.tcp, http, named pipes etc
/// Data contract type used to build the xml to be serialized.
public WcfServiceEndpointConfigurer(WcfSerializationMode mode, WcfDataContractType contractType)
{
this.mode = mode;
this.contractType = contractType;
}

///
/// Configures the service end point so that it can receive requests from clients.
///

public void ConfigureServerEndPoint()
{
ServiceHost host = new ServiceHost(typeof(T));
if (mode == WcfSerializationMode.AsNetRemoting) WcfAsNetRemoting(host);

host.Open();
}

///
/// Based on the uri, create a transparent proxy casted to type T and create communication channel to establish connection.
///

/// The type of the instance to cast the transparent proxy to.
/// The uri used to create a channel for connecting to the server endpoint.
/// An instance of type T which can be interacted with in a request-response fashion.
public T ConfigureClientEndPoint(string uri)
{
ChannelFactory factory = new ChannelFactory(uri);
Binding unknownBinding = factory.Endpoint.Binding;
if (unknownBinding is NetTcpBinding) factory.Endpoint.Binding = FixNetTcpBinding((NetTcpBinding)unknownBinding);

// This is the only way to globaly switch out DataContractSerializer.
if (contractType == WcfDataContractType.NetDataContractSerializer) SwitchToNetDataContractSerializer(factory.Endpoint);

return factory.CreateChannel();
}

// If the configuration file is completely misconfigured(or not configured for WCF running as .Net Remoting :), then this method should not do anything.
private void WcfAsNetRemoting(ServiceHost host)
{
foreach (ServiceEndpoint endPoint in host.Description.Endpoints)
{
Binding unknownBinding = endPoint.Binding;
// If binding should be NetTcp to get as close as .Net Remoting as possible, then we need to fix a lot of constraints on the binding
// so that we can serialize object graphs of some decent size.
if (unknownBinding is NetTcpBinding) endPoint.Binding = FixNetTcpBinding((NetTcpBinding) unknownBinding);

// This is the only way to globaly switch out DataContractSerializer.
if (contractType == WcfDataContractType.NetDataContractSerializer)
{
SwitchToNetDataContractSerializer(endPoint);
}
}
}

private void SwitchToNetDataContractSerializer(ServiceEndpoint endPoint)
{
foreach (OperationDescription desc in endPoint.Contract.Operations)
{
DataContractSerializerOperationBehavior dcsOperationBehavior = desc.Behaviors.Find();
if (dcsOperationBehavior != null)
{
NetDataContractSerializerOperationBehavior serializer = new NetDataContractSerializerOperationBehavior(desc);
serializer.MaxItemsInObjectGraph = Int32.MaxValue;
int idx = desc.Behaviors.IndexOf(dcsOperationBehavior);
desc.Behaviors.Remove(dcsOperationBehavior);
desc.Behaviors.Insert(idx, serializer);
}
}
}

// Setting most data-transport constraints on the tcpBinding to their respective maximum limits.
private NetTcpBinding FixNetTcpBinding(NetTcpBinding binding)
{
binding.ReaderQuotas = new XmlDictionaryReaderQuotas
{
MaxArrayLength = Int32.MaxValue,
MaxBytesPerRead = Int32.MaxValue,
MaxNameTableCharCount = Int32.MaxValue,
MaxDepth = Int32.MaxValue,
MaxStringContentLength = Int32.MaxValue,
};

binding.MaxReceivedMessageSize = Int32.MaxValue;
binding.MaxBufferPoolSize = long.MaxValue;
binding.MaxBufferSize = Int32.MaxValue;
return binding;
}
}


With this in place, then I saw that I could serialize the same object graphs in WCF much faster then in .Net Remoting! How is this possible? The NetDataContractSerializer even builds up an xml document before serializing the stream of data? To turn this question around, maybe the old BinaryFormatter has a sloppy implementation and there are a lot of possible improvements and easy wins which would have made it much faster? Shouldn't it be more difficult to create a xml representation of an object graph than a binary one?

Anyhow, I wondered if this speed difference would change if I created a really, really deep graph, hundreds of objects deep. When the test was in place I started the sample application but the WCF service crashed. The error message said something about message size or object depth exceeding the configured limit. Some additional googling on the error message pinpointed me to a solution where one increases the various settings on the NetTcpBinding as seen from the code excerpt above. With this small modification I am able to serialize really deep object graphs. This is perfect for old .Net Remoting applications which move big object graphs. Even though you won't have interoperability, you may at least move into a supported framework and maybe enjoy some of the tools for monitoring WCF services.

Here is a print of the speed differences:

Remoting serialization cost in ms: 147.0147
-===-WCF serialization cost in ms: 200.02 -===-
Remoting serialization cost in ms: 291.0291
-===-WCF serialization cost in ms: 245.0245 -===-
Remoting serialization cost in ms: 420.042
-===-WCF serialization cost in ms: 349.0349 -===-
Remoting serialization cost in ms: 658.0658
-===-WCF serialization cost in ms: 466.0466 -===-
Remoting serialization cost in ms: 800.08
-===-WCF serialization cost in ms: 574.0574 -===-
Remoting serialization cost in ms: 1074.1074
-===-WCF serialization cost in ms: 643.0643 -===-
Remoting serialization cost in ms: 1332.1332
-===-WCF serialization cost in ms: 700.07 -===-
Remoting serialization cost in ms: 1500.15
-===-WCF serialization cost in ms: 789.0789 -===-
Remoting serialization cost in ms: 1850.185
-===-WCF serialization cost in ms: 981.0981 -===-
Remoting serialization cost in ms: 2024.2024
-===-WCF serialization cost in ms: 1038.1038 -===-
Remoting serialization cost in ms: 2299.2299
-===-WCF serialization cost in ms: 1284.1284 -===-
Remoting serialization cost in ms: 2520.252
-===-WCF serialization cost in ms: 1342.1342 -===-
Client Remoting and WCF tests is complete...
Press Space to end process...

The interesting thing about this printout is that the object graph is getting deeper and deeper as the root object (an instane of a User) is requested, a number of objects are being added to the user before the request returns. The user instance on the server side will thus grow and get larger and larger, deeper and deeper and thus more difficult to move. Both Remoting and WCF has their own user which they call repeatedly. The impressive thing when reading this output is that WCF has a slight time increase as the root object's footprint increases, while .Net Remoting's time increase is that of WCF multiplied with 2 to 3. In the end, WCF is about twice as fast as .Net Remoting. Don't you believe me? Please download the sample and see for youself!

For additional tips or comments about how to make this blog post better, please contact me.