XmlSerializer - the first deserialization is very slow

Question

I have a solution with two projects; an asp.net MVC application, and a class library. Let's call them project MVC and project CLS.

In the project CLS, there are two different versions (V1 and V2) of an XSD file that I have used to create two serializable classes with the same name, but under different namespaces (V1 and V2) using xsd2code.

In the MVC project, when the user uploads an XML file, the CLS.dll is used to deserialize the XML into an object. When the XML file is of type V1, the deserialization is very fast, but the XSD file for the V2 version is a lot more complex, and the deserialization can take up to a couple of minutes, only the first time (it's very fast afterwards, until the application is run again).

I used the Sgen.exe tool to create a serializer assembly (CLS.XmlSerializers.dll) for the CLS.V2 type in order to eliminate the first-time creation of the assembly on the fly, and therefore improving the performance.

I have successfully managed to add the Sgen Task to the Post Build events, and the assembly CLS.XmlSerializers.dll is created every time I build the project. Also, I have used the unit test code in this post to make sure the assembly is loaded, and it does. The test passes susscessfully.

However, still, the first time the XML file is deserialized, it takes a long time. So, something still should be wrong. But, I don't know what. Please help.

UPDATE:

I used Fuslogvw.exe as was suggested in the comments, and I can see that the CLS.XmlSerializers.dll is being loaded successfully. Then, how come the first time the XML file is deserialized it takes around one minute, but every time after that takes less than a second?

UPDATE 2:

One of the differences between the two XSD files is that the second one (V2) has a reference to a very big XSD file that containes definitions of some xs:enumeration types that are used in the main file. And, that's the reason the deserialization took a long time. Since all I need to do is to deserialize the XML files into objects and do not need to validate the values of the attributes and elements against those enumerations, I ended up removing the reference to that XSD file, and replacing all the enumeration types with their base types (in this case, xs:string). Now, V2 is deserialized as fast as V1, and I don't even need to use Sgen.exe. I guess Sgen.exe only helps in situations where you need to deserialize a very large XML file. In my case, the XML files are always very small, but the desrialization is (was) complex.

Answer 1

In order to increase performance of XML serialization, assemblies are dynamically generated each time XmlSerializer is instantiated for the first time for a specific type. It happens only once in the application lifetime, but that makes its first usage slow.

When you instantiate an XmlSerializer you have to pass the Type of the objects that you will attempt to serialize and deserialize with that serializer instance. The serializer examines all public fields and properties of the Type to learn about which types an instance references at runtime. It then proceeds to create C# code for a set of classes to handle serialization and deserialization using the classes in the System.CodeDOM namespace. During this process, the XmlSerializer checks the reflected type for XML serialization attributes to customize the created classes to the XML format definition. These classes are then compiled into a temporary assembly and called by the Serialize() and Deserialize() methods to perform the XML to object conversions.

Full Content: Troubleshooting Common Problems with the XmlSerializer

More Info: XmlSerializer Constructor Performance Issues

Answer 2

It is a known issue of x64 jit compiler, it can be very slow in some cases. That's why you have much better performance when running the deserializtion the second time when code is already compiled.

Try to use .net 4.6 or higher, it features a new version of x64 jit compiler (RyuJIT). If it is not possible to update .net version then take a look at this thread.