I spoke at the Charlottesville .NET User Group this week and at the Raleigh Code Camp. I cheated and did the same presentation to both groups. Call me lazy but, in the middle of planning our own Code Camp in Richmond, I really didn't want to prepare two separate talks. I did a talk back at CodeStock 2009 on a similar topic back in June 2009 but it's evolved a lot since then based on my own growth and understanding. You can find the code and slides below.
This talk takes about an hour and covers the basic concepts of metaprogramming and how the DynamicMetaObject class, the IDynamicMetaObjectProvider interface and the DynamicObject helper class work in the .NET Framework 4.0. The talk is very demo-centric, focusing on the code that support the concepts outlined in a few slides. The first demo focuses on helping you to understand the dispatch mechanism in the DLR and how you can implement your own classes that participate in the call binding process. It's a fairly simple fluent XML parser implemented using the DynamicObject helper class. While it's simplistic, it helps to introduce developers who are new to the DLR to the concepts of dispatch and binding in C# before diving into things Pythonic.
The second demo focuses on performance measurement. Having learned about the delegate and rule caching that goes on in the DLR's CallSite<T> and ActionBinder classes in the presentation, these tests show the effect of all that awesome work that Microsoft has done for us to enhance the performance of dynamic code. The result is that we see the caching yields about 5,000 times the performance of traditional static to dynamic language interoperability. Rather than the hundreds of transactions per second that we are accustomed to when doing dynamic language invocation from a context like C#, we now get millions of transactions per second. To put it all in context, the same transactions performed in pure C# are still about 4 times as fast as the DLR with delegate and rule caching. However, you have to admit that what the DLR provides today is pretty amazing. For many, this level of performance makes the use of languages like Python for scripting a large application feasible for the first time.
The question I most often get about this is, "When will the dynamic language performance be just as good as 'real' .NET code?" OK, so I take exception to the term 'real' there but I know what you all mean. I don't know the answer and it doesn't really matter that Ruby and Python performance be just as good as C#. Different languages have their own strengths and sometimes that includes performance. But when pure C# generates 6 million transactions per second, traditional thunking across to dynamic languages performs at a rate of several hundred per second, the fact that adaptive Inline Caching in the DLR helps me to yield close to 2 million transactions per second is pretty darned good. We're well on the way to my other favorite language, Python, becoming a first-class citizen in the .NET family. It takes education, of course, for people to know just how wonderful the DLR is which is the real purpose of my evangelism. I just love the DLR, if you can't tell already.
I am doing a Channel 9 geekSpeak next week on this topic and I'll refer to some of this material. And my talk at the October 2009 NoVA Code Camp will use an even further evolved version of this talk. Drop me a note and let me know what you think about this topic. Do you have success stories? Horror stories regarding dynamic language integration?
Slides for DLR Performance Talk September 2009 [PPTX] (184 kB)
Slides for DLR Performance Talk September 2009 [PDF] (680 kB)
Code for DLR Performance talk September 2009 [Updated 21 Sept 2009] (16 kB)