Threading and Throttling differences between WCF and ASMX
I know I havent been posting deep technical stuff that I used to do. Contrary to what people think my current role entails, keeping abreast of the technology landscape is what I am supposed to do and what I enjoy and so when colleagues joked with me when was the last time I booted up Visual Studio, for example, I enjoyed seeing their shocked faces when I told them: "oh - just last night. why you asked ?" i-wink.
I dont dwell deep like I used to but I still code up decent projects which I implement within my own developmental testing environment (yes, I have one running latest versions of Active Directory, SharePoint, Exchange and all the other goodies), driving the houshold crazy when I think of new and different ways to document expenses, publish a Book or CD library, home automation projects using all sorts of different technologies (yes, that includes Ruby-on-Rails) or in my new Windows Mobile 6 Device. Of course, I admit I dont post topics deep like I used to. It is not so much the content but more so, the limiting factor of time.
Recently, I was involved in some internal technical discussions with regards to the issue of scale comparisions between _WCF_ and ASMX. Below are some discussions:
If you have a web service that is going to be IO bound, you would definitely want it to be scalable and almost every resource in the world tells you to implement ASP.NET asynchronous pattern (BeginSomething/EndSomething, etc) on it so to go easy on the thread pool. ASP.NET uses an IAsyncHttpHandler to handle the request, which means the worker threads are not blocked while the IO-bound operation executes somewhere else. Sounds good so far.
If you make a WCF version of it with webHttpBinding (which actually means you can invoked it AJAX-style) following the same async pattern for the methods, you may find that each invocation of the WCF service eats up two threads – one for its ASP.NET HttpModule.ProcessRequest and the other for the actual IO. Ouch! You may think that this means your WCF implementation may end up eating all threads reserved for ASP.NET, which would indeed scale down the server
Is this true OR are we missing the complete picture ?
While the scenarios explained above are reasonable observations, it doesnt paint the complete picture. WCF does perform better scalability than ASMX.
- Threading:
For ASMX, when a request comes in, it would be queued up immediately for async ASMX. So the thread is released for that request and a new thread will pick up the work item later.
For WCF, when a request comes in, we queue it up in WCF and let an IO thread handle the request. At the same time, the request thread-pool thread is held to wait for the request to complete.
Yes, WCF uses more threads than async ASMX. But there is a reason for this. Using asynchronous ASMX is dangerous and not really a good practice (and I have hinted at this many times in the many Web Service/ASMX presentations I have done over the past few years). While it does well at what it is supposed to do, it does trick the developer into a "false sense of security". Essentially, if you dont know how the ASP.NET blackbox works, you may find yourself thrown against the car wall when you take a hidden, unsuspecting corner at high speeds. It does not provide enough throttling for client loads. Basically the server takes all items and queue them up for later processing. The server does not have a good throttling mechanism to control the number of work items. To everyone else, it seems that the server is quite friendly to all clients. However, if the number of clients is unbounded, this is really bad. First of all, the server working set would grow unlimited due to unlimited requests queued up. Secondly, many client requests would become obsolete when it’s picked up by the server from the queue. The latter accounts for a a good set of problematic scenarios I have come across in my past consulting gigs with regards to high-load and high-transactional ASMX asynchronous implementations before I joined the borg.
Think of it as a side of the brain (that tells you that you are about to be full) not functioning properly when you sit down at a buffet table. You eat and eat and eat without knowning when to stop and then your ingestion/digestion system starts kicking in, you actually hit the wall. Hard. Literally.
- Server Throughput
When you measure scalability, the most important measurement is the server throughput. That is, how many requests the server can handle per time unit? For async ASMX, it would be pretty fast at the initial phase. However, like the ingestion/digestion analogy I was referring to above - Once the server is in a steady phase (as when CPU is fully loaded), the throughput will go down because the server capacity has reached. You can compare the data between async ASMX and sync ASMX over the long run to see what I mean.
Also you would see higher memory usage of the async approach.
- ASP.NET Throttling
That said, ASP.NET does have a throttling mechanism that is used for sync ASMX, which is the threadpool thread limit. The number of threads used to handle requests are bounded (http://support.microsoft.com/kb/821268). WCF uses this fact to throttle incoming requests. You can always change the configuration settings to increase number of threads to be used to allow more work items to be queued up.
The max number of threads follows the following formula:
MaxWorkerThreads x #CPU – MinFreeThreads
This is 12 by default on a single-proc machine.
- Two-level Throttling for WCF
WCF leverages the ASP.NET threadpool throttling to throttle client requests. At the same time, WCF has its own item queue throttling. The former is throttled by the setting mentioned in the immediate above point, while the latter is controlled by WCF throttling settings (maxConcurrentCalls etc). ASP.NET can automatically adjust threads based on CPU loads so that you would always get full load of the server.
In this way, you may experience client failures because the requests are rejected at ASP.NET layer beforehand. So you can increase the ASP.NET throttling to get better experience. But eventually you would still be bounded by the physical server capacity, no matter whether you use async ASMX, sync ASMX, or WCF as mentioned above.
There is improvement work done in .NET 3.0 SP1 and of course, .NET 3.5 (beta 2 here), with the use of prioritized item queues. Do expect even-better WCF performance even in some of the common scenarios. Fine tuning minWorkerThreads will even give us even better results.
Thanks to Wenlong for helping out with the guidance and explanation. The complete scenario and the design principles for it will be published in greater detail in a MSDN whitepaper later. Do watch out for it.
