I have a WCF net.tcp service hosted with the builtin ServiceHost, and when doing some stress tests I get a strange behavior. The first time i send a bunch of requests, 5 to 10 requests are answered quickly, and the rest are returning at about 2 second intervals. The second time i send the requests, 10 - 20 are returned quickly, and rest with 2 sencond intervals.
The above repeats until I can get over 100 requests returned quickly, but if I wait a minute or so the memory usage of the service goes down and the requests go back to 5-10 returning quick.
The service I am testing has a small delay, so that I can get many open connections at the same time, if this delay is removed the requests return so quickly that i have perhaps 2-5 connections open at the same time. This delay is to simulate DB connections and other outgoing stuff.
From the behavior it looks like the ServiceHost is allocating something, threads, class instances, but I can not figure out what it is.
I could have a timer in the client that calls the service to keep it working, but that seems like a bad solution.
If I have a high sustained load to the service it will crunch all requests quickly, but if I have a period of low activity and then a surge of connections comes in the service will be slow.
I guess my question is WHAT is it the get allocated during high load of the WCF service, and HOW can I configure the service to preallocate more of the things that get allocated.
EDIT:
I did some more testing, and looking at the taskmgr for the process I can see that when the servicehost is 'resting' there are 10 threads open, but when I start sending requests, the threadcount goes up. As long as the threadcount is high the servicehost can process incoming requests quickly, but if I pause sending the requests, the open threadcount decreases, and subsequent requests starts taking longer time to process.
Now, how can I tell the servicehost to keep a bunch of threads open? Or more than the 10-12 that it keeps by default?
Well, after lots of googling, it seems that the problem is the threadpool. The CLR threadpool allocates a few threads, and when they are used, it throttles the creation of new threads, and after a time it also deallocates unused threads.
There is some confusion about a bug that meant that the ThreadPool did not honor the SetMinThreads call.
http://www.michaelckennedy.net/blog/PermaLink,guid,708ee9c0-a1fd-46e5-8fa0-b1894ad6ce0f.aspx
I am not sure if this bug is solved, or what, because when I modify the ThreadPool settings, the problem persists.
The thing that determines how may request are handled simultaneous is the ServiceThrottlingBehavior. There are a number of different threasholds that will limit the amount of request being processed. This also depends on the binding your are using, for example wsHttpBinding defaults to sessions on while basicHttpBinding uses no sessions and the default session limit of 10 is no problem.
See http://msdn.microsoft.com/en-us/library/ms735114.aspx for more details.
The bug you referenced is fixed in .NET 3.5 SP1. That may have had something to do with the problem, I think it's more likely (much more likely) that throttling is your problem rather than thread as Maurice keyed into.
<system.serviceModel>
<service name="???" >
<endpoint ... />
</service>
</system.serviceModel>
What's the throttle limit for this "empty" config? 10 session, 16 concurrent calls! Beware.
Here's more on the threading:
http://www.michaelckennedy.net/blog/2008/08/20/ThreadPoolBugInNET20SP1IsFixed.aspx
This feels like a hack but it seems to solve your issue. The problem is that the threadpool will take time to start up a new thread, so what you really need is threads waiting on standby. Add a constructor to your service and set the minimum number of threads you would like.
public YourService()
{
int workerThreads;
int portThreads;
ThreadPool.GetMinThreads(out workerThreads, out portThreads);
ThreadPool.SetMinThreads(200, portThreads);
}
Related
Please excuse the Obvious Self-Q/A, but this information is widely misunderstood, and almost always incorrectly answered. So I Wanted to place this information here for people searching for a definitive answer to this problem.
Even so, there's still some information I haven't been able to nail down. I will put this towards the end of the question (skip to that if you are not interested in the preamble).
How do I correctly configure a WCF NetTcp Duplex Reliable Session?
There are many questions and answers regarding this topic, and nearly all of them suggest setting inactivityTimeout="Infinite" in your configuration. This doesn't really seem to work correctly, particularly for the case of NetTcp (It may work correctly for WSDualHttp Bindings, but I have never used those).
There are a number of other issues that are often associated with this: Including, Channel not faulting after client or server unexpectedly disconnected, Channel disconnecting after 10 minutes, Channel randomly disconnecting... Channel throwing exception when trying to open... Unable to configure Metadata on same endpoint...
Please note: There are two concepts that are important below. Infrastructure messages are internal to the way WCF communicates, and are used by the framework to keep things running smoothly. Operation messages are messages that occur because your app has done something, like send a message across the wire. Infrastructure messages are largely invisible to your app (but they still occur in the background) while operation messages are the result of an action your app has taken.
Information I have figured out, through hard won trial and error.
Infinite does not appear to be a valid configuration setting in all situations (and certainly, the visual studio validation schema doesn't know about it).
There are two special configuration converters, called InfiniteIntConverter and InfiniteTimeSpanConverter which will sometimes work to convert the value Infinite to either Int.MaxValue or TimeSpan.MaxValue, but I haven't yet figured out the situations in which this appears to be valid as sometimes it works, and sometimes it doesn't. What's more, it appears that some libraries will allow Infinite in the config, while others will not, so you can succeed in one part of a configuration, but fail in another.
You must configure BOTH inactivityTimeout and receiveTimeout, on both the client and the server. While these values do not HAVE to be the same, they probably should be as they will probably cause confusion if they are not. (technically, you can leave inactivityTimeout to its default value if you want, but you should be aware of its value, and what it does)
inactivityTimeout should NEVER be set to a large value, much less Infinite or TimeSpan.MaxValue.
inactivityTimeout has two functions (and this is not widely understood). The first function defines the maximum amount of time that can elapse on a channel without receiving any "infrastructure" or "operation" messages. The second function defines the time period in which infrastructure messages are sent (half the time specified). If no infrastructure or operation messages have been received during the timeout period, the connection is aborted.
receiveTimeout specifies the maximum amount of time that can elapse between operation messages only. This value can be set to a large value, such as TimeSpan.MaxValue (particularly if your channel runs internally over a trusted network or over a vpn). This value is what defines how long the reliable session will "stay alive" if there is no activity between client and server (other than infrastructure messages). ie, your client does not call any methods of the interface, and your server does not call back into the client.
setting a short inactivityTimeout and a large receiveTimeout keeps your reliable session "tacked up" even when there is no operational activity between your client and server. The short inactivity timeout (i like to keep the default 10 minutes or less) sends infrastructure "ping" messages to keep the TCP connection alive while the long receive timeout keeps the reliable session active. while at the same time providing a reasonable timeout in case of disconnection.
If you set inactivityTimeout to a large value, then the reliable session will not be reliable as it has no way to keep the Tcp connection alive, nor does it have any way to verify the integrity of the connection. It won't know if a user has disconnected unexpectedly until you try and send a message to that client and find out the connection is no longer there. This is why many people who use Infinite for this setting resort to creating a "Ping" method in their service, which is completely unnecessary if you've configured these settings correctly.
If you set inactivityTimeout to a value larger than receiveTimeout then it will likewise also be unreliable, as you will still be governed by the receiveTimeout for operation messages. ie. if you forget to set receiveTimeout and leave it at the default 10 minutes, then if the user is idle for 10 minutes, the connection will be aborted.
When the client or server unexpectedly disconnects (app crashes, network failure, someone trips over the power cord, etc..), the other side may not notice right away. I have attached various ChannelFaulted event handlers in various test situations, and sometimes the connection is faulted right away... other times it doesn't seem to fault at all. What i have discovered through trial and error is that the when it doesn't seem to fault, it will actually fault after the inactivityTimeout expires on that end. (so if it's set to 10 minutes, then after 10 minutes it will call the ChannelFaulted event).
I have not yet figured out why in some situations it notices the disconnection right away, and others it waits for the timer to expire. In both cases, I notice internal first chance communication exceptions thrown and handled by the framework, and there are calls to Abort the connection... but somehow the call to the event handler gets lost and it must wait for the timeout. My suspicion is this is somehow thread related.
When trying to configure Metadata to work across the NetTcp channel, I have had sporadic results. Sometimes it works, sometimes it doesn't. I've read many reports that Metadata doesn't work over NetTcp and that you have to use an Http channel for the Metadata, but I have in fact had it work on several occasions using the net.tcp:// url to generate the proxy. Then I would change something, recompile and it would no longer work. Changing things back, it wouldn't work again. So it was very confusing what magic incantation was necessary to make Metadata function over net.tcp, shared with the endpoint on the same port (obviously with a different address).
When configuring both a NetTcp and Metatdata endpoint on the same service, and specifying non-default settings for connection parameters like listenBacklog, and maxConnections, you also need to make sure the Metadata endpoint uses the same settings, which typically means you have to define a custom binding, since these settings are not available from the standard tcp mex binding. This includes setting listenBacklog and maxPendingConnections on tcpTransport, and groupName and maxOutboundConnectionsPerEndpoint on connectionPoolSettings.
The default setting for the Ordered setting of ReliableSession is True. This uses a lot more overhead than turning it off. If you don't need ordered messages, i would suggest turning it off (need to set this on both sides)
-
Configuration I still need to understand:
How do I correctly configure the shared net.tcp Metadata endpoint? (I will add an example when I get a chance) Currently, i'm specifying an http get url to bypass the problem. It's so inconsistent as to why it sometimes works and sometimes does not. I kept getting the error `The URI Prefix is not recognized' when generating the proxy in Visual Studio.
Why does WCF sometimes Fault the channel immediately upon disconnect, and sometimes waits for inactivityTimeout to expire? What controls/causes one vs the other behavior?
I'm trying to nail down a performance issue under load in an application which I didn't build, but have become very familiar with the workings of.
The architecture is: mobile apps call an ASP.NET MVC 3 website to get data to display. The ASP.NET site calls a third-party SOAP API using WCF clients (basicHttpBinding), caching results as much as it can to minimize load on that third party.
The load from the mobile apps is in the order of 200+ requests per second at peak times, which translates to something in the order of 20 SOAP requests per second to the third-party, after caching.
Normally it runs fine but we get periods of cascading slowness where every request to the API starts taking 5 seconds.. then 10.. 15.. 20.. 25.. 30.. at which point they time out (we set the WCF client timeout to 30 seconds). Clearly there is a bottleneck somewhere which is causing an increasingly long queue until requests can't be serviced inside 30 seconds.
Now, the third-party API is out of my control but they swear that it should not be having any issues whatsoever with 20 requests per second. So I've been looking into the possibility of a bottleneck at my end.
I've read questions on StackOverflow about ServicePointManager.DefaultConnectionLimit and connectionManagement, but digging through the source, I think the problem is somewhat more fundamental. It seems that our WCF client object (which is a standard System.ServiceModel.ClientBase<T> auto-generated by "Add Service Reference") is being stored in the cache, and thus when multiple requests come in to the ASP.NET site simultaneously, they will share a single Client object.
From a quick experiment with a couple of console apps and spawning multiple threads to call a deliberately slow WCF service with a shared Client object, it seems to me that only one call will occur at a time when multiple threads use a single ClientBase. This would explain a bottleneck when e.g. 20 calls need to be made per second and each one takes more than 50ms to complete.
Can anyone confirm that this is indeed the case?
And if so, and if I switched to every request creating it's own WCF Client object, I would just need to alter ServicePointManager.DefaultConnectionLimit to something greater than the default (which I believe is 2?) before creating the Client objects, in order to increase my maximum number of simultaneous connections?
(sorry for the verbose question, I figured too much information was better than too little)
What is "GridInterceptingMessageHandler"? I did a search and I can find no mention of this on nservicebus.com. Also, I see the samples have the line:
.LoadMessageHandlers(First<GridInterceptingMessageHandler>.Then<SagaMessageHandler>())
What does that do exactly?
If you look at the source and its documentation you'll see the following:
Intercepts all messages, not allowing any through if the endpoint has had its number of worker threads reduced to zero.
GridInterceptingMessageHandler
NSB allows you to dynamically tune the number of work threads and endpoint is using to process messages. If the number of work threads has been reduced to zero, the endpoint becomes disabled and will not continue to process messages. The tuning of threads is useful if you would like to increase the speed of message processing(assuming everything else will scale as well) while not having to restart the endpoint.
This is especially helpful if you want to slowing drain the system of messages so that you can perform upgrades or other maintenance duties. By default this is wired up for you, you would only reference it if you decided to override how the message handlers are loaded(as in the example).
Is there something in WCF configuration that defines a timeout for executing a request at service side? E.g. WCF service will stop executing request after some time period. I have a service which make some work depending on client input. In some cases a such call may take too much time. I want to limit the execution time of such requests on service side, not client one using SendTimeout. I know about OperationTimeout property, but it doesn't abort the service request, it just tells a client that the request is timed out.
In general terms, there's nothing that will totally enforce this. Unfortunately, it's one of those things that the runtime can't really enforce nicely without possibly leaving state messed up (pretty much the only alternative for it would be to abort the running thread, and that has a bunch of undesirable consequences).
So, basically, if this is something you want to actively enforce, it's a lot better to design your service to deal with this so that your operation execution has safe interruption points where the operation can be terminated if the maximum execution time has been exceeded.
Though it's a lot more work, you'll likely be more satisfied with it in the long run.
I'm using perfmon to examine my service behaviour. What I do is I launch 6 instances of client application on separate machines and send requests to server in 120 threads (20threads per client application).
I have examined counters and maximum number of instances (I use PerSession model and set number of instances to 100) is 12, what I consider strange as my response times from service revolve around 120 seconds... I thought that increasing number of instances will cause WCF to create more instances, and as a result response times would be quicker.
Any idea why WCF doesn't create even more instances of service?
Thanks Pawel
WCF services are throttled by default - it's a service behavior, which you can tweak easily.
See the MSDN docs on ServiceThrottling.
Here are the defaults:
<serviceThrottling
maxConcurrentCalls="16"
maxConcurrentInstances="Int.MaxValue"
maxConcurrentSessions="10" />
With these settings, you can easily control how many sessions or concurrent calls can be handled, and you can make sure your server isn't overwhelmed by (fraudulent) requests and brought to its knees.
Ufff, last attempt to understand that silly WCF.
What I did now is:
create client that starts 20 threads, every thread sends requests to service in a loop. Performance counter on server claims that only 2 instances of service object are created all the time. Average request time is about 40seconds (I start measuring before proxy call and finish after call returns).
modify that client to start 5 threads and launch 4 instances of that client (to simulate 20 threads behaviour from previous example). Performance monitor shows that 8 instances of service object are created all the time. Average request time is 20seconds.
Could somebody tell me what is going on? I thought that there is a problem with server that it doesn't want to handle more requests at concurently, but apparently it is client that causes a stir and don't want to send more requests concurently... Maybe there is some kind of configuration option that limits client from sending more then two requests at one time... (buffer,throttling etc...)
Channel factory is created in every thread.
You might want to refer to this article and make adjustment to your WCF configuration (specifically maxConnections) to get the number of connections you want.
Consider using something like http://www.codeplex.com/WCFLoadTest to hit the service.
Also, perfmon will only get you so far. If you want to debug WCF service you should look at the SvcTraceViewer and SvcConfigEditor in the Windows SDK.
On your service binding what have you set the maxconnections to? Calls to connect will block once the limit is reached.
Default is 10 I think.
http://msdn.microsoft.com/en-us/library/ms731379.aspx