Problem Statement

Some broker implementations require creating a copy the message forwarding it over to the backend. The broker also might slightly modify things like addressing headers etc. on the message for proper message routing within the DMZ. The problem is that we see a very high CPU cost in creating this copy message and this also results in lower throughput.  Note: Streamed transfer mode is not in scope for this article.

Simulation

For all performance issues we need to measure and profile and to investigate this issue we initially try to simulate the pattern of the broker by just copying over the message and then forwarding it to a backend dummy service. We then take profiles of this to understand how much the actual cost of copying is.

  An 8% cost for copying seems to be acceptable considering the value of making the copy and able to do other things if required. But then again this was not what is being observed. In the profiles from the actual broker we notice about 40% cost for creating a copy. This means that almost half the time is spent in creating a message copy. So effectively your throughput would almost drop to half when the broker is configured to create a copy of the message. This is excluding costs like logging etc. Evidently our simulation is not accurate so we need to isolate this further. We take in more functionality from the broker so that we hit this expensive path. One of the key observations was that the message is copied just before it is being forwarded. This also means that there are a bunch of manipulations that was done on the message and in our simulation we didn't perform any manipulation. So to get this closer we need to probably change some things on the message. To keep it simple we did something like removing some header and adding another header to the message since most brokers modify headers before forwarding it over.

  Eureka!! We observed our throughput went down and this was in line with what we were seeing in our broker. So we can see that CreateMessage and CreateBufferedCopy have increased in cost quite a bit.

  So this was performance data we collected.

  Now that we have identified the root cause we also need to identify the solution so that the broker can achieve the functionality without taking up so much CPU.

Solution

The solution is actually quite simple "Modify your message after you create the buffered copy". I wanted give the solution before the analysis since most of you would probably not be interested in the analysis but if you are then the rest would be interesting.

Analysis

The most common way to create a copy your message is using Message.CreateBufferedCopy(int).

1. The default case of creating a buffered copy creates a BufferedMessage from an underlying BufferedMessageData
2. This is optimal in the following cases
   1. Message headers are not modified (no update in buffered header values)
   2. BufferedMessage headers haven't been captured (this happens for e.g.  when the user inserts a header in the first location)

  If headers have been modified then CreateBufferedMessage takes an alternative path using the DefaultMessageBuffer. The reason is that a fully copy of the message has to be created if any buffered header has been modified. An internal property called headers.ContainsOnlyBufferedMessageHeaders is used to distinguish if the faster BufferedMessageBuffer can be used to create the buffered copy or not. If there are any modified headers then this means we need to assure that the message is fully marshaled over and the buffer itself cannot be copied(e.g. the user can add a reference type to the header) and so we fall back to a path that would fully reparse the message and create a fully deserialized copy of the modified message. The main point here is a copy should always be a deep copy and any kind of modification should not result in a message with shallow copied message parts. When you copy and create a message from the original then your message objects get its own copy of headers that it can play around with without affecting the original incoming message. Message copy by itself is a fast operation as you can see from the above profile and copying a modified message can be very CPU intensive when using buffered transfer mode.

For greater flexibility our router can be something like a pass through router. If we are just calling a backend service then we can use a generic contract to receive and forward messages to the back end service as shown below.

Here we create a copy of the message to consume locally on the broker incase we want to validate some parts of the message or log etc. Ideally the fastest would be to just directly forward it over but application sometimes require all incoming messages to be logged or validated at the entry point of the DMZ.

Pros

1. Loosely Coupling
2. Potentially can avoid a lot of serialization and deserialization cost. (Encoders need to match for this)
3. Changes in the backend usually do not require changes in the Broker provided the broker uses generic client contracts.

Cons

1. Synchronous pattern has its overhead and not scalable.
2. The router has to be of very high capacity to support high loads even though backend machines may block on IO.

Best Practices

1. Match your encoders – Encoder mismatches can cause a heavy serialization and deserialization on the router. This is because any change between the backend and the frontend encoders on the binding would require re-encoding of the message and hence a full read and write at the encoder layer.
2. Avoid having to create message copy. If the backend can create the copy/validate you save the routers CPU for more messages per second.
3. Make sure you use fast message copy  - Buffered messages have a very optimized message copy path that WCF would take provided you haven’t changed parts in the message. I will talk about how to hit this fast path next.

Next – How to Optimize Message Copying using CreateBufferedCopy?

I use the term broker and router very loosely here since they follow very similar guidelines as described here - WCF Broker Overview. Apologies for not being very rigid with these terms.

I will dive into best practices of building a router by progressing from a very simple implementation to a robust one through different scenarios and varying degrees of complexities.

The easiest implementation is by using a strongly typed contract with message forwarding as shown below.

[ServiceContract]
public interface IOrderService
{
    [OperationContract]
    Order[] GetOrders(int numOrders);
}


class OrderService:IOrderService
{
    Order[] GetOrders(int numOrders)
    {
      return backendProxy.GetOrders(numOrders);
    }	
}

Pros

1. Very easy to implement.

Cons

1. Tight coupling between router and backend.
2. The router needs to serialize and deserialize the whole message and hence very inefficient.
3. Any change to backend would require changes to router as well.
4. Not a scalable solution since higher loads would choke the server due to heavy serialization issues.

Next – Router Implementation – Message Forwarding – Copy/Pass through

A broker is usually a central point for message forwarding and pass-through for clients and backend services. There are many types of brokers that come into mind

1. Security broker – Usually a gate check for incoming message which transitions from one security protocol to another without mucking with the message most of the time.
2. Protocol transition broker (http to tcp)
3. Central work load balancer – Routing of some form
4. 
   Message transformer and service busses
   1. Aggregators/ Scatter gather/Composition & other patterns
5. Encoding transition – SOAP/XML/Rest/Binary etc.
6. Relaying persistent brokers (Persistent stores that holds on the message and later forwards it)
   1. Glorified queuing
7. The list goes on… and you can derive a ton when using a combination of Message exchange patterns -

But they mostly fall into the below basic model unless the client manages to bypass the broker with a P2P communication with the backend itself.

Router Implementation – Strong Typed with Message Forwarding

Update:

Aug’5’10

Here are 2 really good articles on MSDN that I would recommend for the functional aspects to architect your your router.

1. Building a WCF Router, Part 1 (Addressing semantics)
2. Building A WCF Router, Part 2 ( Scenarios and Message Exchange patterns)