1

I run an algorithm that receives out-of-process messages of different types. The incoming messages are actually byte arrays and each byte array is pre-pend by a byte array flag indicating the message type. I like to understand whether it is possible to setup an IPropagator<byte[], byte[]> that processes the incoming byte arrays, interprets the byte array flags and then streams the byte array to a specific corresponding linked ActionBlock.

For example lets say I have 2 different message types and I have 2 different corresponding ActionBlocks that should only receive messages that match with the intended message type they are supposed to receive. I believe if I just link the IPropagatorBlock to both Actionblocks that both ActionBlocks will receive the same message? How can I correctly allocate each message depending on its flag (do not worry about the flag, the identification is trivial, lets assume I know at any time to which ActionBlock IPropgatorBlock wants to stream the message)? I am struggling with correctly setting up the data flow structure. I would love to be able to link the data blocks directly to each other rather than having to Post(). Is that possible?

Any help in that regards is much appreciated.

Matt
  • 7,004
  • 11
  • 71
  • 117

1 Answers1

1

This depends on the IPropagatorBlock that you're using. If it's a custom one, you can do anything, including for example recognizing which target block to use based on the order they're linked (or something more reliable).

But assuming the block is actually a normal TransformBlock (or something like that), I think the best option is to use the overload of LinkTo() that takes a predicate, and adding the flag to the output type (which means changing the type of the block to IPropagatorBlock<byte[], Tuple<FlagType, byte[]>>, or a custom type instead of the Tuple). If you do this, then you can be sure that the target will receive the message only if the predicate for that target matches the message.

Also, what happens if you link one source block to more target blocks depends on the source block. In most cases, it sends each message to exactly one target: it first tries the first target, and only tries the second one if the first one declines or postpones the message. The exception to this rule is BroadcastBlock (and the similar WriteOnceBlock), that always tries to send each message to all targets. Again, a custom block can behave any way it wants.

svick
  • 236,525
  • 50
  • 385
  • 514
  • I will test it myself but what is your hunch how fast the approach of using predicate/filter is vs doing the allocation manually within the transform block and then simply posting to Actionblocks depending on what the flag is. (Would obviously not need a transformblock then but an actionblock that posts to other actionblock. Obviously not the prettiest way but do you think it would be faster/slower than using predicates? – Matt Jul 20 '12 at 15:00
  • I think that using predicates will have very small performance penalty. If you have lots of targets, then the linear search for the correct target could be a problem, but that doesn't seem to be your case. – svick Jul 20 '12 at 15:12
  • I tried it: Sleek solution and quite fast. Thank you very much. The reason I put up the question because I have incoming messages using ZeroMQ as messaging system. I either have just one incoming socket and allocate different message types to its respective action block for processing or I set up a socket for each message type and have to write a lot of code to write event handlers and forward the events to another class. I like to try this approach because I want to process concurrently anyway. – Matt Jul 20 '12 at 15:26