[evla-sw-discuss] Delay Model Distributor

[Kevin Ryan]

Having read that Java's new IO package (nio) was designed with speed  
and efficiency in mind, I used it for Delay Model redistribution (the  
part of the VCI that accepts models from Executor and distributes  
them to the appropriate CMIBs).

After doing this, it was decided to use UDP multicast to send the  
models from Executor.  Unfortunately, nio does not yet do multicast.   
Since model redistribution can get intensive in the worst case, I was  
reluctant to abandon nio for the old DatagramPacket and  
DatagramSocket methods.

I did however try it and attempted to measure their timing correctly  
(I am a little unsure of this because of the difficulty with the read  
block - what I did was start a nanoTime just before sending the  
packet and then stopped the time just after receiving it in another  
thread.  Since both threads were running in the same system, they  
used the same clock so I THINK the test is valid).

The results were that nio and the old methods were almost identical  
in time.  I would venture to guess that this is because of the  
'atomic-ness' of a UDP packet and that the efficiencies of nio are  
better noticed when dealing with streams (or that its efficiencies  
are a pack of lies :).

So we can do UDP multicast for the delay models if we want and not  
lose sleep for not doing it the best way.

I have the pseudo VCI running an the future OTS Prototype CPCC called  
'cpcctest' down in James' area and it can send to either an in house  
CMIB or to those in Penticton.  It has been tested by sending a model  
from my machine and verifying that it is correctly parsed down in the  
bowels of the CMIB.

If Barry gives me his multicast address I will configure it for that  
and he can send his own models.

Kevin