[evla-sw-discuss] Delay models to station boards

[Kevin Ryan]

On May 14, 2008, at 1:43 PM, Sonja Vrcic wrote:

> I completely agree that there is no need to send Delay Models via  
> MCCC.
>
> The multicast address for Delay Models could be specified as a part  
> of VCI subarray configuration (a different multicast address can be  
> specified for each station).
>
> If we can agree that there is no need to send Delay Models via  
> MCCC, we should implement things that way from the beginning.

I don't think this is something that we can just agree on, it must be  
determined if it will work.

Like I said earlier I did a small amount of testing by sending UDP  
packets*:
1) from the VLA to my office,
2) between two machines on different switches in my office,
3) between two machines on the same switch in my office.

(note* these tests are not official and could very well be flawed so  
we should not base any design decisions other than more testing on  
them.)

In one test, I sent 2000 datagrams of the same size every 100ms then  
increased the datagram size and repeated until the maximum datagram  
size.  In the next set of tests I did the same thing but instead of  
one datagram per 100ms, I sent 4 back-to-back every 100ms.

The observations are:
datagram sizes > ~47k do not work at all on this network (the maximum  
datagram size for IPV4 is 64K but this is network dependent and not  
guaranteed).  In fact, it is written that some industrial strength  
systems do not even exceed the Ethernet's MTU size (~1500 bytes) for  
their datagram to eliminate fragmentation altogether.

Non 'burst mode' did well with 100ms between datagrams; for datagram  
sizes up to 16k 2000 of 2000 datagrams arrived, at 32KB 1939 of 2000  
made it, and at 46KB 1814 of 2000 were received.

Burst mode did not fair as well with losses from .001% to 50%  
depending on size.
Between 2 machines on the same switch:
rxNum: 1999, packetNum: 2000, size: 15926,  burst Sz = 4, interval = 100
rxNum: 1560, packetNum: 1999, size: 32566,  burst Sz = 4, interval = 100
rxNum: 1000, packetNum: 1998, size: 46641,  burst Sz = 4, interval = 100

The IP layer fragments datagrams into MTU sized pieces and, if one of  
these fragments is missing on the receive side, the whole datagram is  
discarded.  So it makes sense that the larger the datagram size, the  
better chance it has of being discarded.

For sending models the datagrams would fan out rather than funnel in  
as the test's burst mode attempted to simulate so maybe it will not  
be a problem.  Even in the best case however, I think we can rest  
assured that we will drop models.  And I think it will be in our best  
interest to find out how many with all Station Boards in place.

Kevin