🚀 go-pugleaf

I know the T414s are antiquated, but at least they are readily accessible
to me.  In order to collect some reasonable performance results, I need to 
know some details about the timing of transmission on the T414 links.

I adapted an example that came with the Logical Systems Transputer Toolset
(actually example 4), which very simply put a process on each transputer to
send its own node number along the boot channel, and to receive and forward
similar messages from other transputers.  I modifed this so that each such
message was acknowledged by sending it back along the channel it came on,
and by having the processes continue indefinitely.  

The root transputer reported the time for numbers of iterations of receiving
and echoing this message.  The results were as follows:

Number of iterations		Length of messages		Average time
per report							per iteration

	   1				2 bytes			0.2602 	secs
	  10				2 bytes			0.02468 secs
	 100				2 bytes			0.00241 secs
	1000				2 bytes			0.0009773 secs
	2000				2 bytes			0.0009417 secs
	5000				2 bytes			0.0009092 secs
	2000				4 bytes			0.00093381 secs
	2000			       10 bytes			0.00093384 secs
	2000			      100 bytes			0.00096140 secs

The early results (with frequent reporting) are obviously dominated by the
time to send the report over the network to the remote machine recording the
figures.  For larger number of iterations between reporting, the numbers seem
to stabilise.

On my rough estimate, each iteration consists of the transmission of 2 bytes 
in each direction, making a total of 4 bytes.  The transmission of each byte
includes header and trailer bits and acknowledgement bits.  Even allowing 
100% overhead in the number of bits transmitted per byte, the transmission
costs amount to no more than 0.1% of the 10 Mbps channel capacity.

So where does the bulk of the remaining time go?  

I assumed that the compute time for such a simple program would be negligible 
beside the transmission time.  Is this a wrong assumption on the transputer?

How would the above measurements and conclusions change for the later
transputers?

I think I need to know the above details in order to determine how to 
optimise a distributed simulation package, which runs surprisingly slowly.

Thanks in advance for any suggestions.
--
Charles Lakos.                               C.A.Lakos@cs.utas.edu.au
Computer Science Department,                 charles@probitas.cs.utas.edu.au
University of Tasmania,                      Phone:    +61 02 20 2959
Sandy Bay, TAS, Australia.                   Fax:      +61 02 20 2913