To: Luchini, Kristi
Subject: RE: real-time response of CA on coldfire - CORRECTED TIMES

 

-----Original Message-----
From: Luchini, Kristi 
Sent: Friday, March 03, 2006 8:43 AM
To: Luchini, Kristi
Subject: RE: real-time response of CA on coldfire - CORRECTED TIMES

Sorry - my board's clock was somewhat off
(and it's late...)

The corrected times are

    maximum 2400us
    average 1800us


T. 

-----Original Message-----
From: owner-lcls-controls@lists1.slac.stanford.edu [mailto:owner-lcls-controls@lists1.slac.stanford.edu] On Behalf Of Till Straumann
Sent: Thursday, March 02, 2006 11:49 PM
To: lcls-controls; Akre, Ronald A.; Smith, Steve; Johnson, Ronald G.; Medvedko, Evgeny A.
Subject: real-time response of CA on coldfire

I did a quick and dirty test to get an estimate
of latencies that we have to expect if we try
to ship real-time data over TCP/IP/EPICS/Channel Access:

I let the coldfire write a 'synchronization byte'
to its RS232 port when it fires a CA monitor
(i.e., when it has generated 'new' data)

A MVME6100 monitors a waveform on the coldfire
measuring the time difference between the 'sync byte'
and 'data' arrival.

Below, a block diagram is shown.

Both CPUs are otherwise completely idle. The coldfire
is connected to the MVME6100 by a dedicated, fast
point-to-point ethernet connection. No other network
traffic is present.

The measured latencies are:

   Avg: 2300 us
   Max: 3200 us

No fine tuning has been done but I'd say this backs
the reservations I have. It's not surprising - the
added complexity of TCP/IP, a few copy operations
on a slow CPU...

-- Till


PS:


BTW: with 100baseT, a 1000 byte ethernet packet
(more or less what we need to read 1 BPM) spends
100us on the wire alone. I believe the readout from the
device takes about another 100us. So the fastest
possible w/o any other overhead is > 200us
(although, in my approach I'd try to let the
ethernet chip DMA directly out of the FIFO).


Test diagram:

                 COLDFIRE-IOC
                 ============

              -- scanned at 0.5s
              |
              V

   [ waveform: 1000 shorts ]
               process ---> devsup writes char to RS232 (115kb/s)

              /\                                      ^
              ||                                      |
              ||                                      |
              ||  dedicated                           |  RS232
              ||  100baseT                            |
              ||  Ethernet link                       |
              ||  (TCP/IP/CA)                         |
              \/                                      v

                            MVME6100 IOC
                            ============

   [ waveform: 1000 shorts; ]
     INP: coldfire:wf CP

         'slave' waveform                         helper task
         monitors coldfire,                       resets clock
         reads clock when                         when char arrives
         data arrives

                 LATENCY = data arrival time - sync char arrival time