[cadynce] end to end measurements in cadynce experiments and NTP clock sync

[Jonathan Walsh]

gthaker at atl.lmco.com wrote:
> Adam,
>
> As you start to set up to run GT-4 and start to measure the end to end
> time, to what extent are you worried about clocks of the starting and
> ending nodes of the application strings being sufficiently synchronized so
> as to get decent end to end time measurements? With NTP the test Blades,
> assuming all are connected via gigabit net to a local NPT time server,
> should stay within 1msec to 5msec of each other. However, ATL's Jonathan
> Walsh, who has been attempting to test the "Case 32" 150 SCI setup and to
> take end to end measurements finds that at least at Emulab (Utah's) when
> the nodes comeup they are quite far and after an "ntpdate" they still
> don't really get well sync'ed up. (We are going to gather some measurement
> data to get better handle on what we are seeing). ANyway, I thought I
> would ask if this is an issue in the Maryland testbed or not.
>
> Gautam
>
>
> _______________________________________________
> Cadynce mailing list
> Cadynce at list.isis.vanderbilt.edu
> http://list.isis.vanderbilt.edu/mailman/listinfo/cadynce
>   
The amounts I am off on mean E2E time are upwards of 60ms in some 
cases.  I am recording the on-node times and they are correct to the 
amount of computation done.  I have two simple 4 node SCI chains going 
in opposite direction and the times are off by a similar but opposite 
amount.  (ie 1->2->3->4 vs 4->3->2->1).  I am off both on a fresh reboot 
after a time sync using ntpdate as well as after 2-3 hours of idle 
system with ntpd running.

-Jonathan