[ntp:hackers] Pool scum

David L. Mills mills at udel.edu
Mon Dec 3 18:31:25 GMT 2007


Danny,

At the moment I use all of the five addresses returned and accumulate 
all non-duplicative ones returned from multiple commands spread over 
geographic regions. I don't try the same region more than once, as the 
last time I did that the same addresses came back. That might be 
different now.

You might be thinking about a single pool command that operates in 
rounds until a maximum threshold ("maxpool"?) is reached. Presumably, if 
the number of survivors falls below maxclock, an automatic pool 
acquisition would bump the total population to at least maxpool. Gotta 
be careful here. The new rascals might not wind up as survivors and even 
to be considered a survivor will take some time. Life is so deliciously 
intricate.

The scheme works very will using the tos maxclock 5 floor 1 ceiling 3, 
which results in only stratum 1 and stratum 2 servers. The really useful 
thing would be able to launch a pool command or equivalent routine call 
during regular operation, as in the manycast client. This might in fact 
not be too hard, but finding the right call in a mound of abstractions 
is not easy.

Dave

Danny Mayer wrote:

> David L. Mills wrote:
>
>> Guys,
>>
>> I noticed some suboptimal behavior with pool servers. The client code
>> correctly does not cast off reachable servers when the number of
>> survivors is not greater than maxclock, but does cast off servers that
>> have become unreachable or unfit. The problem is that some time or other
>> it is common that some server or other goes to sleep for awhile, so the
>> number of survivors slowly shrink. I modified the code so that, if the
>> numberof survivors is not greater than maxclock, no server is cast off,
>> whether reachable or not.
>>
>
> I have plans to help with this. If you haven't used all of the addresses
> returned by DNS, then you can go get the next untried one. If there are
> no untried ones available it will then go and perform a new DNS lookup
> and add to the list of available servers. Then you can try a new address
> to replace the rascals. This will take a bit of time to implement.
>
> Danny
>
>> The number of survivors at or slightly above the maxclock limit tend to
>> be very similar in quality. These rascals are near the end of the
>> distance list and tend to have their watchcat timers reset at different
>> times. The result is that the watchcat timers seldom reach the limit and
>> the rascals tend to hang around a surprisingly long time. This is not a
>> bug, rather a feature.
>>
>> Dave
>



More information about the hackers mailing list