On Nov 17, 2009, at 17:01, Hal Murray wrote:
First of all: It's great that this is being worked on. I hope the OpenNTPd
will work on this stuff, too - but I'm not holding my breath.
>
> Is the pool project happy having a knob so they can select how many servers a
> system using the pool config line actually gets to use? (If you want to
> change to 4 or 6, just change the number of servers that you return when
> answering a DNS query.)
I think for the pool the ideal would be if it used some subset. That way we
can keep the 'time to live'/caching of the DNS records reasonable and spreading
the load a little anyway. For example use 4 or 5 records and then we can bump
the number of records returned up a bit.
I'd probably add the 'non-numeric' zone to the missing names and have the
"non-numeric" zone return more records than {0,1,2,3}.pool.ntp.org.
So for example when Fedora upgrades to an ntpd with the new stuff they could
use fedora.pool.ntp.org; and OpenBSD if they'd stop being jerks about it could
use openbsd.pool.ntp.org. (Peter, I don't know if you have any influence on
that; but it'd be nice to get fixed).
There are a few things I can think of that ntp implementations could do better:
0) Do less DNS requests. I only marginally care about this honestly; because
scaling the DNS servers to handle the requests isn't too bad. (About a billion
requests a month now; almost certainly 95% or more from sntp implementations).
The 'pool' feature addresses this though which I appreciate.
The two other things that I'd really like to see because they address things
that are hard or impossible to address outside the ntpd implementation:
1) If a server turns into a false ticker or stops responding; after X time (2-3
hours?) drop it and refetch from DNS to pick another. Be sure to deal with
duplicates as the returned results from DNS likely will include servers that
are already being used. This will make us better able to deal with unreliable
servers and it'll give users a way to actually make the traffic to their server
stop in a hurry.
2) Get new IPs to work with on a regular basis from DNS (weekly?). This
should, IMO, also be done for 'regular' servers so server administrators have a
better chance for changing IPs. This will allow users to stop participating
in the pool while still providing ntpd service. (SNTP clients and new ntpd's
will stop querying them but they will continue to provide service to old ntpds
and users who got their IP in another way). It will also balance out any
effect from 1) that'd skew the load towards servers that works[1].
For this part it's important to keep in mind that the goal, for me, is to
rebalance the load between our servers and to give more flexibility and ability
to get out of the pool.
3) Change the default minpoll/maxpoll settings. I know this is controversial,
but I think it'd make sense to have a simple "I'm an end-user" setting that
distributions can use that'll turn down the accuracy and network load a bit.
For servers that are being used by others this should of course not be used;
but for your average laptop, desktop, basic webserver or digital billboard...
A really fancy version would by default or with a configuration option
distributions could use notice that it's not getting requests and automatically
do this.
- ask
[1] For the individual client that's of course a good thing; but it's not
necessarily for the server administrator.
_______________________________________________
timekeepers mailing list
[email protected]
https://fortytwo.ch/mailman/cgi-bin/listinfo/timekeepers
