In <[EMAIL PROTECTED]> wayne <[EMAIL PROTECTED]> writes: > Back in April, I created few scripts to monitor ntp usage, see: > http://www.schlitt.net/scripts/ntp/ > > I asked people, if they wished, to run these scripts and to post URLs > to these stats here on the list. Here is a list of URLs that I can > quickly find: > > [list of URLs snipped]
I went through these URL to the ntp stats to see which ones also recorded the pool_stats.log file. This is the file that contains the number of active clients and whether the particular ntp server is in the DNS pool. I only found three usage sets of data: http://www.schlitt.net/ntpstats/pool_stats.log http://ntp.vuntz.net/stats/pool_stats.log http://ntp1.pulsation.fr/ntpstats/pool_stats.log However, I think this is probably enough data to reach some conclusions. I looked at all the the spikes that occur when the ntp server gets put into the pool for the month of August. I found that: 1) They all had very different base number of active clients when not in the pool. I strongly suspect that this is due to how long the server has been involved with the our project, and therefore how long it has had to accumulate steady, long term, clients. 2) When put into the pool, all of the servers had spikes in the range of around 700 to around 2000 new clients. 3) When the servers were removed from the pool, these extra clients very quickly disappeared. I was very surprised at the size of the spikes. Eyeballing the data over the months, I would have guess more like 800-1000. I was very surprised at the variance in the size of the spike. Looking at the data, I don't see an obvious correlation between the size of the spikes and the time of day or anything. These spikes appear to be caused by some kind of client that does a DNS lookup before each query. This really concerns me because the size of the DNS lookup is actually significantly larger than the size of the NTP packets and we are going to have a *much* harder time distributing the load on the name servers. So, these clients, if the continue to grow in numbers, threaten the scaling of the pool name servers, rather than the pool ntp servers. This is something I really hadn't seriously considered before. It might be worth our while to track down these clients and see if there is anything we can do to fix them. If, say, they are largely caused by some linux distribution that has an ntpdate script in a loop, we might be able to get that distribution to fix their software. Right now, there appears to be only 700-2000 of these clients that float around from ntp server to ntp server, which isn't too bad. I would guess that the average pool server has something like 2500 clients, and there are around 350 servers in the pool, so we are serving about 875,000 "clients". How many of those are duplicated because people (correctly) use several servers is hard to say. -wayne _______________________________________________ timekeepers mailing list [email protected] https://fortytwo.ch/mailman/cgi-bin/listinfo/timekeepers
