[ntp:questions] Re: stratum

Terje Mathisen terje.mathisen at hda.hydro.com
Mon Sep 26 13:30:52 UTC 2005


Richard B. Gilbert wrote:

> The "maximum" number of clients a server can handle depends on many
> things.  Perhaps the most  important consideration is network
> bandwidth.   I've read somewhere that rackety.udel.edu, a stratum one
> server, run on a Sun IPX (antique hardware (12-15 years old??), serves
> something like 672 clients and uses a large fraction of a T1 line

That can't be right!

I'm almost certain the real number of clients is much higher, maybe an
order of magnitude or two?

The results from 'ntpdc -c monlist' is limited by the 600-700 entry
monitoring array, which is in the right ballpark to come up with your
672 figure, but doesn't have to indicate that this is the actual number
of clients:

Worst case each client will poll the server every 64 seconds, each
packet is also about 64 bytes (or a little larger, but this doesn't make
too much of a difference here), so each client generates about 1
byte/second of IO (in both directions).

A T1 line should be able to sustain at least 150-180 KB/sec, giving a
similar number of clients before the network is saturated.

Note however that most well-behaved clients will settle down to polling
much less frequently than every 64 seconds, on my machines most
stabilize at 1024 second poll intervals, with short-term drops to 256 or
512 before returning.

> (1.5Mbits/second).   Modern hardware, served by a T3 line (45Mb/second
> and ca. $8000 US/month) should be able to handle forty-five to fifty
> thousand clients, assuming that the clients are well behaved.

More.

Worst-case clients have been seen to poll every second, in which case a
T3 will saturate at ~80 thousand packets/requests per second.

Terje
-- 
- <Terje.Mathisen at hda.hydro.com>
"almost all programming can be viewed as an exercise in caching"




More information about the questions mailing list