[ntp:questions] Nice fanless high-perf NTP server: Fitlet!

Terje Mathisen terje.mathisen at tmsw.no
Fri Jan 16 07:02:09 UTC 2015


Paul wrote:
> On Thu, Jan 15, 2015 at 4:40 PM, Terje Mathisen <terje.mathisen at tmsw.no>
> wrote:
>
>> Not in my msg, but in the subject of the entire thread. :-)
>
>
> I'm so used to nomail at example being wrong I had a knee-jerk reaction.  My
> bad.
>
> This does give me the chance to ask what a high-perf NTP server might be.
> Given the constraints of the protocol I've always thought of performance in
> terms of queries per seconds.  What metric are you using?
>
That's correct, the highest performance ntp server I've ever worked on 
was a multi-threaded version that ran on a quad or 8-core Intel Xeon 
server cpu, with multiple Gbit/s interfaces.

It turns out that it is exceedingly hard to manage Gbit wirespeed for 
the extremely small ntp packets, but probably possible now if you can 
allow any core to service any network request, with no need for a single 
core to be dedicated to each network interrupt handler. (The box above 
was fixed at an older Linux kernel version which had this limitation.)

Anyway it is definitely possible to get into the 100K to 1M 
requests/second range.

As I noted above the real problem isn't in the actual packet processing, 
which can be made very efficient indeed for the normal case of client 
mode request/reply, but in all the HW/driver/OS overhead that's occured 
before you get those packets in&out of the ntpd process.

Re. the Fitlet: With a 3.9 to 4.5 W power budget this box will never get 
into those ranges, but even handling 1K requests/second with sub-ms 
jitter and delay would still be a very nice Pool server.

Terje

-- 
- <Terje.Mathisen at tmsw.no>
"almost all programming can be viewed as an exercise in caching"



More information about the questions mailing list