[ntp:questions] Testing throughput in NTP servers

Terje Mathisen "terje.mathisen at tmsw.no" at ntp.org
Sat Sep 15 08:57:54 UTC 2012


Rick Jones wrote:
> Terje Mathisen <"terje.mathisen at tmsw.no"> wrote:
>> You're talking 10 GiGE wire speed, right?
>
>> That's more than 100 M requests/second!
>
> If I recall correctly, the maximum number of minimim sized frames per
> second for 10GbE is something like 14.7M or 14.8M each way.  I think

Ooops, I managed to skip the bits-to-bytes step.:-(

Anyway, the CPU cycles needed for wire speed is obviously easy to get 
then, the bottlenecks should be in the NIC/bus/hw parts.

> Luigi Rizzo has gotten 10+ Mpps with an Intel NIC (just basic
> networking) using his netmap stuff.
>
> Using a two-socket server with E5-2680s, and a non-Intel 10GbE NIC
> I've seen around 2.89 M pps each way on a single port with aggregate,
> concurrent, burst-mode netperf TCP_RR tests. If I use non-burst mode
> and many more concurrent netperf's it is more like 2.5 (from memory).
> I suspect it becomes more of a context-switching benchmark at that
> point.  Haven't quite gotten around to driving both ports at once.

Right.

Terje

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



More information about the questions mailing list