[smokeping-users] Submillisecond RTT measurement accuracy

[G.W. Haywood]

Hi there,

On Fri, 15 Jan 2010, Thomas Klein wrote:

> i have to setup LAN measurements and reporting for a 0.9 ms RTT Service
> Level Objective ...

Traffic over the public Internet will be orders of magnitude slower.
Presumably this is traffic on private networks/leased lines?

> Anybody on the list who has done something similar already or can
> give me some hints about the accuracy i may can achieve with Smokeping?

Smokeping itself will have no trouble with that.  You have great
flexibility in the tools it uses.  I routinely monitor LAN traffic
with RTTs in the region of 200 microseconds, see for example

http://www.jubileegroup.co.uk/JOS/misc/VORTEX_last_21600.png

> I am also wondering what method will be best, icmp ping, TCP syn, UDP
> echo? Hints are welcome.

The more the OS stacks have to do the longer will be your RTT, but
before anyone can give you better advice about that I'd suggest that
your Service Level Objective might need to be specified more clearly.
Pinging a machine with icmp will tell you if a machine is reachable,
but it one might argue there's little point rejoicing in the knowledge
that a machine is reachable if it isn't accepting connections... :)

Be aware that you can have issues with the event timers on some (and
not necessarily the older) hardware, especially with multi-processor
architectures.  Read about the different timing devices and their
limitations; Google for things like PIT, TSC and HPET; find out
exactly what devices you have in your hardware, and if they are being
used correctly by your OS.  Most of the time you'll be OK, but look in
your kernel config for things like CONFIG_X86_TSC_DISABLE and decide
whether you need to use a kernel boot parameter or two.

Above all, look critically at the results you're getting and don't
simply believe what you see.  Take time to become familiar with your
networks, and how they behave.  If you take this seriously, you're in
fior the long haul.

--

73,
Ged.