[ntp:questions] 1Gb/s network issues - was: Re: OT: GPS18x LVC failure
Kasper Pedersen
spam at kasperkp.dk
Mon Oct 12 17:01:34 UTC 2009
On 10/12/2009 12:00 PM, David J Taylor wrote:
>
> "Unruh" <unruh-spam at physics.ubc.ca> wrote in message
> news:QsBAm.47742$Db2.29202 at edtnps83...
> []
>> Actually those Gb ethernet switches seems to have some really bad and
>> variable
> so whether being very simple and unmanaged is likely to mean that it's
> performance is better or worse I leave to speculation or measurement!
I did some tests on Linux. As expected there's variable latency and
jitter in each direction, in addition to what the network provides.
Something that gives really large offsets (obviously) is when the
network stack decides/has to do an ARP, since the other end has one
second to answer. I measured purely at the receive path: there is the
jitter from interrupt coalescing in the nic, plus the kernel interrupt
latency. The receive path is interesting since the kernel provides
packet receive timestamps, you don't have to wait for the packet to make
it up to userspace.
I built a device that spits out priority tagged ethernet frames with
<200ns error to UTC, and did a number of tests with different networks
and nics, the result of which is in a table here:
http://n1.taur.dk/etherpps/
On switches that understand priority tagging it makes a world of
difference. On the corporate network where I had 14us unaccounted
latency and 18us jitter with the packet prioritized above all other
traffic, when the prioritization was disabled I got another 200us jitter
at peak load time (users and iscsi traffic sharing the same path).
I suppose there's a way to make the kernel priority tag outbound ntp
packets, but I did not find it.
/Kasper Pedersen
More information about the questions
mailing list