loosing packets

[Rick Jones]

Tom Schmitt <TomSchmitt at gmx.de> wrote:
> Most times a DNS-request is answered properly, but sometimes you'll
> get a timeout instead. I looked into the problem and found that I'm
> loosing UDP-Packets (UDP-Buffer-Overflow). Looking at "kstat -p
> ip:0:ip:udpInOverflows 10" showed me, that I'm loosing about one
> packet out of 100.

> Increasing the Buffersize from 2 MB to 4 MB didn't change anything.

Are you certain that named isn't making a setsockopt() call of its own
to set the SO_RCVBUF size?  Might want to fire-up truss and trace the
system calls being made at startup.  Unless there is another way to
see the SO_RCVBUF size of a given socket.

> Googeling for this problem brought the suspicion, that the load is
> too much for the server so it cannot handle all the requests. So I
> tried to confirm this suspicion: I did a load-test with queryperf
> and put ten times of the normal load on the server, doing more than
> 100 thousend queries in about 25 seconds. But what did I see? There
> was no increase of packet-loss. Instead, in this short period of
> time, there was not one single packet lost. So it seems to me that
> this is not a simple load-problem.

How many queries did queryperf have outstanding at one time?  While it
may indeed be something else, one has to make a distinction between
servicing a steady-state load and the size of a burst that can be dealt
with.

Was your queryperf load from one system or many?  From how many
different systems does your production system receive load?

rick jones
-- 
oxymoron n, Hummer H2 with California Save Our Coasts and Oceans plates
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...