Oracle SQL*Net & tcp keep alive

In article <CJMEun.CAL_at_vtm.be> gerrit_at_vtm.be (The Administrator) writes:
>
> On our system we have encountered some problems with SQL*Net daemon
 processes
> on the run !
>
> Yesterday a user crashed his client twice and his daemon process
 wouldn't
> stop. In fact they started using a lot of CPU resources (almost 95% of
 the CPU
> was busy for user processes and when I killed them both it dropped to
 less
> than 3%). A "ps -ef" today around 11:00 shows that both processes used
 more than
> 500 minutes of CPU time. The processes started yesterday at 17:53 and
 17:57.
> Which means that both processes have kept the CPU busy for almost 100%.
>
> I called Oracle about this problem and they told me it was due to TCP/IP
 which
> did not signal the SQL*Net that a client had disconnected or that a
 signal was
> given but it was catched. (=> so what he said was : it's not an Oracle
 problem
> but a unix problem. We are trying to find a way around this problem with
> sqlnet version 2...)
>
> As a bonus : because of the processes still running -> dbshut didn't
 work
> properly -> the backup script failed ! (ARGH!!!!!)
>
> Anybody having the same problem ? And even better : anybody knowing a
> solution ????

Yep - we had the same problem. Turns out that in our flavor of Unix (Sequent - dynix/ptx) there are a couple of TCP parameters that can be set to do the following:

1. Enable 'TCP/IP Keepalive'
2. Set two 'threshold values' that determine when:
3. the server goes into 'high traffic' (my term, for lack of a better one) mode, sending rapidly repeated 'are you there?' packets to the client and b) when the server decides the client is gone and kills itself

Setting these two parameters seems to have solved our problem. You should also make sure that the client side code doesn't do anything unusual (like trap signals) which might cause Oracle some heartburn.

Good luck!

--
Clay Jackson - N7QNM
US WEST NewVector Group Inc
Bellevue, WA
uunet!uswnvg!cjackso