TCP Queuing Problem

I am seeing a strange problem on a Sequent system. We are running Oracle 7.1.6 on a clustered Sequent 750 system running Dynix/ptx 4.1.3. We connect to this cluster using SQL*Net 2.1 from application service processes written using the Tuxedo OLTP monitor. There are several hundred of these processes, which run on four Sequent SE-20's running Dynix/pty 2.1.6.

The problem we are seeing is that occassionally (about once per day) we start seeing queueing on sockets in netstat. The database cluster sees bytes queued in the send queue and the client machines see data queued in their receive queues. When the amount of data queues starts getting up into the 1000+ range, we know this problem is starting to happen. Gradually the queues build up and more and more sockets are affected. Eventually we run out of stream buffers and other bad things like that and we have to reboot the client machines. When that happens, the problems clear up (for a while). Until we either boot the machine or start killing off clients (none of which appear to be hung -- though some are idle in Tuxedo), the queued data stays there.

We're working with our vendors and doing our own investigation to solve this, but I thought I would throw the question out here in case anyone has seen a problem like this and/or has any suggestions. If anyone has a handy Sequent tool for mapping TCP/IP endpoints/sockets back to process slots/ids, I'd also appreciate hearing about it.

Thanks in advance, arenn. Received on Wed Nov 13 1996 - 00:00:00 CET