ipcs queues full, system slows down

[kestens jan]

Hi all,

I'm currently testing network management software on HP WS (C3X00, B2600, B180) &K-servers (HPUX 10.20 + general release patches dec 2001).

I noticed that when a certain message queue (ipcs) fills up to almost its maximum value the system slows down (GUI, dcp's,....).
I talked with the developers and they suspect HPUX.

Does anyone have some experience with this ?

I included a part of an ipcs trace.

Greetz

Jan K

[Chris Wilshaw]

Developers normally try to point the finger at the part of the system for which they are not responsible :)

Does the message queue fill up if the NMS is not running? If not, the NMS is to blame.

If the queue fills up when the NMS is not running, are there any other non-system processes running at the time?

[kestens jan]

Some more info by the developer :

I find a situation where lisp can't write to a queue
because another queue is full. Once this happens,
attempts to write the empty queue fail.

When the ber module queue is full, there are a few more
writes to the OI queue and then this stops too.

When the full queue is first emptied, writing to the
other queue also resumes.

This problem is reproducible. You can test it by calling

kill -STOP

the RAI queue will slowly fill up. When it is full,
the user interface blocks. When you call

kill -CONT

everything gets back to normal.

This looks very much like it is an HP-UX problem.

During this period, another process (pid=1800) using an
entirely different queue also can't write messages. The output
of ipcs seems corrupt: QBYTES is 554, but QNUM is 0!
We have seen this before, but now I see there is a relation
with the situation above.