Re: Responsiveness of Server at high CPU load

vslabs_at_onwe.co.za (Billy Verreynne) wrote:

>
>How about the other Oracle sessions that are connected? Are they
>responding?

This problem has happened several times during the last year. Every time is happens, coincidencially, lot of work is being done with the databases (these are mostly test instances) because some external contractor needs to get his tasks done. Otherwise and most of the time, the instances are just idle. So I am asked to DO SOMETHING to solve the problem immediately. In this situation, I have no chance to investigate anything.

Last time, though, I noticed that two instances were active. I could hardly connect to any of them. My mistake was not to try the quiescent instances too... So I can't tell now whether ALL instances were suffering the same problem.

But definitely, at the OS level, one could still more or less work fluently.

>Are you using MTS or dedicated server or a mix?

Dedicated server only.

>If a process is consuming CPU resources to such an extent that the
>system stops responding, then it is a kernel issue IMO. Unless of
>course the app ups its process priority to something like real-time.
>But then Oracle does not muck about with process priorities - Oracle
>simple forks/threads with default process priority.
>
>If it is indeed a kernel issue (aka bug), then the *complete* system
>will be unresponsive - i.e. no telnet /ftp connections accepted,
>existing telnet session commands and response very slow, etc.
>
>Is this the case? If so, then IMO you have run into a Linux kernel bug
>or something along those lines.

As I said, Oracle commands would last forever to complete, OS commands were just slow.

>Lastly, if the existing Oracle sessions are still responding, then
>make sure that you establish a sysdba connection up front. Use that
>session to pop the hood and dig around the v$ tables (especially
>events and waits) to see what Oracle is doing and what it it waiting
>for during the time that no new Oracle connections are accepted.

There is no way to investigate anything, the system just won't answer.

>Oh yeah - is there something equivalent to truss on Linux? That can be
>a very useful tool at times like this.

I tried pstack <Oracle_process_id> and waited, waited.... until I gave up.

Bye
Rick Denoire Received on Wed Dec 17 2003 - 18:17:40 CST