Re: Binding a process to a processor

From: Simon Travaglia <simon_at_mrjollylivesnextdoor.cc.waikato.ac.nz>
Date: 1997/10/08
Message-ID: <61elge$nio_at_netserv.waikato.ac.nz>#1/1


Kevin Gee <geek_at_cse.uta.edu> wrote:
> Thomas Ray wrote:
> > Hi everyone,
> > We have an RS6000 J40 with 2 CPUs running AIX 4.2 and Oracle 8.0. Our
> > batch processes seem to be contending with the onlines for cpu and we're
> > struggling to meet batch deadlines. So, we've decided to go with another
> > dual cpu card in order to relieve the bottleneck. My question is.... Is it
> > possible, through AIX, to bind the online Oracle processes to one set of
> > cpus and batch to the other? (This way, online can only take up to 50% of
> > the total cpu.) Is there a better way? --
 

> You can bind an individual process (and all its threads) to one
> processor with:
> bindprocessor cpu# PID

For the full 10 points who can spot a potential problem here? Me.
You might want to bear in mind that binding your report servers could cause DB degradation. If you've got a low priority job and it gets chucked out of the DB's service because of that, it's possible it will hold locks on data which it can't release until it gets DB service again. Result: Patchy performance or even Database Deadlocking.

Also bear in mind that it's sometimes unwise to bind to the master processor. (Note that the 'master' is not master in the master/slave send, but is the processor that does all I/O to Uniprocessor-only (and low throughput) Drivers. So, if you bound to the master processor (currently configured in sys/processor.h to 0), you could get some pretty crappy response.

We don't bind, but run some of our Ingres Database servers at a lower priority within Ingres. They're only report servers and are set as 'readlock=nolock' so that we get the performance from both angles - users get their work done first and the report servers run with what's left over. No lock contention. (Of course, there is the possibility that a report might mis-report something if the value of it changes at the exact time it's being read-no/lock)

-- 

Simon Travaglia,  Email without auto spam rejection at spt_at_waikato.ac.nz
http://mrjolly.cc.waikato.ac.nz Ph: +64 7 838-4709 Fax: +64 7 838-4066
  University of Waikato, Private Bag 3105, Hamilton, New Zealand..
Received on Wed Oct 08 1997 - 00:00:00 CEST

Original text of this message