IDS Forum

Re: Efficiency of the Informix Virtual Processors

Posted By: Cesar Inacio Martins
Date: Friday, 21 September 2012, at 7:17 a.m.

In Response To: Re: Efficiency of the Informix Virtual Processors (JASON HARRIS)

Hi Jason,

Well, about the proc affinity , I have a complete inverse experience...
when defined the affinity the performance goes down.
But this was with Sun T2 processor (ugh!), not tried with this IBM P7 +
SMT4.

Your questioning about KAIO , drop/add CPU VPs keep the KIO thread
working, I can say what happen here : after drop/add cpu dynamically,
looking with "onstat -g ioa" , the kio appears for this new CPU VPs
added and the numbers to show they working "properly". (if can affect
others "parts" of the engine , is other mystery).

About the efficiency versus I/O wait, Networking , memory and etc, I
believed this is not an issue in our environment.
We have a good I/O throughput with our storage , network and about
memory we are using large pages and there is no problem with the amount
already allocated (this LPAR have only the database working over it, so
no concurrency with others softwares).
The context switching is something what is hard says if you have a good
number or not (or exists some formula? for database dedicated environments).
I have all historical (sar) since we using other production machine
(Power 6) and the point of view the proportional growth of the context
switch VS # of Procs (where the new/actual machine have 3 times more
CPUs/COREs) , the number is good because they grew less of 60% .

I will think about test the affinity , just not sure when...
Thanks for this tip.

On 20/9/2012 20:45, JASON HARRIS wrote:
> Hi Cesar,
>
> It may seem odd, but I have found that Informix works better with SMT if you
> use affinity. Otherwise, SMT tries to push every thread of one oninit
> processes to a different logical processor. We want them all to stay on the
> same one, setting processor affinity does that.
>
> I don't recall off hand, but when you add a CPU VP does it get a KIO thread?
> And if the ones you dropped had poll threads on them, do the ones you add get
> new ones? These could affect the performance.
>
> To understand thread efficiency, ask yourself this question: Why does it take
> so long (thread time) to get so little (CPU time) done? Typically, slow I/O
> either network or disk can contribute, contention with other processes on the
> server, lock of memory/swapping, too much context switching.
>
> Consider a system with slow disk, the threads won't do much work sitting
> around waiting for data to come it. This will result in a low effective
> percent.
>
> A low effective percent is not always something that can be fixed with simple
> configuration changes, or even something that needs to be fixed.
>
> It does indicate that there is room there for improvement. It probably
> indicates that adding CPU won't improve the performance.
>
> Hopefully, John will write some more about it in the future.
>
> Cheers,
>
> Jason
>
>
> *******************************************************************************
> Forum Note: Use "Reply" to post a response in the discussion forum.
>

Messages In This Thread

• Efficiency of the Informix Virtual Processors
  VIKAS HIVARKAR -- Monday, 17 September 2012, at 3:02 a.m.
  • Re: Efficiency of the Informix Virtual Processors
    KERNOAL STEPHENS -- Monday, 17 September 2012, at 9:36 a.m.
    • Re: Efficiency of the Informix Virtual Processors
      Fernando Nunes -- Monday, 17 September 2012, at 11:40 a.m.
  • Re: Efficiency of the Informix Virtual Processors
    Cesar Inacio Martins -- Tuesday, 18 September 2012, at 11:48 a.m.
    • Re: Efficiency of the Informix Virtual Processors
      JASON HARRIS -- Wednesday, 19 September 2012, at 10:56 p.m.
      • Re: Efficiency of the Informix Virtual Processors
        Cesar Inacio Martins -- Thursday, 20 September 2012, at 7:34 a.m.
        • Re: Efficiency of the Informix Virtual Processors
          JASON HARRIS -- Thursday, 20 September 2012, at 7:45 p.m.
          • Re: Efficiency of the Informix Virtual Processors
            Cesar Inacio Martins -- Friday, 21 September 2012, at 7:17 a.m.
            • Re: Efficiency of the Informix Virtual Processors
              JASON HARRIS -- Friday, 21 September 2012, at 9:14 p.m.
  • Re: Efficiency of the Informix Virtual Processors
    ANDREW FORD -- Thursday, 25 October 2012, at 4:30 p.m.

IDS Forum is maintained by Administrator with WebBBS 5.12.