RE: AFF_SPROC + AFF_NPROCS
Posted in 2000
Dominic,
My particular experiences has seen a significant performance improvement at
all the sites I've set the AFF's parameters. I set AFF_SPROC to 1 rather
than 0 - just in case there are special requirements for the first CPU and
AFF_NPROCS to one less than the total number. I recommend this as a starting
point and then run heavy monitoring for a week or two (appl_jobs and oninit
Unix system secs vs Informix vp secs)- i.e. lots of ps -ef's and onstat's
and even more awk's to join everything together into readable N min buckets.
Geoff Poole
> -----Original Message-----
> From: D.Penfold@Libertel.nl [SMTP:D.Penfold@Libertel.nl]
> Sent: 10 February 2000 15:18
> To: informix-list@iiug.org
> Subject: AFF_SPROC + AFF_NPROCS
>
> Does anyone out there have any warnings/advice re: setting the above
> parameters? I'm administering a system at the moment that has 6 cpus, and
> NUMCPUVPS set to 5, but neither of the AFF parameters are set. onstat -g
> glo
> shows very uneven cpu vp activity, and top shows very uneven cpu activity.
> I
> immediatley suggested setting:
>
> AFF_NPROCS 5
> AFF_SPROC 0 or 1>
> Someone else panicked at the suggestion, saying that it meant that
> informix
> would take up too many resources and not leave the in house application
> with
> enough, which may need up to 50% of the cpu activity. The only problem I
> could think of was if some of these individual in house application
> processes take up a lot of cpu time, and may need to be individually bound
> to specific cpus. This is apparently not the case, its lots of small
> processes. Accordingly, I said that the way to deal with this is to
> allocate
> a lower NUMCPUVPS, 4, 3 or even 2, and spread these around evenly with the
> AFF parameters, not to just let them all bunch up together.
>
> Someone else has put these changes on hold, saying they want to look into
> what cpus perform intterupts or any other vital function.
>
> Any advice on any of this? e.g. my approach, anything I've missed, and
> also
> what means/tools could you use to determine what cpus perform special
> unique
> functions and need to be protected.
>
> Thanks in Advance
>
> Dominic