kernal parameters

[p7]

hi all

we have a superdome partition: 11i,16 cpus, 30gb mem, lots of filesystems on xp1024 disks.
we re consilidating from many smaller hpux systems. im getting one shot to tune the kernal properly before its almost impossible to get downtime.
it will have many oracle databases, samba, nfs, and older cobal telnet based apps.
any thought on tuning the kernal properly would be appreciated.
(by the way, yes i already got the paramaters from individual vendors, but since this has so many different apps on one system i was wondering if any of u have encountered similiar)

thx in advance

[A. Clay Stephenson]

There are just not enough data to give you any meaningful hints. One thing I would do is essentially decouple all of the formulae related to the pseudo tunable maxusers and insert values directly. On a system this large, the formulae are almost certain to be wrong. Also, ninode (unless you have more than the /stand filesystem as hfs) should be limited to a few hundred - 1000 is very generous. Because you have mentioned Oracle, make sure that you do not set timeslice to 1. Leave it at 10 or a value very close to 10. Be generous with values like maxvgs, nfile, shm*, sem*, msg*, nkthread, *npty*, and maxfiles. Obviously, pay more attention to the static tunables than the dynamic ones.

I suspect that you are running a number of legacy 32-bit applicaztions so give some real thought to memory windows otherwise your global 32-bit memory space is going to become scarce.

[TwoProc]

well, number of semaphores and nfile (number of files ) will certainly be scalar with the number of databases (if we can assume they are of the same size). So multiply these by the number of databases - also look at maxusers (as it is part of the formula base for many of other things).
If all of the databases are launched by the same user (eg oracle), then the maxfiles_lim and maxfile would have to be updated. If it's all separate users (best if it would be all separated) then you don't need to fool with that one.
You might also have increase nkthread. I'd prefer you be able to "fool" with this parameter to see what you need - but if you only get one shot - multiply it times the number of databases. This is a low-probability one - but I've been tripped up with it before (including it mainly b/c of your one-shot rule).
What about number of shared memory segments? Probably need to scale up with that one too.

[Kent Ostby]

This is slightly off topic, but you want to make sure that you have at least 2x RAM worth of disk swap space.

Depending on your previous oracle experience this may need to be even higher.

On the flip side, give careful consideration to dbc_max_pct and dbc_min_pct. You don't want to leave this to 50% with a high memory system (i.e. leaving the defaults to 0).

[TwoProc]

Re: previous post - HPUX can only use up to 32Gig of swap. So just set it at the max if you desire, and psuedo-swap will take care of the rest.

[Don Morris_1]

No, HP-UX can use up to 32Gb of swap *if swchunk is left to the default* (and to get that on 11.11 or 11.0 you have to increase maxswapchunks to the maximum, 11i v2 doesn't have maxswapchunks anymore). If you increase swchunk, you can go up to 1Tb.

[Ted Buis]

Have you considered using hard partitions or vPars? In consolidation there are always trade-offs between isolation and maximum efficiency.