Control-M and kernel parameters.

[Simone Benzi_1]

Hello everybody,

We're managing a hp-ux 11.0 system (L1000, 2 CPU, 3Gb RAM, 2.5 Gb swap) with an oracle 8.1.6 instance and a Control-M batch scheduler (with a Sybase db).

Due to various problem with Control-M, customer ask to improve this kernel parameters:

nbuf (from 0 to 646)
shmseg (from 32 to 36)

At present tunable parameters are:

STRMSGSZ 65535
dbc_max_pct 10
dst 4
max_thread_proc 256
maxdsiz 0X20000000
maxfiles 2048
maxfiles_lim 2048
maxswapchunks 1290
maxuprc ((NPROC*9)/10)
maxusers 200
msgmap (MSGTQL+2)
msgmax 32768
msgmnb 65535
msgmni (NPROC)
msgseg 32767
msgssz 128
msgtql (NPROC*10)
ncallout (NPROC)
nfile (15*NPROC+2048)
nflocks (NPROC)
ninode (8*NPROC+2048)
nkthread ((NPROC*2)+16)
nproc (2*(20+8*MAXUSERS))
npty 120
nstrpty 60
nstrtel (MAXUSERS)
num_tachyon_adapters 5
semmni (NPROC*5)
semmns (SEMMNI*2)
semmnu (NPROC-4)
semume 64
semvmx 32768
shmmax 0X40000000
shmmni 512
shmseg 32
timezone (-60)
unlockable_mem (MAXUSERS*10)

Does anybody see any issue in modify nbuf and shmseg? Consider that this is a very important production system so we MUST act as much "conservative" as possible!

Thanks a lot!

Simone

[Pete Randall]

Simone,

I can't see any problem with these modifications.


Pete

[James R. Ferguson]

Hi Simone:

"Tunable" parameters are meant to be just that -- "tunable". Be aware that that setting 'nbuf' to a non-zero value will disable dynamic buffer caching. In general, see:

http://docs.hp.com/hpux/onlinedocs/939/KCParms/KCparams.OverviewAll.html

Regards!

...JRF...

[A. Clay Stephenson]

I would have no problems at all increasing shmseg - you could easily increase it to 256 or so. It's simply the per process number of allowed shared memory segment attachments.

In almost all cases nbuf should be left at zero although it is impossible to tell you without also knowing the value of bufpages. Those two paramters interact. If bufpages is 0, dynamic buffer cache is enabled and nbuf should be left at zero; if bufpages is non-zero, in ALMOST all cases nbuf should still be left at zero because the default behavior is almost always optimal.

I think you are the "victim" of Control-M's generic tuning suggestions.

[Steven E. Protter]

Might want to actaully measure performance to see if shmseg is being utilized. We had ours at 256 for a while and were utilizing less than 10%

I see no problems with the changes, they should not really take things out of whack, just wondering if anyone looked at data prior to making these decisions.

I am attaching a background script that measures performance for a user defined period of time, collects the data in a nice file.

SEP

[Jeff Schussele · Accepted solution]

Hi Simone,

IF you set nbuf & leave bufpages at zero, you do two things:

1) You negate the dynamic buffer allocation.
2) You set bufpages to nbuf*2 by default.

Setting the buffer to a *fixed* value is somewhat of a gamble - I like to let the system allocate the buffer dynamically thereby letting the buffer grow when memory use is low & disk use high, but still letting memory use grow when needed.
So IF you must use static buffer allocation, then you need to look long & hard at just how much you need at all operating levels & set the value so you have enough for heavy disk usage, but not so high that you starve the apps for memory.

My 2 cents,
Jeff

[Simone Benzi_1]

Hello!

Thanks a lot for all your answers.
Is there any way to see how many buffer pages the system is currently (dinamically) allocate?

Probably we do not modify nbuf since it looks like a not very "conservative" change, but we can start with an analysis with it.

Simone

[A. Clay Stephenson]

By far the best tool to view buffer cache is Glance. Just look under "Memory Report" and Buffer Cache is right there. If you do not have Glance, you can install the 60-day Trial Version from your Application CD's.

[Mark Greene_1]

Run gpm and from the "reports" option on the menu bar, select "system table reports". Nbufs is reported as "Buffer Headers", and is the last entry of the top section.

HTH
mark

[Victor BERRIDGE]

Hi Simone,
What version of what ControlM are you running?
If you were to generate a new kernel Id go and double maxdsiz => 0x40000000 and yes pass shmseg to 50 (of course leave nbuf to 0)
What Im more concerned about is by the look of your params you dont seem to have a lot of swap this may be an issue depending what you are running contrlmserver with how many agents, or ctm/Em (can usequite some RAM)


All the best
Victor