system tuning for DB

[MartyB]

I have an rx6600 running 11.23. It's got 8 processors and 64GB of RAM. Also, the system is connected via dual 4GB HBA's to a DMX-3.

This system hosts a ~300GB DB sitting on the DMX. I am running into what appear to be IO bottlenecks whenever maintenance events run on the DB. The system is running a mostly stock 11.23 installation, and I'm figuring there are some tweaks I can implement that would help performance on this system.

So my question is, what are some recommended 11.23 tweaks for a large DB system such as this?

TIA!

[Yogeeraj_1]

hi,

First of all, you will have determine the daily workload on the system. If this is an Oracle database, you can look into the Oracle Enterprise Manager Database control tool.

If there are other processes running during that maintenance period, you may wish to reschedule them or move the maintenance to another slot.

If the problem persists, you will have to idetify the key datafiles/file systems that are more affected and alternatively rebuild them to lesser IO intensive areas.

For tweaks at the OS level, you will have to post your current kernel settings so that we can have some clues and recommend accordingly.

kind regards
yogeeraj

[MartyB]

Thanks for the response Yogeeraj. Well, what I'm looking for are system (kernel, or otherwise) tuning suggestions for a server with a high IO workload. Please assume a default 11.23 HP-UX installation, I don't want to leave out any suggestions.

Also, I'd like to leave DB tuning out of this discussion since I'm tackling that through the DB vendor.

Thanks!

[Tim Nelson]

I would much rather see some stats first but here are some generics.

1) reduce dbc_min_pct and dbc_max_pct to 1 and 2 respectively ( give the memory to Oracle)

2) unset the dynamics of vx_ninode (16,000 is a good number ) this releases more memory for Oracle to use.

The rest all depend on what is really going on, need stats.

3) if disk queuing is occuring look at the scsi_queue_depth using scsictl -m queue_depth for each SAN device ( HPUX defaults to 8 but with luns encompasing many physical spindles something like 32 or 64 may help )

4) if avserv is good ( i.e. < 5-7 ) then there is little you can do as the disks are responding execellent.

5) if avserv is bad ( i.e. > 8-10) start looking at the array performance. How many spindles / meta ? What type of layout Raid 0+1, Raid 5 ?

6) separate datafiles.

7) multi-pathing ? Is there bandwidth issues at the FC controller(s), switches, or HBAs on DMX?

7) Tune Oracle !! ( see metalink.oracle.com )

Once again without seeing some stats the issue(s) could be anywhere.