performance problem

[prasanna wijeratna]

we have a customer(Bank) running HP-UX 11 in a N class system which is runnning Amtrix - this is middleware switch application which you used to interconnect all regional braches. This is highly write intensive application for single transaction it writes minimum of seven files to disk. Altogether it creates average 60000 files to disk and 400-6000 MB log per day. It use oracle RDBMS to store final data.

At present we have file system /home amtrix which create 100% I/O and some times system is not responce we have created the stripe set then system improved but still sometime we faced the problem Now the application engineers tune their application same time

They are planning to create cache file system on the memory I wonder whether we can do this on HP-UX ? please advice me the possibility how we create cache file system on the memory

[U.SivaKumar_2]

Hi,
They are speaking about Dynamic Buffer Cache
governed by following kernel parameters.
dbc_max_pct - maximum percentage
dbc_min_pct - minimum percentage
Increasing Maximum percentage of the Dynamic buffer cache is recommended for
I/O intensive systems like your's.

regards,
U.SivaKumar

[Stefan Farrelly]

Heres how to setup a ram disk on 10.20 - same procedure for 11. You can create one up to 4Gb.

http://www.dutchworks.nl/htbin/hpsysadmin?h=3&dn=71127&q=mknod%20memory%20filesystem&fh

Also, HP now sells solid state disks - which are basically memory blocks so performance is awesome. You could investigate purchasing a couple of these for your logfiles.

[Tim D Fulford]

The dynamic buffer cache could do the trick. If not, you want to go whole hog & use solid-state disks (mirrored).

Other things to consider

striping over LUNs
Using a storage medium that utilises caching, such as VA7xxx
If throughput is a problem (i.e. kB & not IO), utilising a stripe over different controllers.

I also have some questions/suggestion
o How many IO/s does your system recieve
o How much kB throughput
o How much space do you need
o MeasureWare... (see below)

From the above two you can more easily design the storage to meet your needs.

I would reccomend analysing a week or so data using MeasureWare (if you have it)

extract -xp -gd -b today-8 00:00 -e today-1 23:59

You should get two files xfrdGLOBAL.asc & xfrdDISK.asc. You can import these into excel (you will need to sort through xfrDISK.asc to find the wort disk)

From the xfrdGLOBAL.asc you can answer quite a few of the above. From the xfrdDISK.asc, selecting a particular disk, I would plot disk% Vs Phys IO Rt.
This will show a few things,
o Give you an Idea of the service times (slope of the graph)
o How well the disk responds under increasing loads, thus you can see what the maximum IO/s is
o Using yopur figure for max IO/s you can calculate a max kB/s. Or if the measureWare version relatively recent C.03.45, you should be able to see Phys kB Rt.

I hope that some of this helps

Tim

[Mark Landin]

A solid-state disk may be a good solution for you. HP sells some, but so do many other vendors. Just search for "solid-state disk" on google.com to find some leads.

Increasing your buffer cache on your system may or may not help. For instance, the common wisdom these days is that 300MB or so of buffer cache is best ... you get less "bang" for your "byte" with cache sizes over that. Having an overly large buffer cache may starve other processes of memory and cause paging, so you don't want to overcommit on buffer cache.
Further, if your Oracle DB is in a raw partition, then it won't use system buffer cache AT ALL.

I know that Solaris lets you build filesystems in memory ... as someone else pointed out, you CAN do that on HP-UX, but again you must be careful as you take memory away from processes ... you could cause an altogether different performance problem :)

The general rule is: run the fastest SCSI bus you can (I think the N can support Ultra/160 cards, as well as Fiber channel), use the fastest disks you can, use as many controllers as you can, and use as many spindles as you can.