Performance from disks

[Johan Barelds]

Hi all,

On my company there has been e comparison between different systems with the samen query. The HP-UX server had the poorest results. A lot of clues point towards the io-performance from the disks. Therefor i want to benchmark these disks against other HP-UX systems. I am not able to un Glanceplus because i don't have graphical access and the command-line interface looks like s**t on a non-hpterm. Ik would like to know if someone has good ideas how i can perform such a benchmark and which tools to use (prefereable already installed).

Thanks!

[Leif Halvarsson_2]

Hi,
If you want raw disk performance you can get it from the disk vendors data sheet.
If you are interested in filesystem performance you perhaps can use the Postmark filesystem benchmark from NetworkAppliance.
http://www.netapp.com/tech_library/3022.html

[monasingh_1]

I think use sar, iostat while you do dd. Check how much time dd takes to for certain size of disk. Also check that all systems being checked have the same type of striping, mirroring configured. you do not want to check mirrored with not mirrored or striped with non-striped...

Also, besides disk porformance check the network cards, network connection, number of hops involved between systems, switch and net-card settings.

I would also check the kernel parameters for buffer cache. Also check the immediate-reporting on the disk. I will also check if swap resides on the same disk(VG00).
Also check the connectivity to disk subsystem, number and type of channelsSCSI/fiber, striping, mirroring, disk subsystem cache.

These are few things I will start with..
hope this helps..

[Steven E. Protter]

I have a sar backgroud data collection script that will do everything you want for any time period.

Due to the fact that nurse ratchet has me chained to the bed, all I can do is link to a prior post of mine. Thank G-d for cell-modems.

Last post in this thread.

See ya.
http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x296def70e827d711abdc0090277a778c,00.html

SEP

[Sridhar Bhaskarla]

Hi,

How are you trying to do benchmarking?. What is the tool you are using to produce the load?. Is the backend setup exactly the same on all other systems?

There are quite a few things involved on HP's performance. For ex., HP sets a default dynamic buffer cache maximum of 50% of memory. Depending on the load, it may degrade performance. A previous patch reset the kernel parameter timeslice to 1 that would cause the cpu to spend more time in context swithing alone. And many more.

It is also dependent on how you laid out your disks in the volumegroups and filesystem and how the database is laid out on top of it.

Here is a website that can give you some benchmarking across the vendors.

http://www.tpc.org/

If you do not have any additional tools available (I would recommend installing Glance+Pak that will allow you to capture the performance metrics), then you can use sar. You can put sar in cron and let it collect the numbers while you run your tests. Look at the man pages of "sar", "sadc", "sa1" and "sa2".

sar -b and sar -d deal with disk related metrics.


-Sri

[Bill Hassell]

Same query? Hummm...that sounds like a BIG challenge. I am assuming that the query is an SQL query and that the 'same' database executables, recommended patches, kernel parameters, shared memory definitions, etc are all correct. I have very often seen the 'identical' setup on two different machines produce SQL queries that require 2-3 orders of magnitude more disk I/O, which has nothing to do with disk performance. For instance, shared memory for Oracle can be made very small requiring all sorts to be done on disk rather than memory, or too small for row insertions using memory only. Or threading may be disabled on the HP-UX machine or the buffer cache is much too large, etc. sar can be used to track disk I/O assuming that there is nothing else happening on the machine.

Make sure that the query produces approximately the same disk I/O before comparing the results.

As far as Glance, I would get a copy of Reflection for HP (note: not Reflection/X) as the best HP terminal emulator to avoid all the character mode issues with dumb emulators. An alternative is QCTerm (www.aics-research.com/qcterm/). With a real HP terminal, navigation in sam, swinstall, glance and other HP products is immensely easier.