Reporting disk activity between XP512 and applications

[Marco Forese]

Hi ya all, new to this forum, first post.
And to begin I've a nice trouble in hands.

In a SAN structure, XP512 fiber based, I've some application offering bad wio% response.
Not all the applications, only the two heaviest in db sizes and data access.
One is SAP (~800GB db) the other is a Oracle Custom web appl (~330GB db).
Both of them are clusterized, both of them connected directly to XP via redundancy fiber channel adapters.
By the way I've a SAR -U response average between 35 to 70% wio.
Add info: in cluster only the node allocating DB shows this situation (on both appls; and this is something I find acceptable, this is the node who works on the storage with more frequency).

Thanks in advance for any suggestion,
Marco, Italy

[Duncan Edmonstone]

Marco,

How are your primary/alternate links laid out to the ports on the XP, remember that each pair of ports on the XP is controlled by a pair of i960 processors, one of which handles odd LUNs and the other handles even LUNs. If your LUN distribution is such that all the primary links to one pair of ports (A & B, C & D, etc.) is even or odd then you might see poor performance like you describe.

Not sure if you can get at this link, but you should read this doc if you can:

http://www4.itrc.hp.com/service/iv/docDisplay.do?docId=/DE_SW_UX_swrec_EN_01_E/XP-Diskarrays.pdf

HTH

Duncan

[Marco Forese]

Thanks Duncan,
already tried to download that info cause I made a search in forum past posts, but actually it is unavalaible.
By the way I'm supposing it cannot be a miscofiguration in the structure.
These machines are actually working since 2002 without any previous evidence of this kind of troubles.

[Marco Forese]

Duncan or any else , is it possible to reach that doc linked above somewhere else?
And still, excluding Glance and Performance Advisor, any other tool can report a good info on disk performances?

[harry d brown jr]

Glance/measureware is the ONLY tool that will MEASURE performance accurately. There ISN'T anything else that comes close. Break out the check book and purchase it. Why is that concept hard to follow?

Maybe this will do: http://www.amazon.com/exec/obidos/tg/detail/-/B00001ZWV7/104-3206485-0751952?v=glance&s=toys&me=A3UN6WX5RRO2AG&vi=pictures&img=1#more-pictures ??

live free or die
harry d brown jr

[Marco Forese]

Well Harry, the problem is that in my group money is always a problem.
And you know that Glance or Measureware aren't free.
In some minds get's easier to waste system performances and user satisfaction instead than pay a price and have the best from machines.

[harry d brown jr]

You do this:

Fill out a purchase order for glance/measureware/perfview.

hand it to your boss.

the next time someone asks about anything regarding the performance of a machine, you point them to your boss and have them ask your boss it that purchase order was filled so that you can perform your job.

live free or die
harry d brown jr

[Marco Forese]

Thanks Harry,
but sometimes it needs to live tied instead than free.

[TwoProc · Accepted solution]

There's a tool for your XP called performance advisor that you can purchase that will tell you what's going on inside the XP512. However, it's pretty pricey. I find that the XP512 while being great a concurrency, has some undesirable delays for what I'd call single-process single stream reads. We alleviated them on this box with first, a serious tuning effort at all code running in the box - mostly identified with statspack and just watching long running processes that have high cost factors - which usually meant full table scans to get ride of. Secondly, with heavy analysis of what I/O levels we had and were using at peak, it was determined to put a lot more into the db_buffer_cache. Like 5 times more than what we had. Like more than I was comfortable with (we had a 97% cache hit rate- so it just didn't make sense) - but it worked! For what it's worth, we don't see this same sort of problem (or need of resolution) on the XP1024 - but that's a newer, faster storage system (much faster actually) so it makes sense. To sum, the way to extend the life of our XP512 is to tune(supposed to do that anyway), and get more I/O's buffered in the SGA. Consider also buying more cache ram for the XP512, if you look on the used market it's not that expensive - but we found that money in ram in the server did more dollar for dollar then ram cache in the XP so unless you've got a small amount (like 8G or so) - do the server memory option first.

[Marco Forese]

Let me re-submit the question.
Thankin John, for the patience, the actual cache amount on XP is 12 Gb, which exceedes the requirement for a total allocated space of round 4TB.
Are there some parameters in kernel that directly impact the way I/O is managed by the system?

Thanks anyhow