Re: Statspack ratios help

From: Terry Sutton <terrysutton_at_usa.net>
Date: Wed, 7 Jun 2006 21:57:21 -0700
Message-ID: <005901c68ab8$299194a0$6400a8c0@TerrySutton>

Cary, hope you didn't hurt yourself falling overboard. :-)

Yes, I misspoke. I did mean that having soft parses is better than having the same number of hard parses, but obviously having almost as many parses as executions is not good. I should have stuck with my initial statement. Ratios don't tell you much.

While I think more highly of Statspack data than Cary does (partly because I often wind up having to work with that data, without the ability to profile a specific process), I find the data in the Wait Events/CPU Usage section much, much more useful than the ratios.

--Terry

Original Message -----

> though the fact that almost all of your parses are soft is a good
thing

Man, I was onboard 'til that last sentence. :)

I think the right statement about this is the same one Terry used in his first paragraph. Rephrasing slightly, I'd say, "It might be good. But it might not. You can't tell by looking only at this."

To me (normally pessimistic), so many parse calls are a red flag indicating that the application is doing something stupid: either parsing inside a loop, or closing and reopening cursors too often, or doing the 2-tier thing of disconnecting and connecting instead of the 3-tier thing of sharing Oracle sessions.

I think Terry's point is that having a bunch of soft parses probably beats having the same number of hard parses, but my point is that what we're seeing here is probably way short of good enough.

My favorite close-enough definition of "soft parse" is this: "How the Oracle kernel handles a parse call that the application should never have made in the first place." Think about it...

One final point: I HATE doing analysis with wait-wait-don't-tell-me data like the basis of this thread. To Jonathan's point, you simply can't tell from data like this what the performance IMPACT of the change is going to be. If we were looking at profiles instead of Statspack data, then we'd be able to tell exactly how many seconds per business task this behavior is costing someone.

That may sound self-serving because I'm a profiler company guy, but the factual basis behind this statement is WHY I have dedicated my career to becoming a profiler company guy.

Cary Millsap
Hotsos Enterprises, Ltd.
http://www.hotsos.com
Nullius in verba

Hotsos Symposium 2007 / March 4-8 / Dallas Visit www.hotsos.com for curriculum and schedule details...

-----Original Message-----
From: oracle-l-bounce_at_freelists.org
[mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Terry Sutton Sent: Tuesday, June 06, 2006 8:04 PM
To: oracle-l_at_freelists.org
Subject: Re: Statspack ratios help

Sandeep,

You've gotten some comments which may be confusing. I'll have to disagree
with the concept that "a 100% buffer cache hit ratio is bad". That's not
true. A 100% BCHR is not bad. Nor is it good. That's why wise folk tell
you not to use ratios to judge/improve performance. The ratio is mostly irrelevant.

What you want to look at is the section of Statspack that is half a page below these ratios-- the Wait Event and CPU usage statistics. Is your system spending a large amount of time waiting? If it's not waiting, is the
CPU usage high (relative to the number of CPUs you have). If the answer to
these questions is no, then you don't have a system-wide problem. But that
doesn't mean that you might not have problems with individual processes, which may not show up in system-wide statistics. Are your users complaining? Is your application sufficiently fast? If something is running slowly, you want to concentrate on that process (which is where you
use 10046 traces and such, as someone mentioned).

But the ratios below don't tell you what you need to know (though the fact
that almost all of your parses are soft is a good thing).

--Terry

> Hi,
>
> I am running OLTP system on Oracle 10.2 and J2EE weblogic.
>
> I am getting some ratios as:
>
> Load Profile Per Second Per
Transaction
> ~~~~~~~~~~~~ ---------------

> Redo size: 156,162.18
3,020.55
> Logical reads: 26,407.64
510.79
> Block changes: 904.27
17.49
> Physical reads: 0.39
0.01
> Physical writes: 34.01
0.66
> User calls: 5,863.32
113.41
> Parses: 981.47
18.98
> Hard parses: 0.01
0.00
> Sorts: 16.97
0.33
> Logons: 5.54
0.11
> Executes: 1,085.95
21.00
> Transactions: 51.70
>
>
> Rollback per transaction %: 44.17
> Buffer Nowait %: 100.00 Redo NoWait %: 99.97
> Buffer Hit %: 100.00 In-memory Sort %: 100.00
> Library Hit %: 100.00 Soft Parse %: 100.00
> Execute to Parse %: 9.62 Latch Hit %: 99.88
> Parse CPU to Parse Elapsd %: 69.06 % Non-Parse CPU: 91.60
>
> With 100% soft parse, execute to parse ratio is so low. Is it bad, how
> I can I improve it?
>
> I see rollback per transaction as 44.17. We are using Hibernate that
> generates database mapping and produces most of the SQLs. How can I
> invetigate further? But I doubt if application is doing some big time
> rollbacks.
>
> Thanks for help
>
> Sandeep
> --
> http://www.freelists.org/webpage/oracle-l
>
>
>
>

--
http://www.freelists.org/webpage/oracle-l




--
http://www.freelists.org/webpage/oracle-l

Received on Wed Jun 07 2006 - 23:57:21 CDT