Ian,
You have run into something that I've seen in the past which is a
performance problem, but which is not caused by a bad BHR. What I have noticed
are PeopleSoft queries, done by their query tool(YUCK), that have all of their
data in memory and are just spinning around and around the same data over &
over. It's a genuine case of BAD SQL being created with lots of subqueries,
etc... These one can see by looking at top sessions for CPU utilization. The
session in question will almost always be at the top ot the unix top commands
results stack also. There is no disk io associated with the session, just one
heck of a pile of memory manipulation. In actuality the BHR for this session &
the databse in general at that time (PeopleSoft HR) is 100%. So again I agree
with Anjo, BHR is not the silver bullet.
Dick Goulet
____________________Reply Separator____________________
Author: "MacGregor; Ian A." <[EMAIL PROTECTED]>
Date: 8/8/2002 3:48 PM
Sorry hit the return too quickly, Resuming
and
select a.aid, a.chnum, f
from rollup a
where a.effdt=(select max(b.effdt) from rollup b where
a.chnum=b.chnum)
Admittedly they are differ by the aggregate function (sum). The one with the
sum was killed by the user, but statspack captured the following
EXECUTIONS DISK_READS ROWS_PROCESSED SORTS PARSE_CALLS BUFFER_GETS
-------------------- -------------------------
---------------------------------- ---------- ------------------------
-----------------------
1 2922 0
-7833 1 9262161 for the first query
1 1498 60
130 1 11450 for the second.
It is obvious that the first query is more expensive than the second. However
the first query has a better BHR than the second. So it should run faster
:) I trap v$session_waits for active sessions every 5 seconds. Admittedly not
as good as a trace, but it does give me some idea of what was going on when the
query was run.
I trapped 48 wait events for the first query and only one for the second. The
one wait trapped for the second query was not significant, and nearly all the
waits trapped for the second query were not significant either, except perhaps
in their number. I did see a few "interesting" 'direct path read waits'
interspersed with 'db file sequential read' waits indicating I/0 contention.
The query plans for the two statements were identical save the extra group by
for the summation.
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
-------------------------------------------------------------------------------
In conclusion, If I just looked at the BHR for the first state and were of the
opinion. "the higher the better", there should be no need to proceed further;
.999684622 is pretty darn good.
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
I don't pay daily attention to the BHR. Most of my time is spent ferreting out
bad SQL. So every morning I look for statements involved with significant
waits, look for high disk reads, buffer gets, and check to make sure the full
table scans are legitimate. If there are problems less expensive transitive
queries are written or the developer is instructed to "materialize" a complex
join as part of the process. Drop table followed by CTAS can solve innumerable
problems.
Ian MacGregor
Stanford Linear Accelerator Center
[EMAIL PROTECTED] <mailto:[EMAIL PROTECTED]>
[MacGregor, Ian A.] -----Original Message-----
Sent: Thursday, August 08, 2002 2:44 PM
To: Multiple recipients of list ORACLE-L
First, I didn't see the broadcast. I believe the claim that a high BHR may not
indicate a healthy system, has become it is always indicative of an unhealthy
system.
Egad, by BHR is over 90% and no users are are complaining, I better start
tinkering! Isn't the idea that BHR is an unreliable indicator
.
Here are two statements :
select a.chnum, sum(a.f) f from rollup a where a.effdt=(select
max(b.effdt)
from rollup b where a.chnum=b.chnum)
group by a.chnum
-----Original Message-----
Sent: Thursday, August 08, 2002 1:31 PM
To: Multiple recipients of list ORACLE-L
Well, I guess that I disagree. Buffer hit radio does matter as one of the
performance indicators, but
certainly not the only one. Your and Mr. Milsap thesis is that LIO also is very
expensive and its cost
is far from being negligible, so having gazillion of LIOs instead of 100 times
smaller number of PIOs will
not make our system run faster. BHR alone cannot be used to judge to overall
health of the system, but
thebn again, there is no such thing as the "overall health of the system". It's
the users of the system who
will say whether the performance is satisfactory or not, and I'm usually tuning
an application, not an
imaginary "overall system". Low cache hit ratio usually tells me that I do have
a hog who is using lots
of PIOs. By my experience, it usually is a very good indicator that something is
wrong, at least on an OLTP
system. So, after all, I do find BHR a useful indicator, but by no means the
only one or the most important
one. Event 10046, SQL_TRACE (level 1 of 10046), explain plan and v$session_event
still are the tools
I need most, but I still do need BHR as an indicator.
Mladen Gogala
Oracle DBA
Phone: (203) 459-6855
Email: [EMAIL PROTECTED]
-----Original Message-----
Sent: Thursday, August 08, 2002 1:05 PM
To: Multiple recipients of list ORACLE-L
Moi wrong ;-) Jeeh, human after all
To summarize the webcast:
db-block-buffers do mattter. Too many LIO do matter. Too many PIO do matter. But
Buffer Cache Hit ratio doesn't matter ....... End user satisfaction does matter.
I am always willing to clarify any points that I made, you just have to ask me l
....
Anjo.
----- Original Message -----
To: Multiple recipients of list <mailto:[EMAIL PROTECTED]> ORACLE-L
Sent: Thursday, August 08, 2002 5:43 PM
Guys,
I had this dream that I missed the webcast - which I did. However, someone said
it wasn't very interesting but the conversation of the people (gurus) left over
was very interesting as there was good solid evidence that he was incorrect and
db_block_buffers do matter. Kind of inline with the discussion about redos
yesterday and my indexing/partition issues - hmmm.
-----Original Message-----
Sent: Thursday, August 08, 2002 1:38 AM
To: Multiple recipients of list ORACLE-L
Www.precise.com, go to Events->webcasts...
On 2002.08.08 00:53 Madhusudhanan Sampath wrote:
> Are transcript documents available anywhere?
>
> Regards
> Madhusudhanan S
>
>
>
> _________________________________________________________________
> MSN Photos is the easiest way to share and print your photos:
> http://photos.msn.com/support/worldwide.aspx
<http://photos.msn.com/support/worldwide.aspx>
>
> --
> Please see the official ORACLE-L FAQ: http://www.orafaq.com
<http://www.orafaq.com>
> --
> Author: Madhusudhanan Sampath
> INET: [EMAIL PROTECTED]
>
> Fat City Network Services -- (858) 538-5051 FAX: (858) 538-5051
> San Diego, California -- Public Internet access / Mailing Lists
> --------------------------------------------------------------------
> To REMOVE yourself from this mailing list, send an E-Mail message
> to: [EMAIL PROTECTED] (note EXACT spelling of 'ListGuru') and in
> the message BODY, include a line containing: UNSUB ORACLE-L
> (or the name of mailing list you want to be removed from). You may
> also send the HELP command for other information (like subscribing).
>
--
Mladen Gogala
--
Please see the official ORACLE-L FAQ: http://www.orafaq.com
<http://www.orafaq.com>
--
Author: Mladen Gogala
INET: [EMAIL PROTECTED]
Fat City Network Services -- (858) 538-5051 FAX: (858) 538-5051
San Diego, California -- Public Internet access / Mailing Lists
--------------------------------------------------------------------
To REMOVE yourself from this mailing list, send an E-Mail message
to: [EMAIL PROTECTED] (note EXACT spelling of 'ListGuru') and in
the message BODY, include a line containing: UNSUB ORACLE-L
(or the name of mailing list you want to be removed from). You may
also send the HELP command for other information (like subscribing).
Text Item
Description: Binary data
