Oracle FAQ Your Portal to the Oracle Knowledge Grid
HOME | ASK QUESTION | ADD INFO | SEARCH | E-MAIL US
 

Home -> Community -> Mailing Lists -> Oracle-L -> RE: see higher CPU usage after increase SGA

RE: see higher CPU usage after increase SGA

From: Cary Millsap <cary.millsap_at_hotsos.com>
Date: Thu, 10 Jun 2004 21:40:58 -0500
Message-ID: <016101c44f5d$8a119f20$6701a8c0@CVMLAP02> 





Answers in-line below.




Cary Millsap


Hotsos Enterprises, Ltd.


http://www.hotsos.com


* Nullius in verba *



Upcoming events:

- Performance Diagnosis 101: 6/22 Pittsburgh, 7/20 Cleveland, 8/10 Boston
- SQL Optimization 101: 5/24 San Diego, 6/14 Chicago, 6/28 Denver
- Hotsos Symposium 2005: March 6-10 Dallas
- Visit www.hotsos.com for schedule details...






-----Original Message-----


From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org]
On Behalf Of zhu chao


Sent: Thursday, June 10, 2004 1:40 AM


To: oracle-l_at_freelists.org


Subject: Re: see higher CPU usage after increase SGA



Thanks Cary.



I am not quite agree with your definition: Utilization and Usage.
Utilization = usage_during_time_initerval/capacity_during_time_interval.
During the same time interval, capacity is the same. So utilization
increases  is just the same as usage increase. Right?
Before and after my tune effort(increase data buffer by 1GB), the statspack
report time interval was the same:


9am to 15pm.


And unix CPU usage report was also between 9am to 15pm.
So I think Utilization increased by 5% is the same as CPU usage increased by
5%. Do you think so?



[Cary Millsap] Yes, if you're measuring over the exact same time period, AND
if the fact that your "tuned" job is finishing sooner doesn't lead to more
[other] work being squeezed into the second 15-minute interval. My suspicion
is that after you reduced the response time of the job you're interested in,
the 15-minute interval now contains a bit more work (that is, "usage"
defined my way) than it did before. In other words, I'm proposing that think
your CPU usage increased, but not because of the job you're trying to
measure. And if this is true, then it's not a fair analysis, and you've
fallen prey to a ratio fallacy.



Let's forget the above and go on analyze why CPU usage increased.



Possible Reasons that CPU usage can decrease: Physical IO dropped by
30%.(that is, About 100-120DiskRead/Second)
Possible Reasons that CPU usage can increase:  Larger data buffer, same
db_block_lru_latches (=4), cache buffers chains latch is same with 10gb/11gb
data buffer. So each latch does manage more block buffers.



[Cary Millsap] Maybe more CPU usage consumed spinning for latches. Maybe.
But you won't know until you measure it.



My answer to your question:


1. Sorry I did not measure the application response time. As there are too
many connections to the database.



[Cary Millsap] But response time is practically the only metric that really
matters...



    I *GUESS* there won't be noticable speedup in the program, as database
layer is just maybe 30-50% of total application response time, and the
application WAS pretty fast, most of them finished in  0.001-0.05Second.
2.  What is most important to me is: Scalability, that is, to support more
user, more transactions. As this is difficult to measure how much more
business we can support, we use System Resource Usage percent to estimate
it. So sorry, the answer to your second question is: Ratio.  Do you have
some suggestion about measure capacity other than the resource utilization
ratio? Thanks.



[Cary Millsap] Jared's right. Ratios will never work for you in this
pursuit.



    
  
  1.   Fewer LIOs means better scalability.

  
  2.   Fewer latch acquisitions means better scalability. (How to acquire fewer
latches? See #1.)

  
  3.   Fewer dbcalls means better scalability. (For example, process arrays
instead of one row at a time.)







Regards


Zhu Chao.




The cache buffer chains latch activity report during 9am-15:00.

                                Get                                  Spin &
Latch Name                    Requests         Misses      Sleeps Sleeps


1->4


(11gb)


cache buffers chains        1,868,578,608     449,111      13,588
435738/13175

                                                                  /186/12/0
cache buffers lru chain         3,838,481       2,460         476


1985/474/1/0



(10gb)


cache buffers chains        1,874,057,944     588,231      16,539
572097/15757

                                                                  /367/10/0
cache buffers lru chain         4,499,478       3,228         122


3107/120/1/0






> Zhu,


>




> I think what’s happening to you is strictly a function of how you’re

> looking at your measurements.


>




> First, I like to distinguish between the terms “usage” and

> “utilization.” (You’ve used them in your post as synonyms.) This

> distinction helps me make a point. So, first, some definitions:


>




> Usage - the total duration for which a resource is “in use.”


>


>        Utilization - usage divided by capacity for a specified time



> interval.



>


> Now, “utilization” is a concept inherently susceptible to ratio

fallacies.


> This is because utilization, of course, IS a ratio. I think the ratio

> fallacy that’s victimizing you right now is this one...


>


>        Conjecture: If CPU utilization went up, then performance must have



> gotten worse.



>


> This conjecture is false. It’s a commonly-believed rule of thumb, but it

’s


> just not right. Here’s what I think has happened to you. When you reduced

> the [so-called] physical I/O count, you probably reduced the total

run-time


> of your program. But of course, you have kept your CPU usage constant (you

> haven’t reduced the LIO count by making your buffer cache bigger). So

look


> at the impact of what you did upon utilization. Remember,


>




> Utilization = (Usage) / (Capacity for a specified time interval).


>




> You have kept usage constant, but you’ve reduced the specified time

> interval. Therefore, your utilization went up.


>




> The fallacy is that when CPU utilization goes up, it can also be an

> indication that you’re simply able to get more work done now in the same

> time interval as before.


>




> Now, my questions for you:



>


> 1. Did your action reduce your total response time by some small

percentage


> (maybe 5-6%)?

> 2. What’s more important to you: (a) Utilization statistics? Or (b) total

> runtime?


>




> I hope your answer to question 2 is (b), because (a) is a flawed metric;

> utilization is NOT a reliable measure of the quality of PERFORMANCE.

Answer


> (b) does not leave you vulnerable to the ratio fallacies that (a) does.


>
>




> Cary Millsap

> Hotsos Enterprises, Ltd.

> http://www.hotsos.com

> * Nullius in verba *


>




> Upcoming events:

> - Performance Diagnosis 101: 6/22 Pittsburgh, 7/20 Cleveland, 8/10 Boston

> - SQL Optimization 101: 5/24 San Diego, 6/14 Chicago, 6/28 Denver

> - Hotsos Symposium 2005: March 6-10 Dallas

> - Visit www.hotsos.com for schedule details...


>
>




> -----Original Message-----

> From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org]

> On Behalf Of Mark W. Farnham

> Sent: Wednesday, June 09, 2004 9:04 AM

> To: oracle-l_at_freelists.org

> Subject: RE: see higher CPU usage after increase SGA

>


> okay, I'm a sucker for mystery puzzles with partial information. Take a

look


> at Hotsos' site to get an idea of figuring out where you're burning your

> time. In the meantime, from what you've written I have these observations:

>


> 1) Neither disk farm/io channel capacity nor CPU utilization is what

limits


> your performance to 76.54/sec.


>




> -- How can I say this? Well, in the boundary case that you exactly shifted

> the bottleneck from disk to CPU I'd be wrong,

> but otherwise you previously had CPU to burn and now you've got io

capacity


> to burn, and the rate didn't change.



>


> 2) Probably the increase in CPU utilization is just searching longer

memory


> structures in Oracle, although I suppose it could be something like a bit

of


> increased paging. OR, now that you've marginally reduced waits for disk

> reads you're pounding a bit harder on the actual bottleneck.


>




> Someone might have an educated guess at your new bottleneck, but without

> data about where your processes are spending time and which resources are

in


> short supply compared to load, I'm guessing all they'd be doing is

guessing.

>




> Now, trying to be actually useful, I think your next task is to find out

> where time is being spent. If analysis shows that the big consumer(s)

is/are


> already doing as little work as possible, then you make the restrictive

> system component faster (or discover that it is not cost effective to make

> the bottleneck resource faster.) If analysis shows that big consumer(s)

> is/are doing more work than required for the task at hand, you work to

> improve the big consumer(s).


>




> mwf


>




> -----Original Message-----

> From: oracle-l-bounce_at_freelists.org



[mailto:oracle-l-bounce_at_freelists.org]On


> Behalf Of zhu chao

> Sent: Wednesday, June 09, 2004 9:31 AM

> To: oracle-l_at_freelists.org

> Subject: see higher CPU usage after increase SGA


>
>


> Hi,

>     I once saw Jonathan said at metalink that huge SGA does not help in

many




> case, But no further discuss at that topic later.  Last night we added 1Gb

> to oracle sga and we see fewer disk read but higher CPU usage.


>


>     Fewer disk read of course cut CPU usage, but larger buffer cache



> management in unix and oracle, seems caused higher CPU usage. Has someone

> also have similar experience? How to explain the higher CPU usage?


>


>     We have a 16GB memory sun 880 with 10G data cache. As disk read get



> higher and higher , and not much SQL to tune we deciede to increase data

> buffer from 10G to 11GB, as there is still 1.5G free memory on the host.


>


>     We expect to see some CPU usage drop, as disk read drop by 30%. But



> after 1 day's run, we saw higher CPU usage then before we increase the

SGA.


>


>     http://www.cnoug.org/attachments/LDBn_cpu.bmp (the Excel picture that

> shows the CPU usage before and after increase sga).

>     The following Statistics from Oracle shows the load profile before and



> after SGA increase:


>


>                   LIO                    PIO            Transaction/Second

> CPU usage in oracle

>       10gb        47,990.70              448.68               76.54

> 177.9

>       11gb        47,707.28             325.95                76.54

> 187.9

>       Change:     Nearly same           Disk read dropped   Transaction

rate


> CPU used increased.

>                                           30%               keep

consistent




> by 5%


>
>


>       Time I measure： 9 am - 15pm.

>       Oracle: 5% increase.

>       Unix:    6% increase.

>
>


> ----------------------------------------------------------------



> Please see the official ORACLE-L FAQ: http://www.orafaq.com

> ----------------------------------------------------------------

> To unsubscribe send email to:  oracle-l-request_at_freelists.org

> put 'unsubscribe' in the subject line.

> --

> Archives are at http://www.freelists.org/archives/oracle-l/

> FAQ is at http://www.freelists.org/help/fom-serve/cache/1.html

> -----------------------------------------------------------------


>
>


> ----------------------------------------------------------------



> Please see the official ORACLE-L FAQ: http://www.orafaq.com

> ----------------------------------------------------------------

> To unsubscribe send email to:  oracle-l-request_at_freelists.org

> put 'unsubscribe' in the subject line.

> --

> Archives are at http://www.freelists.org/archives/oracle-l/

> FAQ is at http://www.freelists.org/help/fom-serve/cache/1.html

> -----------------------------------------------------------------


>


> ----------------------------------------------------------------



> Please see the official ORACLE-L FAQ: http://www.orafaq.com

> ----------------------------------------------------------------

> To unsubscribe send email to:  oracle-l-request_at_freelists.org

> put 'unsubscribe' in the subject line.

> --

> Archives are at http://www.freelists.org/archives/oracle-l/

> FAQ is at http://www.freelists.org/help/fom-serve/cache/1.html

> -----------------------------------------------------------------

>


>





Please see the official ORACLE-L FAQ: http://www.orafaq.com




To unsubscribe send email to:  oracle-l-request_at_freelists.org
put 'unsubscribe' in the subject line.

--
Archives are at http://www.freelists.org/archives/oracle-l/
FAQ is at http://www.freelists.org/help/fom-serve/cache/1.html
-----------------------------------------------------------------

----------------------------------------------------------------
Please see the official ORACLE-L FAQ: http://www.orafaq.com
----------------------------------------------------------------
To unsubscribe send email to:  oracle-l-request_at_freelists.org
put 'unsubscribe' in the subject line.
--
Archives are at http://www.freelists.org/archives/oracle-l/
FAQ is at http://www.freelists.org/help/fom-serve/cache/1.html
-----------------------------------------------------------------


Received on Thu Jun 10 2004 - 21:39:24 CDT












Original text of this message













  HOME |
  ASK QUESTION |
  ADD INFO |
  SEARCH |
  E-MAIL US