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: ASSM in 10g RAC doesnt seem work that well

Re: ASSM in 10g RAC doesnt seem work that well

From: Martic Zoran <zoran_martic_at_yahoo.com>
Date: Tue, 29 Mar 2005 05:10:56 -0800 (PST)
Message-ID: <20050329131056.21473.qmail@web52606.mail.yahoo.com>

Just add-on.

The script down is for normal heap table with PK and not IOT as I said.
I also made the mistake by producing 900 duplicates, so the difference was not that huge as it should be.

I changed the script and do it again without duplicates that are making things bad for both ASSM and non-ASSM :)

I have got this on 1GHz Sun with 10.1.0.3:

CPU used by this session               
CPU used when call started             

ASSM - 164
non-ASSM - 731

Almost 5 times more CPU when doing bulks

On HP-UX 11.00, 360MHz 9.2.0.5:

CPU used by this session               
CPU used when call started             

ASSM - 361
non-ASSM - 518

On Solaris 7, 450MHz 9.2.0.5:

CPU used by this session               
CPU used when call started             

ASSM - 194
non-ASSM - 324

Solaris 9, 450MHz, 10.1.0.3 RAC with ASM:

CPU used by this session               
CPU used when call started             

ASSM - 312
non-ASSM - 1521

Solaris 9, 450MHz, 10.1.0.3 RAC with ASM:

CPU used by this session               
CPU used when call started             

2 processes running on different instances ASSM - 440
non-ASSM - 2359

1 process
ASSM - 369
non-ASSM - 1562

CONCLUSION


  1. ASSM is using less CPU over non-ASSM in all tests Different tests, the same behaviour, ASSM is just spending less CPU. I did, where I pushed it with the fastest bulk inserts and so removing other bottlenecks causing this difference to be deminished
  2. 10g improved ASSM or worsen something else, because the difference in 10g is 4-5 times while in 9i is not even 2 :) Also probably because of additional statistics overhead in 10g it is slower (using more CPU) then 9i as we all know as true :)
  3. In high concurrent environments (as nicely Jonathan said earlier) more cpu spinning causing CPU time to increase. It increased for me as coming from 1 to multi session environment. But it increased for both ASSM and non-ASSM.

I could not correlate any statistics consistently to why is this the case when comparing ASSM and non ASSM. When comparing 1 sessiong against multiple I can find out some diffferences. Basically I had worse statistics for ASSM in most cases as expected per Jonathan's mail telling about the algorithm behind (like buffers, ...).

do not forget that these numbers in your special case should be evaluated.

But I learned the lesson how drastically these two things can change the performances (CPU at least) of high speed DML:
1) ASSM
2) spinning (thanks Jonathan, I totally forget to get this into account while thinking about CPU time even reading that small Internal's book a few times :)

Regards,
Zoran


Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
--
http://www.freelists.org/webpage/oracle-l
Received on Tue Mar 29 2005 - 08:14:42 CST

Original text of this message

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