Thanks Ric. The gathered values for cpuspeednw do appear to vary; accross several identical servers I see values varying by +/- 100 or so. Research by some of my colleagues concludes that cpuspeednw is not particularly important because its influence on the overall cost calculated by the optimizer is very small. Its variance accross different types of CPUs (say 400 to 2000 or so, depending on the type and age of the box) is not enough to be a significant cost factor. The single block read times are generally accurate, as the SAN throughput has not changed much over the last few years. So mreadtim, sreadtim stats gathered several years ago are still ok - this is the critical number (from system statistics) when calculating cost; not cpuspeednw/cpuspeed. We will test this by regathering on a non-production environment and replay lots of SQL to confirm that the execution plans dont change. Mark ________________________________ From: Ric Van Dyke [mailto:ric.van.dyke@xxxxxxxxxx] Sent: 16 January 2010 00:43 To: Teehan, Mark; oracle-l@xxxxxxxxxxxxx Subject: RE: Impact of an incorrect CPUSPEEDNW I've been running some tests here on my laptop and the CpuSpeedNw value is not constant as I run the gather stats at different times with different loads going on. It's not fluctuating by much, about 2% from the lowest value to the highest value. The fact that it fluctuates is a strong indicator that Oracle is "doing something" to calculate the CpuSpeedNw value and is not just asking the CPUs how fast they are. Testing done on a Dell Latitude with T2500 dual core CPU 2GHz, Oracle 11.1.0.7 EE Lowest value - 1009.655 Highest Value - 1031.383 ----------------------- Ric Van Dyke Hotsos Enterprises ----------------------- Hotsos Symposium March 7 - 11, 2010 Be there. ________________________________ From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx] On Behalf Of Teehan, Mark Sent: Thursday, January 14, 2010 11:22 PM To: oracle-l@xxxxxxxxxxxxx Subject: Impact of an incorrect CPUSPEEDNW I am reviewing system statistics for a number of database servers running various 10g releases. Most are using default NoWorkload settings; but I've noticed something odd: some databases have unexpectedly low values for CpuSpeedNw even though they are running on identical servers. This is because they were migrated from older servers and sys stats were not regathered for the new servers. The database will not automatically recollect system stats unless they are explicitly deleted. This seems an oversight: given that some values like CPU_COUNT are rechecked every time the database starts; I wonder if CpuSpeedNw should also be rechecked automatically. In some cases the numbers are about 30% what they should be; which means that the optimizer calculations are incorrect. I wonder if regathering them - I don't want to do this unless I can confirm that it will make a significant difference - after all the database has been subsequently tuned to run with a lower CpUSpeedNw setting. Something else that I'm checking- if you do an RMAN clone, or an import FULL=Y, does it pull over the system stats too? I also posted this to forums.oracle.com. Any ideas? Mark ======================================================================== ====== Please access the attached hyperlink for an important electronic communications disclaimer: http://www.credit-suisse.com/legal/en/disclaimer_email_ib.html ======================================================================== ====== =============================================================================== Please access the attached hyperlink for an important electronic communications disclaimer: http://www.credit-suisse.com/legal/en/disclaimer_email_ib.html ===============================================================================