There have been several discussion at SHARE (in the EXCP project) over the past couple of years on just this topic. The variability of CPU time has been seen up to 20-30% different from run to run. The hardware itself may be part of the issue. The processors (from z990s to z9s) may slow down when they get too hot, and depending on which instructions are being executed and if the data has been fetched at the same time, the cycle time for the instructions can be very different. You could have a good conversation with Bob Rogers about this when next you see him. EWCP was concerned from a Capacity Planning and Chargeback position. It has been recommended that "banding" be used in Chargeback systems to account for the variability. It was also suggested that this variability will get worse before it gets better. So business as usual needs to be changed to cover this issue for Chargeback. And Capacity planning has just become more complicated.
-----Original Message----- From: IBM Mainframe Discussion List [mailto:[EMAIL PROTECTED] On Behalf Of Dan Gillis Sent: Tuesday, February 05, 2008 SYSN 10:12 AM To: IBM-MAIN@BAMA.UA.EDU Subject: Re: CPU time differences for the same job Some previous posters on this thread discussed running in LPARs with different weights. I wanted to point out that in addition to an MP-effect at the box level, there is also an MP-effect at the LPAR level depending on how many CP's are defined to that LPAR. So in the z9 BC example previously discussed, if you have less CP's defined for your test you will get more throughput per CPU second, resulting in less CPU time to do the same work (of course based on other factors like cache hit already discussed). Back in 2005 Gerhard Adam demonstrated this internal LPAR effect at our site by running test jobs on a 2-way and a 23-way LPAR both running on the same z990. Net, I believe that the only way you will get consistent results is to set up a repeatable benchmark on a dedicated machine, or accept the variance as unavoidable. Dan ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
