Ted, I'd advise a "phased in" approach. And a thorough review of your WLM policy.
1. For any given CEC in a CEC PLEX, determine the minimum number of LCPs you want any image to have. Don't set your minimum weight to so low that IRD can possibly justify varying off LCPs past your threshold. For example, on a 2-Image, 6 PCP CEC where the weights add up to 1000, and you want at least 2 LCPs online to each image, don't set the minimum weight for each to less than 1.01 physical [or (1.01 / 6) * 1000 = 168]. IRD will always keep enough LCPs online to meet your minimum weight, rounded UP to the next whole number of LCPs. In the same example, if you want to keep 3 LCPs online, set your minimum to 2.01 PCPs, or a weight of 335. 2. For any given CEC in a CEC PLEX, determine the maximum number of LCPs you want allowed to each image. This might be all of them. It might be some. If, similar to the previous 6 PCP 2-Image example, you wanted to never have more than 4 on to any Image (thereby guaranteeing a max LCP over PCP of 8/6) then only put on LCPs 0 through 3 and leave 4 and 5 offline. Put too many LCPs online to too many images in a CEC PLEX and you _will_ run short on engines, guaranteed. IRD will not completely prevent short engines - it seems to like a buffer. Even if both images were running 100% and screaming for more, and you allowed each to have 3 or more LCPs, IRD would not back them down to 3 each. You'd get 4, and the short engine effect you'd expect. RMF REPORTS(CPU) is very revealing in these matters. 3. If you ever want IRD to favor workloads of importance 3 and below, you have to "sandbag" your importance 1 and 2 work so that it's goals are 0.8 PI at 100% CEC busy while unconstrained. If you have "important" batch, that really is critical, move it to importance 2. This bit us repeatedly, until we moved the batch up. Ever seen a CPU ready queue climb over 60 and stay there until you manually forced IRD to allow more work through (even though the other LPAR had space to give up)? I have. The phased in approach is to turn on IRD, and set the min/max weights to what you currently have now. PR/SM will behave the same. Then add a few LCPs, and give IRD some "flex" room in the min/max weights, and watch it for a while. I'd say no more than 20% of the CEC worth of flex. Fix any WLM problems you identify, and widen the flex. The theory is that eventually you can turn on ALL LCPs, and set the min weights to 0 and leave the max blank. We're not there yet. Good luck to you. Best regards, Gary Diehl -----Original Message----- From: IBM Mainframe Discussion List [mailto:[EMAIL PROTECTED] Behalf Of Ted MacNEIL Sent: Monday, September 26, 2005 7:00 PM To: [email protected] Subject: IRD (+)ve R (-)ve We are seriously looking at the weight management portion of IRD. Channels and CPUs are not worth it when you're running flat out. And, even with weight management, you need 'white space'. During our peak, we have a bunch of reporting jobs that can wait. ---------------------------------------------------------------------- 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
