IBM Mainframe Discussion List <[email protected]> wrote on 06/13/2007 12:37:38 PM:
> Ordinarily, as PR/SM dispatches CPs to LPARs, a Logical CP may have a > task dispatched to it, but beneath it, it has no physical CP. > Milliseconds later, the CP may well be back and the task actually runs. > But when we had our 2064-104 and it's cap kicked in, the CP was gone for > awhile, not for a few seconds at a time but for minutes, and response > time suffered noticeably. It seems that, with fewer engines, and a > steeper climb to exceed the cap, an engine may be gone for quite awhile, > compared to a several-engine system whose cap is closer to machine > capacity. It was also being capped during lunch time, when customarily > our usage dropped anyway, so we did not logically need that 4th CP. But > the distinct impression we got was that MVS was not aware that one LCP > had become only a phantom, and MVS had not varied a CP offline, but > instead continued to attempt to dispatch work to it. > > 1. Given that an over-busy LPAR looses one or more CPs from being > dispatched for more than milliseconds, does MVS know that the CP is > gone? (Currently we have a 2094-S08.) As the MVS scheduler dispatches, > does a 4-engine LPAR's MVS treat the situation internally as if it still > had 4 engines, while one of them is out-to-lunch due to capping, or is > MVS instead informed, varies the CP offline internally, and dispatches > its work on just the remaining 3 CPs? It would seem that, if the > situation has become significant enough for an engine to be offline long > enough (minutes to an hour) but without MVS's knowledge, then it might > be more beneficial if (perhaps by automation) an engine was officially > varied offline, allowing work that is dispatched to actually run, rather > than having the TCB imagine that it is going to run on a LCP whose PCP > is not there anymore. Of course the follow-up question wonders whether > the now-3-CP environment will have one of its PCPs taken away as PR/SM > figures that it has to actually pinch somewhere to fix its average? > Observations? And how has the group-of-capped-LPARs worked? LPAR dispatching does not work the way you are suggesting. LPAR spreads a zone's share equally among the online Logical CPs in the zone. You may see a Logical CP getting dispatched only once every 500 milliseconds, but that is because MVS has it in a wait due to Alternate Wait Management. If the Wait bit is not on in the PSW, a Logical CP gets an equal share of the zone's share. Jim Mulder z/OS System Test IBM Corp. Poughkeepsie, NY ---------------------------------------------------------------------- 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
