Tracy, I believe you are seeing the effects of queuing. Consider the opposite issue. Suppose you are having response time problems and you wish to resolve them. Sometimes adding just a small amount of resource will reduce your response time dramatically. The converse is true: removing a small amount of resource can increase your response time dramatically.
The reason the simple math model does not reflect reality is because each of your transactions response times consist of the sum of times needed of each resource, plus the sum of the times waiting for that resource. So even though the resource usage is unchanged, the wait time for the resource (in this case, the CPU), goes up. Sometimes you can see this in queues at the grocery store or bank: adding a single checker or teller can quickly reduce your time in line. The time to process your transaction is unchanged, but the time you spend waiting got the process to occur is reduced, and you are happy. IBM has a solution for this: Capacity on Demand. You configure your system similar to what you are proposing, but when you need additional capacity, it is provided. It is not free, however. Tom Harper IMS Utilities Development Team Neon Enterprise Software Sugar Land, TX -----Original Message----- From: IBM Mainframe Discussion List [mailto:[email protected]] On Behalf Of Adams, Tracy Sent: Thursday, March 26, 2009 3:14 PM To: [email protected] Subject: Performance problems Okay, this is a continuation of a previous post... First of all we have an 88 mip cpu that is not constrained in any way. RMF cpu intervals are 20% during the day and during the 3 hours of batch 100% like a good MVS system can do. So with the rising cost of software, mainly CICS, we are looking to cut the mainframe's capacity in half. Now in the simplest math, batch should double in time and daily rmf stat intervals will increase but still not hit 100%, as long as no other constraints are revealed. Some basic tests have revealed results that I can't explain. Response time in our IDMS transactional system during the day (as record via PMDC writing smf records translated by MXG). A typical SAS model of performance for a given online transaction would be 95% < .5, 4% < 1, 1 % > 1. When I set a hard cap at 90% the model looks more like 70% < .5, 15% < 1, 10% < 2 and 5% > 2 of that 1% > 3. When I set the hard cap at 75% the model looks more like 50% < .5, 15 < 1, 20% < 2 and 7% > 2 and 3% > 3. And when I set the hard cap at 50% the model looks more like 40% < .5, 25 < 1, 25% < 2 and 10% > 2 and 3% > 3. And the users now users are really complaining now. RMF type 70 records (cpu) for all four scenerios (100%, 90%, 75% and 50%) show averages in the 20% utilized. RMF type 74 records (IO) show avg resp in single digits. UIC hasn't fallen below 255 in 10 years. Batch... completed in the same time frame set at 25% as it did at 100%. So if the hard cap sets the amount of Service units consumed not the actual speed of the processor, why is response time in the online going so far south when the CPU is still running unconstrained? Why did batch not slow down? ---------------------------------------------------------------------- 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
