[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference cpdev::opencp

Title:	Capacity Planner for Open Systems

Moderator:	BCMPQN::GRANT

Created:	Tue Aug 17 1993
Last Modified:	Tue Jun 03 1997
Last Successful Update:	Fri Jun 06 1997
Number of topics:	708
Total number of notes:	2503

667.0. "VAX/Alpha CPU utilization descrepencies on model" by NZOV02::LAYMALCOLM () Sun Feb 02 1997 23:35

    Hi,
    		I've built a model from some DECps data I collected of a
    VAX 4300 running OpenVMS V6.1. This machine has about a 70% utilization
    factor across a typical days workload BUT has a high (10-15) COM queue
    at the same time according to DECPS-PA results.
    
    Most of the PA output revolves around the CPU bottlenecks  / high COM
    queue etc. which is fine because I know its there.
    
    However, when I change the CPU in the Capacity Planner to something
    like a Alphaserver 1000 to model the effects of upgrading the CPU,
    utilization  goes down to about 4%. This is fine (?) except I have this
    nagging feeling that the COM queue is still there as throughput etc
    doesn't go up at all.
    
    Has anyone had experience in modelling a scenario like this and if so,
    are these results about what you'd expect to find? I'm having trouble
    giving these fugures to a customer as I don't really beleive them
    myself.
    
    Capacity Planner is V3.2-07
    
    Thanks in anticipation,
    
    Malcolm Lay.

T.R	Title	User	Personal Name	Date	Lines
667.1		BSS::JILSON	WFH in the Chemung River Valley	`Mon Feb 03 1997 10:15`	10
	This the inherent danger of modeling a saturated time period. The modeler has no idea how much demand is not being serviced. Restart this exercise by taking a time period that doesn't have a cpu queue and build a model from it. Then increase the load on this model to represent a peak time period and then change the cpu. Your results should be very different. The big task is finding a non-saturated time period that still has the same workload mix as a saturated time period and figuring out how much to increase the model so that it then represents a peak period. Good luck. Jilly
667.2	Alpha/VAX CPU descrepency #2	NZOV02::LAYMALCOLM		`Mon Feb 03 1997 23:05`	16
	Thanks for your quick response. I modeled a time period in the evening when COM Q was averaging about 3 which was as low as I could find with the same workload mix. I managed to increase the load so that I/O's /CPU util% etc. was "roughly"the same and then I switched to the Alpha CPU and - same results! Thruput doesn't go up, just CPU util right down. If I saturate the model the first thing to blow out is disk which is already a bit unbalanced/over utilized (but thats another story). So either I accept that a AlphaServer 1000 reduces util from 70% to 4% or I'm still missing the point somewhere? Malcolm.
667.3	...	BCMPQN::GRANT	Roger Grant	`Tue Feb 11 1997 11:15`	8
	Malcolm, Could you make available your .merg and .memory files? Regards, Roger
667.4		STM009::FRASER		`Wed Feb 12 1997 18:27`	12
	re: .2 I think you are missing the point, if what you said about "blowing out the disk"then the model is right in saying that utilization goes down. The real problem appears to be as you stated the need to balance the disk/File I/Os aacross more spindles and/or buses. Therefore I would suggest balancing the I/O and/or adding spindles first, then rerun the model to show even better utilization. Don..