If you build a response time model for processors, - AND you have a
target response time not to be exceeded, it is easy to show that 1
processor responds worse at 80%, than two at 80%. Equivalent response
time is expected when the two processors are at 90%. So the source of
the question is really "batch" mentality vs the "response time" mentality.
MP effect comes from the batch mentality where thruput was the only
measure. The batch mentality will always challenge this, response time
mentality should understand.... If you care about response time in the
Linux/zVm world, you don't run at 100% most of the time.
So the only time the MP Effect question is relevant is when both
processors are running at 100%, which makes the question not relevant on
IFLs. From an accounting perspective, I guess you could use the z/OS
numbers, which would likely under-charge the Linux user for CPU
consumed, since using those numbers a CPU second consumed is not charged
as a full CPU second.
Schuh, Richard wrote:
I would expect that some would challenge your conclusion based on the
idea that the MP effect does not even appear unless you are running at
or near capacity. If I have two cpus or IFLs and 1.1 cpu's worth of
demand, will I notice the MP effect? Probably not. I probably will see a
better service level than when I was trying to service the same demand
with only 1 cpu. The question is, if n tasks causes a single engine to
run at 100%, will 2 engines be able to service 2n tasks as well as 1
serviced n? I think that under normal circumstances, the answer is that
the 2 engine machine will only be able to service somewhat less than 2n.
Regards,
Richard Schuh
-----Original Message-----
From: The IBM z/VM Operating System
[mailto:ib...@listserv.uark.edu] On Behalf Of Barton Robinson
Sent: Tuesday, February 03, 2009 10:57 AM
To: IBMVM@LISTSERV.UARK.EDU
Subject: Re: Correcting Statements From Marketing
Ok here's some heresy that I've presented to IBM and maybe
was communicated to their sales folks. From a capacity
planning and service level perspective, adding a CPU gives
you MORE than 100%, not less than.
Really, BUT ONLY if you actually care about service levels.
From a service level perspective, i know that i can provide
on ONE IFL a given service at 80% CPU utilization. If I ADD
an IFL, and more work of a similar nature, I now have TWO
IFLs, and I know that I can provide that SAME service at 180%
CPU Util.
So, I went from ONE IFL, to TWO IFLs, and increased my target
CPU utilization by 1.25 times.
On z/OS if you just run at 100% all the time, and run batch
to soak up cycles, then add a CPU and you don't get 100% of
one CPU more work done.
That is the only time MP factors should matter.
And this heresy is why it is much easier to deal with
installations running multiple IFLs, because the performance
will be better at higher utilizations than single IFLs at
lower utilizations. Adding a second IFL more than doubles
your usable capacity. Adding a 3rd or 4th is less dramatic.
From a historical perspective, we used to have the MASTER
PROCESSOR effect where adding a CPU added much less capacity.
Installations today do not see this impact.
Schuh, Richard wrote:
This got no response when posted under a different topic:
"Yikes, We have someone from IBM Marketing now making the
statement,
"I have confirmed...no MP factor with IFLs....". That is the entire
statement, all of the dots included. I did not replace
anything with
ellipses. Somehow, that does not ring true. I mentioned that the
rating of an IFL is the same as that of an ordinary CPU and someone
went to marketing for "the real answer". Perhaps they should have
said, "No different MP factor for IFLs than for regular
CPUs, they are
the same in that regard." That would make more sense.
Anyone from IBM
care to comment - you will probably be quoted."
I am not considered an authority on the topic, especially when I
disagree with an interpretation of a statement made by IBM
marketing.
I need to disabuse someone of their notion because it will
affect the
capacity planning process. They do not seem to believe that running
the same O/S on two systems, one with n standard CPUs and the other
with the same number of IFLs will produce a result of equal
MP effect.
Barton, you are also invited to respond. At least one of
the people on
the other side of the fence will take your word for it.
Regards,
Richard Schuh
