Ted MacNEIL wrote:
Also, for the poster that asked about CPU usage.
Who cares? This entire CICS sub-system is using less than 5% of the processor.
The only one being impacted is this sub-system.
CICS cannot sustain that rate very long without response implications.
We need to know the cost per I/O to size the work effort to repair or remove.
the science center had done a whole lot of work on performance ...
http://www.garlic.com/~lynn/subtopic.html#545tech
lots of work on instrumenting much of the kernel activity and collecting
the statistics on regular intervals (every ten minutes, 7x24 ...
eventually with historical data that spanned decades).
it also did a lot of even modeling for characterizing system operation
and it did the apl analytical system model that eventually turned into
the "performance predictor" on hone.
hone was an online (cp/cms based) time-sharing system that supported all
the field, sales and marketing people in the world (major replicas of
hone were positioned all around the world).
http://www.garlic.com/~lynn/subtopic.html#hone
eventually all system orders had to be processed by some hone
application or another. the "performance predictor" was a "what-if" type
application for sales ... you entered some amount of information about
customer workload and configuration (typically in machine readable form)
and asked "what-if" questions about changes in workload or configuration
(type of stuff that is normally done by spreadsheet applications these
tdays).
i used variation off the "performance predictor" was also used for
selecting next configuration/workload in the automated benchmark process
that I used for calibrating/validating the resource manager before
product ship (one sequence involved 2000 benchmarks that took 3 months
elapsed to
run)
http://www.garlic.com/~lynn/subtopic.html#benchmark
the performance monitoring, tuning, and management work also evolved
into capacity planning.
there also were some number of instruction address sampling tools ... to
approx. where processes were spending much of the time. One such tool
was done in support of deciding what kernel functions to drop into
microcode for endicott's microcode assists. the guy in palo alto that
had done the apl microcode assist originally for the 370/145 ... did a
psw instruction address sampling application in 145 microcode and we
used it to help select what part of kernel pathlengths to drop into
microcode for kernel performance assist. minor reference
http://www.garlic.com/~lynn/94.html#21 370 ECPS VM microcode assist
However, another technology we used was to take interval activity data
and run it thru multiple regression analysis ... assuming we accounted
for the major activity variables, we were able to come up with pretty
accurate cpu time correlation with different activities. Doesn't take a
whole lot of effort ... just a far amount of recorded activity over
range of activity and feed it into a good MRA routine. At the time, we
primarily used MRA from the scientific subroutine library. There are
some free packages you can find on the web ... but to get one that can
handle both a large number of variables and large number of data points
... you may have to consider a commercial package.
In subsequent years, I've used MRA to find anomolous correlations in
extremely large applications ... unusual activity correlations that
conventional wisdom maintained couldn't possible be correct ... but
actually accounted for major resource consumption. This is analogous to
one of the other postings in this thread about finding things happening
in a large application; that the responsible application people couldn't
possibly believe was happening.
trivial things MRA may turn up is cpu use and i/o activity by major
functions, cpu use per i/o, etc.
----------------------------------------------------------------------
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