DASD Response Time on antique 3390 4360

David Day

recent post regarding number of users supported by vm370 on old 32mbyte 3081 system Personal Computer

i had written an analysis that disk relative system thruput had declined by an order of magnitude over a 10-15 year period ... or otherwise the 3081 system could have supported on the order of 3000 users ... i.e. the number of users supported actually increased proportional to the increase in disk thruput ... not cpu or memory resources. initially the disk division performance group was buttigned to refute my statements ... however, after several weeks, they came back and observed that i had slightly understated the problem. they eventually turned the analysis around into share presentation on how to configure disks and system to improve thruput (somewhat implicit implication that it was compensation for disk thruput bottleneck).

from long ago and far away

SHARE 63 Presentation B874

Acknowledgements

This review makes liberal use of the computer science literature. As usual, the views expressed in this report are those of the author. Many contributed facts and ideas, but the selection and presentation are the author's responsibility, including any mistakes. I am especially indebted to Lynn Wheeler for pointing out how the relative speeds of things have changed over the years, to Brian J. Smith for helping me through many of the intricacies of attachment modeling, to Bill O'Brien for suggesting this review, and to my manager, Steve Goldstein, for his patient support throughout these activities.

... snip ...

somewhat related to that there was general trend thru-out the last half of the 70s with the increasing use of electronic storage to compensate for disk thruput bottleneck.

we had done some kernel hooks to capture record numbers from disk accesses for modeling things like file activity caching. this was installed in a variety of different systems in the san jose area ... some interactive online systems .... some much more commerical batch oriented. the traces were then run thru model of various kinds of caching configurations and strategies. one of the results was for a fixed amount of electronic storage ... and all other factors being equal .... a single global system file cache always provided better performance that subdividing the electronic storage into channel, controller, and-or drive caches.

you do find electronic storage in controllers and-or other places .... in addition to use of system memory for global system cache ... possibly because there are different kinds of memory technology being used in the different locations (invalidating the model buttumption regarding all other things being equal, possibly things like controller memory being less expensive than global system memory). another scenario is that there have difficencies shown up in some of the global system management strategies ... which can be conpensated for by having (less efficient) parbreastioned management of the resources.

on of the other thing that the detailed i-o record traces started to show up was that some amount of activity had periodic clustering .... i.e. collections of files tended to be used together on periodic basis ... one any specific member of the collection was accessed, others in the collection tended to be used also.

somewhat related recent post regarding doing i-o with zero buffer copies Baby MVS???

some past post mentioning the record traces and cache modeling work distribution (Was: Re: A primeval C compiler) distribution (Was: Re: A primeval C compiler) mainframe?" outer cylinders still faster than thereof