The Pankian Metaphor 3132

re:

sort of the evoluation was that computer systems were built to just run. as things got more complex ... especially with various kinds of multiprogramming ... there was a realization that varous kinds of optimization and performance turning could improve throughput.

the problem was that much of the state-of-the-art around the time that i was (re)releasing dynamic adaptive scheduling for the research manager (11may76, not to mention the earlier incantation done in the late 60s) was that there was no really good understanding of the science of performance tuning. some set of performance settings were specified ... and sometimes it improved things and sometimes it didn't. part of the issue was that performance settings could be workload and configuration specific ... with static settings and workload that possibly dynamically changed minute to minute ... there would be no ideal setting.

in any case, the prevalent state-of-the-art at the time for (re-)releasing the dynamic adaptive stuff (11may76) was to try and identify all thruput related decisions in the systems and attach various kinds of control parameters to each of the decision points. You build an enormous specification of all possible control parameters and the types of decisions that they affected. customers then were encouraged to have significant resources devoted to studying thruput, (possibly randomly) changing turning parameters, evaluated the result and presenting detailed reports at user group meetings (like SHARE and GUIDE) about specific customer experiences (randomly) modifying the numerous tuning parameters.

part of the pressure from corporate hdqtrs was that the (other) mainstream operating system product had a significant subculture and folklore around performance tuning (requiring large amount of resources devoting to performance tuning was felt to be representative of an advanced, state-of-the-art customer installation)

the concept that you could have a science of thruput and deploy a dynamic adaptive resource manager based on such principles, was incomprehensible.

as a result i had to come up with the ruse of having "people set" tuning parameters and allow the dynamic adaptive control mechanisms "compete" with the people specific static settings.

part of this is because the science center had spent a lot of effort on instrumenting systems and capturing the data for detailed study and analysis

and was well on its way to evolving things like capacity planning based on the work

--