winscape 2345

the science center had done a lot of work on performance monitoring ... and created some performance monitoring tools that accumulated data 7x24 ... first on the science center machine

and then on hundreds of other machines (eventually accumulated something like decade of information). this help motivate the work on synthetic workloads, benchmarking, performance modeling, performance profiling and the work that eventually evolved into capacity planning.

on of the applications was an apl-based analytical model that evolved into the marketing support tool available on hone called performance predictor

that allowed marketing people to input information about customer configuration and load and ask what-if questions (changes to load and-or hardware).

when i was getting ready to release the resource manager & fair-share scheduler

a wide spectrum of synthetic workloads and configuraitons were evaluated for calibrating the operation of the resource manager. then a modified version of the performance perdictor was created which examined a broad range of customer workload and configuration data and created an automated benchmark specification (configuration, system options, synthetic workload characteristics). the benchmark would be run and the results feed into the modified predictor. The predictor validated the expected results for the resource manager and then also used the information as input for choosing the next benchmark configuration. This process itereated eventually for 2000 benchmarks that took three months of elapsed time to run.

note that the overall system technology was not only used for a lot of commercial and gov. installations ... but also used as the basis for a number of commercial time-sharing offerings (including the hone system that provided world-wide field, sales, and marketing support):

some minor drift