Multi-modal optimization, old post from 7yrs ago:

Multi-modal optimization, old post from 7yrs ago:

El-Erian is discussing bimodal distribution. The Only Game in Town: Central Banks, Instability, and Avoiding the Next Collapse

https://www.amazon.com/Only-Game-Town-Instability-Avoiding-ebook/dp/B0165I3V4C/

loc3334-36: Remember the dominance of IBM on the eve of the personal computer (PC) revolution. The company had by far the most powerful brand in technology. Each year it deployed a large R&D budget. And it was profitable. By these metrics, it was in a very strong position to dominate the PC revolution.

loc3336-39: Research shows that IBM executives were aware of the "disruptive technology" aspect of the PC. They discussed the issues involved, recognizing the potential for a bimodal outcome for mainframe customers, their bread-and-butter clientele. Some would be lost permanently to the PC while others would be interested in mainframe upgrades to support new requirements.

log3340-41: But when it came to implementation, IBM appeared to fall into the "active inertia" trap. Rather than pivot decisively to the new approach, they allowed their much more familiar historical behavior to overinfluence their future actions.

Late 80s, a senior disk engineer gets a talk scheduled at annual, world-wide, internal communication group conference, supposedly on 3174 performance ... but opens the talk with statement that the communication group was going to be responsible for the demise of the disk division. The issue was that the communication group had strangle hold on datacenters with corporate strategic ownership of everything crossing the datacenter walls, and were fiercely fighting off distributed computing and client/server (trying to preserve their dumb terminal paradigm and install base). The disk division was starting to see data fleeing the datacenter to more distributed computing friendly platforms with drop in disk sales. The disk division had come up with a number of solutions to reverse the process, but were constantly being vetoed by the communication group.

Trivia: The science center had developed a lot of performance technologies, monitoring, simulation, analytical modeling, workload&system profiling (evolves into capacity planning), etc. One of the APL-based analytical models was enhanced and offered on the world-wide online sales&marketing support system HONE as "Performance Predictor", sales support could enter workload&system profiles and ask change "what-if" questions. I had done dynamic adaptive resource management as undergraudate in the 60s, which was picked up and shipped in CP67. In the morph from CP67 to VM370 there was a lot of simplification and many features were dropped. During the FS period I continued to work on 360/370 stuff (even periodically ridiculing FS activity). VM370 customers were lobbying for re-introduction of my dynamic adaptive resource management and the with the implosion of FS and the mad rush to get stuff back into the 370 product pipelines, contributed to decision to (re-)release my "Resource Manager".

I had also developed some automated benchmarking tools that could vary workload & configuration. As part of the release of the "Resource Manager", 2000 automated benchmarks were run that took 3months elapsed time. The first 1000 benchmarks were selected to methodically cover possible workload and configuration profiles. The last 1000 benchmarks workload&configuration profiles was selected by a modified version of the "Performance Predictor" which was fed results of all previous benchmarks. It would select workload/configuration, predict the results and then compare the benchmark results with the predicted. It would also search for optimal workload+configurations doing "hill climbing" searching for maximums and also attempting to differentiate from "local" maximums verses "real" maximums.?This assumes that there might be an arbitrary number of optimal maximum solutions somewhat analogy to multi-modal (not just bimodal) distribution.