Strobe equivalents

Gilbert Saint-Flour

for some drift ... in the early 70s, the science center

had done a lot of work on performance monitoring and measurement technologies ... some of it later evolved into capacity planning.

there were sort of three kinds of technology

* monitoring & sampling * simulation & modeling * multiple regression analysis

all had their strengths and weaknesses and there were various situations were one of the technologies could identify an issue when the other two couldn't.

there were both software and hardware monitors. when work was being done for selecting what should go into the ecps microcode buttist, the kernel was instrumented with a software monitor and then the person at the palo alto science center responsible for the apl microcode buttist did a microcode based PSW sampler. old standby posting describing ecps microcode buttist analysis

for other drift ... recent post mentioning the apl microcode buttist

there were two or three different simulation and modeling projects at the science center. there was an event driven system model written in PLI ... used among other things for modeling paging behavior ... and an analytical system model written in apl.

we used a variation on the apl system model in the automated benchmarking for validating the resource manager before release.

basically in excess of two thousand benchmarks were run taking three months elapsed time. initially there were something like 1000 different benchmarks defined that had wide range of configuration, workload, and system parameters. the the modified apl system model was feed all results and allowed to select the set of conditions for the next benchmark. these results were fed back into the apl system model and it repeated the condition selection settings for the next set of benchmarks. this was repeated for another 1000 or so benchmarks

the apl system model was also adapted to the HONE system (world wide vm-based system that supported all field, sales, and marketing, by the mid-70s, salesmen couldn't even place a mainframe order w-o having first run it thru one of the HONE configurators)

where it was called the performance predictor and allowed marketing people to input customer configuration and workload information and ask "what-if" questions (what happens if amount of memory is doubled or the workload changes, etc).

another instruction analysis tool was "REDCAP" which had been developed in POK for doing workload instruction traces ... for studying detailed workload instruction end characteristics as aid in processor design. the science center adapted REDCAP for analyzing application end in virtual memory environments. this was used to analyze the port of apl-360 to cms-apl (and end characteristics of the memory allocation and garbage collection change from small 16k-32k byte real memory workspaces to very large virtual memory workspaces). It was also used by a number of application development groups to study their application in the transition from real storage to virtual memory operation (applications like IMS). It was also released as a product called VS-Repack. VS-Repack would also perform cluster analysis of program operation and attempt semi-automated program reorganization for improved end in virtual memory environment.

some recent references to old vs-repack

A few years ago, I ran into a consultant (european) that was doing work using a descendant of the performance predictor (with nearly 20 years of enhancements). During the corporate troubles in 1992, the vendor had acquired the rights to the software and had run it through an APL-to-C language converter ... and then subsequently made additional enhancements.

He was doing some consulting at a large datacenter that had an enormous application that ran across a large number of mainframes ($$ value in the 8-9 digit range). Even a few percentage performance improvements in the application translated into large number of hardware dollars. The application had been studied extensively using standard mainframe monitoring tools and heavily optimized. The consultant, using his enhanced performance modeling tool had identified additional areas that resulted in another ten percent optimization savings.

Remembering the science center experience from the early 70s (nearly 35 years earlier), i wondered if multiple regression analysis could identify opportunities; in fact it turned up something accounting for over 20% of total usage. The issue has been that things like instruction sampling and event modeling tend to turn up things at the micro level ... while multiple regression analysis frequently can highlight more macro level issues. The identified feature was a complex, spaghetti combination of low-level stuff ... which turned out could be optimized (at the macro level) and resulted in 14% total system savings.