virtual memory 4499
Brian Inglis ... Perhaps it went to the same place where people who are unwilling to accept actual data in studies whose results don't conform to their preconceptions banished it. Of course, cavalierly dismissing...
presumably this refers to "local" LRU as opposed to global-system LRU. what was their basis for "LRU" ... was it at the per instruction reference (updating the ordering of all pages as to the least recently used at a per instruction referenced level) ... or was it at a larger granularity?
one of the things that the several emulators did at the science center
was full instruction traces for exact storage reference sequence (at the instruction level) and then compared LRU against things like (belady's) OPT as well as the various implementations approx. LRU ... and then looked at global system thruput of the different mechanisms.
part of this contributed to the analytical model that grew into the sales-marketing buttist tool available on the hone system
called the performance predictor.
one of the issues was that standard (global or local) LRU degenerated to FIFO under various conditions and much worse than RANDOM and-or OPT. part of this was the slight of hand coding that i did in the global LRU approximated that would degenerate to RANDOM in cases when standard LRU degenerated to FIFO (and overall showed a 5-10 percent improvement over straight LRU).
the other issue in various of these studies wasn't so much how close things got to LRU (or belady's OPT) but comparing overall system thruput using different strategies.
leading up to releasing my resource manager on 11may76
one of the things that some of the paging simulation work done at the science center looked at the trade-offs of different sized pages. in (severely) real...
we did thousands of simulations and benchmarks ... that helped calibrate both the stuff in the performance predictor as well in the resource manager. we then did a varioation on the performance predictor as well as a lot of synthetic benchmarks calibrated against workload profiles from hundreds of systems, along with an automated benchmarking infrastructure. the modified performance predictor got to choose the workload profile, configuration, and system parameters for the next benchmark, it then predicted what should happen, the benchmark was run, and the measured results compared against the predicted. the modified performance predictor updated the most recent results ... and then used all the past benchmarks results plus the most recent to select the next benchmark. leading up to the final release of the resource manager ... 2000 such benchmarks were run taking three months elapsed time.
--
Alt Folklore Computers from Newsgroups