[cctalk] Old UNIX Systems

On 06/04/2013 01:52 AM, Paul Birkel wrote: Way Back When, I was detailed along with a co-worker to figure out what to do when a government customer canceled a contract in mid-stride, but was left with many units of bulk core storage, each unit being 4 million words (40 million characters). Like most bulk core of the time, it was rather slow first-access, but ran at full bandwidth once a block transfer got started. We figured that this incredible wealth of random-access memory (you could use it to stare data via special read and write instructions, but otherwise could not execute or use normal memory instructions to access it.), much like a RAM Disk would allow us to use new and exotic sort/merge methods. Curiously, the classical ones from Knuth (I recall the fold-out-pages from the book showing timings of various sorts) involving tapes also applied to this stuff--we couldn't come up with anything that was better (remember the large start-up latency). Fast-forward about a year--we were on a different system whose curious (and idiotic) architecture was vector. The machine had scalar operations, but they essentially were executed as vector operations of length 1. So the darned thing would churn out vector results a super-word (512 bits or 8 words) per cycle--ONCE IT GOT THE OPERATION STARTED. In other words, all data accesses had the same latency, be they a simple 1-word scalar or a 65Kword vector. There, for a lot of things, Knuth went out the window. Selection sorts for about n up to 1000 were fastest using vector instructions. Quicksort was horrible for most practical values of n. --Chuck