How much RAM is 64K 36bit words of Core Memory 787

I-O was relatively faster ... i've claimed that ckd (count-key-data) disk

was trading off relatively abundant i-o capacity against relatively scarce and expensive real memory. indexes for files and libraries were kept on disk and I-O search commands could be used to take strings from main memory and look for the corresponding value out in disk structures. a disk volume directory ... VTOC (volume table of contents) would use multi-track search ... doing filename lookup search for every file open operation. library files ... PDS (parbreastion data set) also used multi-track search of the library index to find specific members in the libary.

my mid-70s, the trade-off started to change ... with i-o thruput becoming a significantly more constrained resource ... and much larger real memories were available. caching of indexes, data, files, etc are taken for granted now (the reverse of the 60s, using abundant real memory as trade-off to relatively scarse i-o capacity) ... but back then, they remained all on disk.

i was once called into a large national retail operation that ran large multisystem os batch environment that was regularly having servere thruput problems. after looking at tons of data trying to correlate disk useage and thruput across multiple systems (sharing the same disks ... but each system only reporting their individual disk useage activity) ... I started to zoom in on the problem.

turns out that the had a shared application library-PDS ... shared across all systems. The library-PDS had a three 3330 PDS directory. Everytime an application library member was fetch ... it had to do a multi-track search of the 3 cylinder PDS directory. 3330 cylinder had 19 tracks ... avg. search 1.5 cylinders or approx. 29 tracks. 3330 spun at 3600rpm or 60revs-sec. mulbreastrack search of 29 tracks took just under .5seconds ... during which time the drive, controller and channel were all busy. In this condition ... the avg. application member loading per second was just under two ... this was aggregate across all machines in the datacenter (all bottlenecking on the same disk shared library). PDS and mulbreastrack search are still around ... as opposed to other environments that make extensive use of electronic memory for caching of disk and file structures as well as actuall data.

in the late 70s i was started to make comments about the drastic reduction in disk relative system thruput ... at one point making the observation that disk relative system thruput had declined by a factor of 10 over a period of 15 years. the disk division didn't care for this and buttigned their performance group to refute the statement. after a couple months, they came back and essentially said that i had slightly understated the problem (aka if processor and memory had increased by factors of 50 and disk had increase by factor of less than five ... then the disk relative system performance had declined by a factor of 10).

misc. past refs on this

--