🚀 go-pugleaf

In article <1487@wpi.wpi.edu>, jhallen@wpi.wpi.edu (Joseph H Allen) writes:
> In article <1257@houxs.ATT.COM> you write:
> >
> 
> Somewhere I heard that some of the high-end IBM mainframes use drum memory
> for swapping space.  Is this true?  And what is the advantage?  I can imagen
> that drum memory might be quite fast since there is a head for every track and
> no moving parts except for the drum...

	This information may be almost a decade out-of-date: I was sort of
	a "370 repairman" in the 70's.

	Two types of drum were the 2305 Model I and the 2305 Model II.
	They looked about the same, but the former had twice the datarate,
	increased price and *half* the capacity.  I don't remember why it
	had only half the capacity; there is some logical explanation.
	Guess: perhaps it was because of 2-heads-per-trk to get 180-degree
	maximum latency and therefore half as many tracks.

	IBM also offered mixed moving/fixed-head disks but let's talk about
	the 2305.  The Model II was relatively common, but because of the
	high price-per-megabyte, very few Model I's were ever sold.

	The paging software could sense the 2305 angular position and used
	it to choose its allocation whenever data was swapped out.  Thus the
	2305 could be kept quite busy and was always attached, by itself, to
	the highest-priority channel.  Many "370 repairman", after a tough
	installation, regarded 2305 activity as the ultimate test.

	With 3.0 Megabytes/sec, the 2305-I exceeded twice the speed of any
	other peripheral and used 16-bit data to the host channel (3 fat
	cables instead of 2.)  IBM's attached floating-point unit was the
	next peripheral to use that interface.

	Since the 2305-I was incredibly expensive, STC competed with it
	successfully with their "6305 (?)" made from dynamic RAM.
	Intel tried also but was less successful, partly because their
	semiconductor "search latency time" wasn't much better than that
	of the 2305-I! (here "search latency" is the firmware overhead
	on their microprocessor controller.)

	Data-intensive applications often read their data in large chunks.
	A 64k chunk takes 22 milliseconds to read even at 3 MBS datarate.
	For this reason, datarate becomes as important as seek latency
	for these high-powered users and this tends to reduce the advantage
	of fixed-head disks.  Bubbles and CCD's were designed into
	"drum replacements" in the late 1970's, but MOS RAM's are now so cheap,
	and conventional disks so fast and cheap, that the intermediate-
	performance technologies no longer seem to attract commercial attention.

- James Allen