30 Jul
2016
30 Jul
'16
7:24 p.m.
On Sat, 30 Jul 2016, Fred Cisin wrote:
From my experience with floppy formats and reasonably fast computers
(i.e. where CPU processing latency doesn't really matter) the best results
are obtained with no interleaving, no sector staggering on head switching,
and single-sector staggering on cylinder incrementing.
I.e. taking the the 1.44MB 3.5" format (18 sectors, 2 sides, 80
cylinders) as an example you start track #0, #1, #2, etc. on both sides
with logical sector #1, #18, #17, etc. Logical sector numbers then
increment modulo 18, plus 1 (as sector IDs start counting from 1), i.e.:
s[n+1] = (s[n] % 18) + 1. This is because with a fast host CPU it takes
negligible time to issue commands to the FDC as results from preceding
commands become available and therefore the FDC will receive a subsequent
sector or track read (or write) command soon enough for data requested not
to have passed under the head in rotary movement yet. And head selection
is instantaneous, made along a read (or write) request.
A seek command requesting the next cylinder however requires the step
motor to physically move the head assembly and it takes enough time to
complete for medium to have already rotated past the next physical sector.
So if that happened to be the logical sector requested with the next read
(or write) command as well, then almost a full rotation would be required
to even start executing the data transfer. So to avoid that missed sector
you need to stagger sectors back by one on cylinder increments, so that
the physical sector lost due to the head movement is actually one you may
need last and the subsequent one is the logically next one.
Empirical data showed that a 1.44MB 3.5" floppy formatted as I described
above took IIRC ~35s to read (or write) it whole whereas one formatted
with the regular DOS utility required noticeably more. Given the figures
I quoted above we have 2+1/18 vs 3 revolutions per two-sided cylinder read
(or written) for the staggered vs unstaggered sector layout, giving ~51s
required for a whole floppy formatted with physical and logical sector IDs
identity mapped. This makes the staggered format only ~1.5 times rather
than twice as good, contrary to what I quoted from my faded memory in my
original statement. Still this was not to be underrated at the time of
regular floppy use.
Maciej
I highly
recommend it if you're still into the floppy business
and don't use Linux. The plain `format' command supplied with DOS gives
you little control really and produces poor performance floppies.
My PCs (original 5150 and IBM 5160 being exceptions) can handle 1:1 , reading
9 sectors per revolution, making no need for interleave.
IBM and Microsoft were not UNAWARE of the concept of interleave.
But, many machines did need it, and most early hard disks, particularly on
5160, called for a a little experimenting based on how YOU used it, to
determine YOUR optimum interleave. (Programs that changed interleave for
performance were sometimes unaware that the optimum one could be different
based on what you were doing with the data)
WRONG interleave could result in needing more revolutions to read a track.