Re: DMA or not DMA...

[dempson@actrix.gen.nz (David Empson)]

<pubpc1@library.ucla.edu> wrote:

[Discussion with Nathan on merits of DMA]
> 
> Since you don't believe me when I said that I noticed no difference,
> I opened up the GS and ran some tests, with DMA off and flipping the
> switch of the Apple HS SCSI card to turn DMA on (powering down each
> time first).  I did this with 13.75MHz/16K cache acceleration and
> without.  Here are the results of the shift-boot times measured from
> the moment the "Welcome to the IIgs" screen popped up to the time
> the Finder appeared.  My copy of System 6.0.1 was installed after the
> volume had been used for a while, so there was fragmentation.
> 
> 
> Time to boot (sec)      with DMA on     with DMA off
> -------------------------------------------------------------------
> @13.75MHz/16K cache             11              12
> @non-Zip speed                  19              21

I just ran the same test on my system (System 6.0.1 with _lots_ of stuff
installed, but I was shift-booting).  You weren't precise about what you
mean by "the time the Finder appeared" - I have timed from the "Welcome
to IIgs" logo to the appearance of the menu bar, as subsequent times
will vary substantially depending on your icon files, partitions, etc.

I have an 8 MHz/16K ZIP in a ROM 3 IIgs with 5MB of RAM, and an Apple HS
SCSI card.  The hard drive is an 850 MB Quantum Trailblazer (about two
years old; 3200 RPM if Chinook SCSI Utilities is making an accurate
measurement, and it has a reasonably large cache).

The numbers are as follows:

Time to boot (sec)      with DMA on     with DMA off         Ratio
-------------------------------------------------------------------
@8MHz/16K cache                 17              25           1:1.47
@non-Zip speed                  26              36           1:1.38

The right side of the "Ratio" column is the "DMA off" time divided by
the "DMA on" time, thus booting takes 47% longer with DMA disabled and
ZIP enabled, 38% longer with both disabled.

As you can see, there is a substantial difference in speed with DMA
disabled.  I would definitely not want to have to turn off DMA for
normal operations.

I don't understand how your system has such small values for all the
counts, unless you have close to a minimum install of the system
software.  The small difference between the values is probably due to
the hard drive being relatively slow.

However, boot times are not the clearest way of demonstrating DMA
benefits, as the boot sequence has considerable periods of CPU
operations.

DMA is most effective when transferring a large amount of data, so that
the operating system is asking for data in large chunks (up to 64K).

Here are some tests I did, duplicating a 1930KB file in Finder on a
ProDOS volume.  There would probably be some fragmentation involved, so
this doesn't give a pure data transfer figure.

Time to copy (sec)      with DMA on     with DMA off         Ratio
-------------------------------------------------------------------
@8MHz/16K cache                 15              40           1:2.67
@non-Zip speed                  26              67           1:2.58

There are noticeable delays while writing, between each 64K chunk - this
is the main observable difference between the ZIP and non-ZIP tests.

It is interesting to note that DMA makes _more_ of a difference in an
accelerated machine than in an unaccelerated machine.  I can't think of
an obvious explanation for this at the moment.

As far as speed throughput is concerned: the best combination above
transferred 257 KB per second, but that includes CPU overhead.

Doing some tests with Chinook SCSI Utilities: when reading 1000
sequential blocks, 24 at a time, my system can transfer 629KB per second
(DMA enabled, ZIP enabled).  With DMA disabled and ZIP enabled, it only
transfers 43KB per second (a ratio of 14.6:1).  Disabling the ZIP makes
no appreciable difference (under 0.5KB/s), so these figures are
reasonably good upper bounds on the I/O transfer rate with and without
DMA.  I think the DMA limit would be higher if the software transfers
more than 12KB at a time.

-- 
David Empson
dempson@actrix.gen.nz
Snail mail: P.O. Box 27-103, Wellington, New Zealand