kms@ecsvax.UUCP (Ken Steele) (08/19/89)
There has been much discussion on the GVP controller versus the DMA
hard disk controllers (HardFrame, A2090A). For those who have
missed the controversy a short summary. The GVP controller uses DMA
to read data into a buffer and then uses the 68000 to transfer the
data from the buffer/cache into memory. The other controllers use
DMA to directly squirt the data into memory. A complaint against
the GVP has been that reliance on the intermediate step
through the 68000 will produce degradation of transfer speed
performance (particularly if other tasks are running that use the
68000 very much).
One question at the user level is how much? I show below
comparative speed data for transfer with and without a cpu-intensive
task running at the same time. The task was the Boxes demo from the
1.2 distribution disk. PerfMon shows that the 68000
is very busy when Boxes is running (approximately 0% idleness
according to the display).
Technical details of the comparison: A stock A2000 (1.3 with
Fat Agnus) with a GVP HardCard (40 Mb 8051S MiniScribe). The only
other task running was PopCLI3. The comparison was made using
diskperf from Fish disk 187. When Boxes was run no change was made
in the size of the default window, but it was moved from the center
of the screen.
Results: R/W Speed (kbytes/sec)
Without Boxes With Boxes
read write read write
buffer size
512 27.8 27.3 3.4 3.6
4096 140.1 113.7 22.7 22.4
8192 145.9 130.7 41.6 40.7
32768 153.8 131.6 60.0 56.8
131072 * * 68.5 61.9
524288 147.7 132.1 69.9 64.9
*-missing
What does it mean?
The results showed that a cpu-intensive task like Boxes produced
about a 50% reduction in r/w speed with the GVP controller.
Performance degradation was very dependent on the size of the read.
The smaller the size, the proportionately slower the read/write
speed. The effect was very noticeable to me below an 8k buffer size.
However with larger buffer sizes, a 50% reduction in speed was still
60 kbytes per sec. One can ask the question of how many users have
so refined a taste in hard disk speeds that they would notice the
speed reductions in more typical use (e.g., a word-processor +
communications program + drawing program running together).
Additional questions:
I would be interested in seeing comparative data from people with
DMA controllers. It would also be interesting to see performance
by those controllers when they are faced with competition from
other processes using DMA.
Caveats, Nota benes, Don't blame me's:
I picked Boxes as a worst case test of the controller. Boxes may
not be the best test. Suggest another that would be better and
readily available. Suggest a real test for DMA controllers also.
The point is practical/user level experience of typical or worst
case situations. And to all the psychologists out there, the
psychophysical question of speed difference limens could be fairly
straightforward. Just send me a dozen different setups and I will
be glad to do it. I will return the machines just as soon as I am
sure that the tests are _complete_ ;-)
Ken
--
Ken Steele Dept. of Psychology kms@ecsvax.[bitnet || UUCP]
Mars Hill College kms@ecsvax.uncecs.edu
Mars Hill, NC 28754