Path: utzoo!utgpu!watserv1!watmath!att!rutgers!usc!zaphod.mps.ohio-state.edu!mips!prls!pyramid!cbmvax!cbmehq!babylon!rbabel
From: rbabel@babylon.UUCP (Ralph Babel)
Newsgroups: comp.sys.amiga.hardware
Subject: Re: GVP Trade-in
Message-ID: <03423.AA03423@babylon.UUCP>
Date: 31 Aug 90 22:47:22 GMT
References: <589@oregon.oacis.org> <38CP09P@dri.com> <02048.002057@thiger.UUCP> <552@DIALix.UUCP> <02123.132132@thiger.UUCP>
Lines: 83

To DMA or not to DMA ...

In article <02123.132132@thiger.UUCP> skraw@thiger.UUCP
(Stephan von Krawczynski) writes:

> the problem is: if dma-transfer from controller to
> fast-ram takes place, is the processor able to access this
> fast-ram at the same time?

As I said before: During DMA to the on-board RAM, the GVP
controller does not block the Zorro-II bus.

> for if it is not, your processor will do ABSOLUTELY
> nothing, because it's cut off the bus (with exception of
> real weird programs that have the luck to be inside the
> instruction cache when dma starts and need no data from
> "outside"-world).

Zorro-II bus arbitration is faster than task-switching.

> what does it help if the data bursts in from your harddisk
> and the controller isn't able to put it to your mem
> (somehow).

Let's do some math: If a controller doing programmed I/O can
access the data port on consecutive addresses, then the
fastest way to transfer data is:

 READ: MOVE.L  (Ax),(Ay)+    ;20 cycles

WRITE: MOVEM.L (Ay)+,REGLIST ;12+8n cycles
       MOVEM.L REGLIST,(Ax)  ; 8+8n cycles

Usually you cannot use MOVEM for reads because the MC680x0
does a one-word prefetch (memory to register) which would
kill a data word (you _could_ fix that in hardware). You can
use at most 14 registers for the MOVEM, since you need a
source register and you probably want to keep your stack
pointer.

On a 7-MHz 68000 (e.g. Standard-Amiga), this results in
maximum transfer rates of ...

1.4 MB per second for  READ ( 20 cycles for  4 bytes)
1.6 MB per second for WRITE (244 cycles for 56 bytes)

... NOT COUNTING _ANY_ OVERHEAD!

BTW: The german ALF3-ad claims "up to 2 MB per second".

>> At the Zorro II bus speed of about 3.5 MHz, this would
>
> 3.5 MHz?

Let's call it the "byte frequency" (at 4 cycles per word).

> you have to take the bus-arbitration into consideration.

A clever DMA controller doesn't arbitrate for and release
the bus on each and every 16-bit word.

>> You can also connect more than one drive without "choking"
>> on the bus, especially during overlapping seeks and
>> transfers.
>
> what do you want to say here, i don't understand this.

Fast and short bursts keep the SCSI bus available for other
devices (DISCONNECT).

> i do not measure a controllers performance via filesystem.
> there are tests (i use speedtest, comes with all
> ALF-controllers) that measure the real
> device-/controller-performance ((read-)throughput).

Wasn't it you who insisted on "real-world" performance
measurements. "DiskPerf" (DOS-level) is much closer to that!

Ralph

cbmvax.commodore.com!cbmehq!babylon!rbabel
"If you keep your mind sufficiently open,
people will throw a lot of rubbish into it."