>>>	Adaptec AHA-1540CF (w/o floppy), AHA-1542CF (w/ floppy). 
>>>               bus-mastering scsi-2 host adapters

>>Uh, these are good cards if you want VERY SLOW performance.  I would
>>only recommend using one of these if you already own one and you CAN'T
>>afford to buy something better.  Not to mention, until it gets fixed,

>I wonder about this; someone told me they upgraded to EISA & didn't see
>much improvement.  With platter speeds of 3-5MB/s, I really wonder if
>you need the extra b/w of a wider bus.
>I don't have specs on hard re: the thruput of ISA but I tend to think
>it's > 3-5MB/s.

Well, I haven't compared EISA to ISA, head-to-head.  But I have seen
in otherwise similar machines, a P90 with an ISA 1542 lose _badly_ to
a P66 with an NCR PCI controller.

And, I do get adequate performance from my EISA controller with a 486
CPU.  Adequate meaning: as good as I would expect from a 486 with an
EISA bus.  I will send my disk benchmarks to you if you're interested.
I have never had the opportunity, unfortunately, to test one or
multiple truly fast drives on my EISA controller, since I simply have
more important things to spend my money on than a 4GB Quantum Atlas.
Buying drives on the trailing edge of technology, where they are the
lowest dollar/megabyte has determined that I'll always have
"satisfactory" drives, but nothing stellar.

True ISA bandwidth maxes out at 8MB/s.  That's 2 8MHz clock cycles per
16-bit word transfer.  (I say "true" ISA, because some boards can run
the bus faster -- maybe ~10MHz -- at the risk of intermitant failures,
or out-right failures of slower cards.)

EISA, in burst-mode (ISA has no burst mode), is 33MB/s: 1 8.25MHz
clock cycle per 32-bit dword transfer.

Just for comparison purposes, PCI in burst-mode is 132MB/s: 1 33MHz
clock cycle per 32-bit dword transfer (I don't know of any true 64-bit
PCI devices, yet).  (This also is assuming a 33MHz PCI clock; PCI can
be clocked anywhere from 25MHz to 33MHz -- a good thing to remember
when picking processor speeds.  In a heavy I/O based machine, a 120MHz
CPU (30MHz PCI bus) could potentially run slower than a 100MHz CPU
(33MHz PCI bus) machine, for example.)

This shows another thing to consider: EISA and PCI have head-room
left, where ISA is saturated, with a saturated SCSI bus.  Just because
your ISA card should theoretically be able to do a full 8MB/s, doesn't
mean you'll really be able to get all of that bandwidth in real use.
You still have to share the bus with the CPU, and possibly other
devices.

>Are there significant latency issues perhaps?  That might convince me.

Other than ISA is loaded with latency? :-) I do think that it's
somewhat significant that you can easily run out of head-room with
ISA.  Unless you halt the processor, or keep it in a tight loop where
it can run constantly out of the cache without a miss, you're almost
certainly going to have the CPU steal *some* cycles away from the SCSI
controller.

There's also "soft" latency to consider, as in the fact that a faster
CPU (Pentium) can feed data through a program, through the filesystem
code, and down to the hardware, at the same time as dealing with
potential context switches quicker, and with less latency than a
slower CPU (486).  So, if the 486 can't keep enough outstanding
transactions in the queue to keep the drive(s) from ever waiting for a
transaction, you will absorb some extra latency there.

I think this is especially worth noting because it seems to me, in
purely un-scientific observations, that the critical line between
keeping a fast SCSI device (or multiple devices) always busy, and
maybe having it sometimes waiting between transactions, kind of falls
in the high-end 486 to mid-line Pentium range.  I'm not talking about
an idle machine running a disk benchmark, but a machine doing genuine
mixed-bag real-world type work.

I think this might be one of the reasons some people express
disappointment in EISA performance versus ISA: their CPU just can't
give them a whole lot more.  Of course, I could be totally off the
mark on this.  But it's definite food for thought.

-----------------------------------------------------------------------------
  Michael L. VanLoon                                 michaelv@HeadCandy.com
        --<  Free your mind and your machine -- NetBSD free un*x  >--
    NetBSD working ports: 386+PC, Mac 68k, Amiga, Atari 68k, HP300, Sun3,
        Sun4/4c/4m, DEC MIPS, DEC Alpha, PC532, VAX, MVME68k, arm32...
    NetBSD ports in progress: PICA, others...

   Roll your own Internet access -- Seattle People's Internet cooperative.
                  If you're in the Seattle area, ask me how.
-----------------------------------------------------------------------------