SCSI adapters and the ISA bus

[mikebw at bilow.bilow.uu.ids.net (Mike Bilow)]

Chuck Young wrote in a message to Mike Bilow:

 CY> With all this talk of scanners and SCSI cards, I was
 CY> wondering... 

 CY> Will an Adaptec 1542 16-bit ISA SCSI card provide decent
 CY> performance over an on-board/PCI EIDE system? I thought the
 CY> ISA bus was pretty slow. I realize the question is vague and
 CY> there are advancements made every day in chipsets.  For
 CY> discussion, lets assume an average P-166 and comparable
 CY> disks running under linux 2.0.36. 

 CY> Generally, is an ISA SCSI adapter any faster than the
 CY> on-board EIDE I/O? 

 CY> Anyone?

This is a complicated question.  First, the bandwidth of the SCSI bus is not
likely to be any faster than the bandwidth of the machine bus.  Standard
synchronous SCSI runs the bus at 5 Mo/s (million operations per second).  Fast
SCSI doubles this to 10 Mo/s, and Ultra SCSI doubles this again to 20 Mo/s. 
Narrow SCSI has an 8-bit data path that transacts one byte per operation; Wide
SCSI has a 16-bit data path that transacts two bytes per operation.  The raw
bus throughput therefore is just a matter of multiplying, and an Ultra Wide bus
has a maximum theoretical bandwidth of 40 MB/s.

The Adaptec AHA-154xCF -- let's assume you are talking about the C series --
can do Fast Narrow, which only goes up to 10 MB/s.  Of course, this is a
theoretical maximum, so acknowledgements and device latency are going to limit
sustained throughput to maybe a tenth of that for a given device.  SCSI is at
its best when used to manage multiple devices in parallel, and this is where
the bus bandwidth over individual device bandwidth provides useful headroom.

While the ISA bus is itself clocked at only 8 MHz, this is usually not a
significant limitation for an AHA-154xCF controller.

EIDE, on the other hand, is an asynchronous protocol which can shove data as
fast as the device can take it, at least in theory, and a PCI EIDE controller
has a theoretical bandwidth accessible to the limit of the PCI bus, or four
bytes per operation at 33 Mo/s, which works out to 132 MB/s.

In practice, device throughput is usually the limiting factor, and the bus
bandwidth is not reached.  SCSI hard drives have more sophisticated and
intelligent electronics on-board, so they usually run with larger caches,
better prefetch and queueing algorithms, higher rotational speeds, and
integrated error handling.  IDE hard drives are usually built to optimize for
cost, so they often are much slower than SCSI hard drives of similar capacity.

However, IDE also has a lot of limitations.  The devices sharing an IDE channel
cannot be operated independently, and it is not possible to use the other while
the first is still pending an operation.  With SCSI, commands can be queued to
one device on the bus and then the bus remains free for other uses until the
first device calls back to notify the controller of status.

IDE controllers, except for the new UltraDMA scheme (which has its own
problems), require the full attention of the main CPU while pushing data back
and forth, while SCSI controllers release the CPU for other tasks.  You will
not see any difference on an operating system which cannot multiplex I/O, such
as Windows 95, but this can be a huge penalty for IDE on Linux.  The AHA-154x
series, which is one of the few ISA busmaster devices ever made, is an even
bigger win relative to IDE on this issue.  This was a much greater concern when
CPU capacity was more limited and the standard server might have been a 486
running at 66 MHz; as raw CPU power increases, it becomes expendable.

The end result is that SCSI drives tend to be faster in terms of sustained data
rate than IDE drives of comparable capacity.  Even if this is not true in the
case of your particular situation, the ability of SCSI to offload tasks from
the main system CPU usually more than compensates.

Finally, however, I have to ask why you would bother with an ISA SCSI
controller in a PCI machine.  You can get pretty respectable PCI SCSI
controllers reasonably well supported under Linux, such as those based on the
NCR/Symbios 53c810 chipset, for about $50.  Whatever disadvantages such SCSI
controllers may have, they are almost certainly going to be faster than an old
ISA SCSI controller.

The real importance of all of these issues ultimately depends upon the uses you
plan for your particular machine.
 
-- Mike


***
Subcription/unsubscription/info requests: send e-mail with subject of
"subscribe", "unsubscribe", or "info" to discuss-request at blu.org