🚀 go-pugleaf

pc532@bungi.com writes:
 > Soon, I hope, the folks working with Mach on the PC532
 > will be able to make their work available to those without
 > UNIX source licences.

The micro kernel is free of nasty licences, the unix server isn't.

There are a number of possibilities for a licence free server.
As I understand it, these are the CMU multi server, the GNU Hurd
and the BSD Net2 based server. The last one has the most hope
of being available early and had been done for the 386.

It wouldn't be too big a job to do the PC532 port, but I don't
know of anyone actually doing it.


 > When that happy day comes, I'd like
 > to use the SCSI bus as an IPC channel between my PC532 and
 > my NeXT.  What do people think of this idea?

Well, we all want to use SCSI as an IPC bus. I've got this
Sun 3/50 I want to talk to.

 > I already plan
 > to write the relevant drivers,

We are using a layered approach to the scsi drivers. There is a
machine dependent part, a machine and scsi device independent part and
a SCSI device (disk, tape, cpu, communications) dependent parts.  The
support for SCSI 2 cpu and communication devices is being added by CMU
people (Alessandro Forin).

Write the machine dependent parts and everything else should work for
free, in theory. In practice, I can't get tapes to work yet.

Our machine dependent SCSI code still needs work. The HUT guys did the
original work and I am trying to clean up/debug it with as much time
as I can spare.  Currently we don't support disconnect/reselect or
target mode. The former is pretty essential from a performance point
of view and the latter is essential for scsi2 cpu and communication
devices.

 > but what schemes are there for
 > sharing compute power between dissimilar processors?

The most promising is NORMA. I forget what it actually stands for, but
basically (from memory) it allows memory objects to be shared over a
network as if it were on the same machine. I think it even fixes up
byte order problems etc. You can, apparantly, write a process to take
advantage of multi-cpu and it works the same whether the CPU's are
tightly coupled or very loosly (NORMA) coupled.

This sounds very nifty. The only published performance figures were
over ethernet. They got linear speed up for some sorts of problems
up to about 4 processors. With SCSI communications, the tradeoffs
would be different of course and maybe NORMA would be useful
over a wider class of problems.

(Disclaimer: I have zero experience with NORMA, I just read a paper.)

Ian