What kinds of things would you want in the GNU OS?

[Steve Lamont]

In article <424 at ladcgw.ladc.bull.com> frank at ladc.bull.com (Frank Mayhar) writes:
>In article <4609 at alvin.mcnc.org> spl at mcnc.org.UUCP (Steve Lamont) writes:
>>As quoted from <422 at ladcgw.ladc.bull.com> by frank at ladc.bull.com (Frank Mayhar):
>>| course).  If you're writing a Real Operating System, why worry about machines
>>| that won't support virtual memory.  Hell, by the time Gnu is ready, non-VM
>>| machines will probably be a thing of the past anyway.  Shared libraries
>>Like the Cray Y-MP or Cray 3?  [...]
> [deletion]
>Actually, IMHO, a Cray is a very poor choice of hardware to run a general-
>purpose OS on.  As I see it, you would want the Cray to run a minimal kernel

Why?  Just because it is very fast?  Sounds good to me!  You don't give
any arguments to support your contention, so that gives me the privilege
of setting up a straw-person argument for the express purpose of knocking
it over (I love arguing with myself... I usually get to win! :-) ).

A common criticism heard (and one that I made until I began working on
interactive Crays some three years ago) is that you don't want to be running
vi or emacs on a Cray.  The character level interrupts will kill the
poor thing.  This is simply not true.  They simply do not happen
frequently enough (from the machine's perspective) to worry about.  If
scheduling is done correctly (that is, you schedule memory instead of
cpus) then the editor tasks are frequently tucked away in some unused crack
in memory that would be idle anyhow.  If priorities are adjusted
correctly, then the task will stay in memory for quite a while before it
gets bumped out.  Meanwhile, your favorite bomb code is happily
simulating the end of the universe unperturbed.

>(primarily consisting of a scheduler) that accepts compute-bound jobs from
>a client machine running some general OS, and executes them.  Basically,
>a high-speed backend that runs those cpu bound jobs that are so hard on
>general-purpose OS's, and that's *all* it does.  The general-purpose OS
>(running on a virtual memory machine :-) performs task scheduling, memory
>allocation, job queueing, etc, and hands the jobs to the Cray.

This is a batch processor model which will work in a number of cases,
for FORTRAN finite element weenies simulating the melt down of the local
nuke plant, or any other long running job, f'rinstance.  However, there
are also a number of situations where the user wishes to interact with
the computation.  In a batch mode, you really can't do that.  I suppose
that a really smart compiler and operating system could divide up the
task into interactive and computationally intensive subtasks, running
the compute-bound task on the supercomputer but that's a non-trivial
thing to do and, to my knowledge, a continuing topic of research.
Perhaps you could migrate tasks from one CPU to another as their
resource utilization profile changed.  I don't know.  All I do know is
that there are a number of fairly successful interactive tools in common
usage on Crays and TCP/IP coupled Cray and workstation systems.
Interactivity is a must!

Consider, if you will, the task of debugging.  In a batch mode, the
cycle is:

  Submit job
  Wait for results
  Edit source file to add PRINTs or fprintf()s.
  Resubmit
  Wait
  Edit...
  Etc, etc, etc.

I know, because that's how I debugged on a batch Cray before I began
working on an interactive Cray.  A tool such as adb or dbx or whatever
your favorite debugger might be tightens the cycle considerably on a
supercomputer in just the same manner as it does on a workstation or a
minicomputer.

>Disclaimer:  I have no direct knowledge of how Crays work.  When I made my
>comment, it was in reference to general-purpose operating systems.  And, no,
>I don't think that a Cray has any business running a Real Operating System.
>(Half :-)

Since you half smiley that one, I'll give you a half serious rebuttal
:-).  A Cray, or, again, any other supercomputer, should not be thought
of as simply a compute server.  It is, indeed, a general purpose system
which, when the operating system is properly tuned, provides
considerable interactivity.  Crays are Real (actually Double Precision)
computers which need Real Operating Systems (do I have to include the
silly LUSENET (TM) here? :-) )

If you still have doubts about this, I think I know where I can find you
a Model 029 card punch, if your really need one for your next Cray run.
( Full :-) with oak leaf clusters ).
-- 
							spl
Steve Lamont, sciViGuy			EMail:	spl at ncsc.org
North Carolina Supercomputing Center	Phone: (919) 248-1120
Box 12732/RTP, NC 27709