[linux-elitists] ME roadshow

Eugene Leitl eugene.leitl@lrz.uni-muenchen.de
Thu Aug 31 00:44:36 PDT 2000

Paul J Collins writes:

 > Linux already has enough latencies, even in 2.4.  More are not needed.

Another major quibble: Linux is not realtime. Realtime nowadays means
response times *well* under 1 us, whereas on 2.0/i486DX2(66)

task                 max min
sched_getscheduler() 165-150 us
open(file)	     300-145 us
close(file)	     190-145 us
write(1 Byte)	     250-150 us
read(1 Byte)	     195-150 us
access()	     280-150 us
shmget()	     200-150 us

While most of this is architecturally artefacted, at least some of it
is due to bloat. RTLinux is no solution. What one needs is a minimal
(OS-independent) kernel, hand-coded in pure assembly, by the designers
of the CPU, which, presumably, have been designing it with above
nanokernel in mind.

 > Ingo Molnar (and possibly others) has been doing work on this, but I
 > can't get a kernel built with his patch applied to work.  Shame.
RT patch?

 > Certainly.  No operating system kernel is going to be appropriate for
 > all situations.  A nice microkernel fits the bill there.
The point is: a QNX system will certainly perform no worse than a
Linux system. And a QNX-type system will do on above million-processor
(btw, IBM's Blue Gene will have over one million processors), whereas
vanilla Linux won't.
 > However, I'm talking about super-servers derived from pre-existing
 > operating systems, which will be roughly the same size as the
 > equivalent kernel.  Running Linux or BSD on a microkernel buys you
 > nothing.
Monolithic superservers are obsolete at the high end. Even with
Myrinet current clustering sucks, but future CPUs will have message
passing in hardware with many GBps and 10-30 ns instead of current
10-30 us. SMP has pretty much arrived, but in near future you ought to
be able to buy a Beowulf with several 10 CPUs in a single tower,
without having to resort to 19" rackmount monstrosities.

 > I shine my torch around, and I see big signs saying ``Paul's Ignorance
 > Starts Beyond This Line''.  Embedded and distributed system are
 > another thing I'm not qualified to have an opinion about.

Look, I'm dumb as dirt, but this doesn't mean I can't
criticize. People *should* be thinking ahead when designing
stuff. Maybe in the year 2000 we ought to be able to obtain a
distributed message-passing real-time system. And if PC hardware
sucks, you can always use DSP clusters.

More information about the linux-elitists mailing list