Mach

[Guy Sotomayor]

On Sun, 2020-01-05 at 15:21 -0800, Chris Hanson via cctalk wrote:
> On Jan 5, 2020, at 2:30 PM, Guy Sotomayor via cctalk <
> cctalk at classiccmp.org> wrote:
> > 
> > It did seem for a while that a lot of things were based on Mach,
> > but
> > > 
> > > very few seemed to make it to market. NeXTstep and OSF/1, the
> > > only
> > > version of which to ship AFAIK was DEC OSF/1 AXP, later Digital
> > > UNIX,
> > > later Tru64.
> > 
> > Yes, a lot of things were based on Mach. One OS that you're
> > forgetting
> > is OS X. That is based upon Mach 2.5.
> 
> Nope, Mac OS X 10.0 was significantly upgraded and based on Mach 4
> and BSD 4.4 content (via FreeBSD among other sources). It was NeXT
> that never got beyond Mach 2.5 and BSD 4.2. (I know, distinction
> without a difference, but this is an issue of historicity.)
> 
> I think only some of the changes from Mach 2.5→3→4 made it into Mac
> OS X Server 1.0 (aka Rhapsody) so maybe that’s what you’re
> remembering.

You're probably thinking about the user space.  I was working on the
OS X kernel from 2006-2012.  I can tell you that most of the kernel
that was still Mach related (most actually got removed...about all that
was left was mach message) was 2.5 based with some enhancements.

> 
> > > MkLinux didn't get very far, either, did it?
> > > 
> > 
> > I think that was the original Linux port for PPC.
> 
> It was the original Linux port for NuBus PowerPC Macs at least. It
> was never really intended to “get very far” in the first place, it
> was more of an experimental system that a few people at Apple threw
> together and managed to allow the release of to the public.
> 
> MkLinux was interesting for two reasons: It documented the NuBus
> PowerMac hardware such that others could port their OSes to it, and
> it enabled some direct performance comparisons of things like running
> the kernel in a Mach task versus running it colocated with the
> microkernel (and thus turning all of its IPCs into function calls).
> Turns out running the kernel as an independent Mach task cost 10-15%
> overhead, which was significant on a system with a clock under
> 100MHz. Keep in mind too that this was in the early Linux 2.x days
> where Linux “threads” were implemented via fork()…

At IBM we spent a *significant* amount of time optimizing the
microkernel performance.  I recall that on a 90MHz 601 PPC, we got
round-trip RPC below 1 micro-second.

I personally spent a significant amount of time optimizing the
Pentium kernel entry/exit code and optimizing the CPU specific
porition of Mach RPC (it actually took advantage of the x86
segmentation hardware).

> 
> I don’t recall if anyone ever did any “multi-server” experiments with
> it like were done at CMU, where the monolithic kernel were broken up
> into multiple cooperating tasks by responsibility. It would have been
> interesting to see whether the overhead stayed relatively constant,
> grew, or shrank, and how division of responsibility affected that.

The IBM microkernel project was *very* multi-server.  There was a
version of AIX and OS/2 that ran on top of the IBM microkernel (which
was a heavily modified version of Mach 3.0) were there were quite a few
OS neutral servers (including most device drivers) that were all in
their own server tasks.

-- 
TTFN - Guy