RE: Rethinking OS Directions for the Next Millennium.

Joe Barrera (
Thu, 16 Jul 1998 17:27:01 -0700

> off my radar until a year after the fact? In fact, why have there been
> zero references to HotOS on FoRK since its inception in December 1995?
> Speaking of which, joebar, I'm looking forward to reading

what huh whereami did someone say my name? (* rubs eyes *)

I'm sorry, but I think of HotOS as little brother to SOSP. I hope you're
aware of SOSP. And I'm sure I've mentioned SOSP on FoRK.

> Rich's main point -- seamless distribution and transparency of
> where objects actually are -- is diametrically contrarian to Waldo et
> al's "A Note on Distributed Computing" position that transparent
> distribution simply doesn't work.

This is a VERY old debate in the OS field. Of course, we didn't always use
the term "object". Rig/Accent/Mach used "port". But ports + RPC were the
basis of network transparent objects. I forget what terminology V used.
Another angle on the debate is network file systems. We take the feasibility
and desirability of network file systems for granted, but this was not
always the case. My non-thesis (Mach kernel support for distributed memory
multiprocessors) could be considered a study of when network transparency is
NOT a good idea.

It's stupid to say that transparent distribution DOES or DOES NOT work. If
you design your abstractions (objects, doo-dads, whatever) with network
transparency in mind, then it works. Is it still transparent if you have to
think about it? I'd say yes. Really, you will be thinking of good design
principles, and these principles will come to include those that make an
object easily and efficiently transparently distributable.

I could go on and on, but the real point is, you're in the right place now,
and you can talk to Rich in person as much as you want, so stop bitching.
You should also talk to the rest of the group. If you can actually get
Rashid to stay on a serious subject, talk to him about the tensions between
transparency and efficiency and error cases and so forth. Now of course
Birman (e.g.) thinks Rick swept most of the hard problems under the rug, and
maybe that's fair. Rick/We/Mach definitely took the "Worse is Better"
approach towards making network transparency happen, instead of worrying
about all the cases where you get screwed. You can also view NFS in the same
way. NFS totally sucks, but it made network file systems real, and it got
everyone interested in fixing it, instead of just saying that network file
systems are intrinsically impossible.

Talk to Rich about Birman's views on whether Mach no-more-senders
notifications are implementable or not.

- Joe

Joseph S. Barrera III
The opinions expressed in this message are my own personal views.