Path: utzoo!mnetor!uunet!lll-winken!lll-lcc!pyramid!prls!gert
From: gert@prls.UUCP (Gert Slavenburg)
Newsgroups: comp.arch
Subject: Re: SPARC and multiprocessing
Message-ID: <11058@prls.UUCP>
Date: 29 Apr 88 16:32:50 GMT
References: <1521@pt.cs.cmu.edu> <28200135@urbsdc> <4921@bloom-beacon.MIT.EDU> <1671@alliant.Alliant.COM>
Reply-To: gert@prls.UUCP (Gert Slavenburg)
Organization: Philips Research Labs, Sunnyvale, California
Lines: 31
Keywords: virtual address cache consistency
Summary: virtual cache consistency solved

> As far as I know, no one has solved the virtual cache coherency
> problem yet...

Coherency can be maintained in a virtual address cache in exactly the same
way as it is maintained in a multiprocessor : pick you favorite 'bus watch'
style multiprocessor consistency protocol. Now use this protocol at the
back (bus) end of a virtual address cache, where requests go out in terms
of physical addresses, applying it TO YOURSELF AS WELL AS TO OTHERS. For
example, if you have an ownership like protocol, aquiring ownership
over a main memory slot (the memory unit that goes into a cache line)
requires 'negotiation' with other caches that hold a copy of this slot,
including yourself if you hold such copies under different virtual address
names.

In order to implement this, the virtually addressed cache needs to be able
to 'observe' (and/or 'interact with' in some ownership schemes) bus 
transactions that occur on those physical addresses of which this cache
holds copies. Again, this can be done in a variety of ways, either involving
a double set of tags or some form of reverse address translation (this may
be a one to many mapping).

The architecture of a multiprocessor with virtual address caches that applies
the above ideas in some unconventional ways was presented at the 13th
Symposium on Computer Architecture in Tokyo (1986) (`Software-controlled
caches in the VMP Multiprocessor', by D.R. Cheriton, G.A. Slavenburg and
P.D. Boyle). At this year's 15th symposium on Computer Architecture in Hawaii,
measurement results of the system, as in operation at Stanford University,
will be presented.

  just thought you might like to know that this problem has been solved,

  Gerrit A. Slavenburg