Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!bionet!agate!saturn!ucbvax!laidbak!mdb@ucscc.UCSC.EDU
From: ucbvax!laidbak!mdb@ucscc.UCSC.EDU (Mark Brukhartz)
Newsgroups: comp.os.research
Subject: Re: Inverted page tables
Message-ID: <6588@saturn.ucsc.edu>
Date: 3 Mar 89 22:15:43 GMT
Sender: usenet@saturn.ucsc.edu
Organization: Lachman Associates, Inc., Naperville, IL
Lines: 43
Approved: comp-os-research@jupiter.ucsc.edu


> The MACH papers discuss some of the difficulties that they had porting
> that operating system to the PC/RT because of the awkward
> implemetation of shared memory on such an MMU.  I do not recall right
> now how they solved the problem.  Perhaps someone at CMU could
> elaborate.
> 
> Too bad the evil empire (IBM) slapped yet another copyright on this
> idea.  It would, however, be interesting to hear how BigBlue does
> System V shared memory on the ROMP MMU.
> 
> 	...ken seefried iii
> 	   ken@gatech.edu

The ETA-10 supercomputer features an associative page table. Its
memory management is based upon the CDC Cyber 205, which has been
around for a while now. The software difficulties of associative
page tables appear to be similar to those of inverted page tables.

Each associative page descriptor contains a virtual address and a
physical address. The hardware sequentially searches the page table
by virtual address, then moves the matching entry to the beginning
of the table. The first sixteen page descriptors are cached in a
fast associative store.

The ETA-10 System V Release 3.0 memory management implementation
builds an associative page table at the beginning of a time slice
from the System V region structures. At the end of the time slice,
it updates the region structure reference and modify bits from the
associative table. This is done in associative table order, since
the table is reordered during execution; the corresponding region
entries can be found by a nearly direct search.

Multiple regions could perhaps be mapped simultaneously with the
machine's lock-and-key mechanism. That might improve context switch
time at the expense of page fault handling time.

Are there other UNIX implementations on machines with unusual page
table organizations?

			Mark Brukhartz
			Lachman Associates, Inc.
			..!{amdahl, masscomp, nucsrl, sun}!laidbak!mdb