Path: utzoo!attcan!uunet!ingr!garth!smryan
From: smryan@garth.UUCP (s m ryan)
Newsgroups: comp.arch
Subject: Re: unconventional architectures
Message-ID: <2844@garth.UUCP>
Date: 13 May 89 02:47:33 GMT
References: <112@centaure.UUCP> <422@unicads.UUCP> <11579@cgl.ucsf.EDU> <89May6.165030edt.10782@ephemeral.ai.toronto.edu> <1989May6.234007.23517@utzoo.uucp> <89May7.001514edt.39763@neat.ai.toronto.edu> <40284@think.UUCP> <13688@watdragon.waterloo.edu>
Reply-To: smryan@garth.UUCP (s m ryan)
Organization: intergraph/apd in sunny california
Lines: 49

>Hear ye!  Segments are evil.  If my usable address space is N, it is a priori 
>*wrong* to assume that I won't mind dealing with segments of any size < N.

Segments are a way to manage the disc image when the virtual image is sparse.

Consider a procedure which adds properties P to some graph and second
procedure which uses property P to add properties Q to the graph. If the storage
used to contain P can be released in the middle of the second procedure, you
have two strategies in a Unix style non-sparse address space:
   -  free P while allocating Q, thus requiring an intermingling of frees and
      allocates which can expand the working set and increase overhead in
      allocation procedures.
   -  use a mark/release which ties up space for P long after it is no longer
      necessary.

In a sparse address space, P and Q can be allocated in two separate regions
simply by having a current base and incrementing it for each allocate. A region
is freed by simply removing it from the virtual address space.

In a nonsegmented system with sparse addressing, you have a problem of mapping
the virtual address to the address of the disc image. For small machines, it
is possible to have the entire virtual space mapped onto the disc, but if it
is sparse, a better technique would exploit that sparseness and reduce the size
of the disc image.

How? Try looking at VSOS for the Cyber 20x. It implemented this, but the map
was limited size so that many programs aborted because the map was too
disorganised and ran out of space.

In a segmented system, the programmer can be required to start each region in
a segment, and each segment has a separate disc image. If the region is
itself fairly compact, it is sufficient to have the entire segment from
base to highest used address mapped to the disc, thus simplifying the virtual
memory <--> disc image mapping. This is used in NOS/VE which I am told is
derived from Multics.

An additional advantage of segments is the segment=disc file philosophy. So
that instead of copying a disc file to virtual memory which may be paged in
and out of another disc file, as on Unix, the paging is directly between
virtual memory and the original disc.

Just as a parallel machine can be used in fundamentally different ways than
a sequential machine, a sparse virtual address space can be used in
fundamentally different ways than a PDP-11's physical address space.
-- 
17. The Aesir filled the otter's skin             Steven Ryan: ingr!garth!smryan
and heaped the gold above its chin.                2400 Geng Road, Palo Alto, CA
But Hreithmar saw a whisker bare    holy corn oil, that is oil from the corns of
and bid them cover every hair.    of holy men, or as it is known in your country