Memory size equivalences

[khs@STONY-BROOK.SCRC.Symbolics.COM (Kenneth H. Sinclair)]

    Date: Tue, 11 Dec 1990 20:37 EST
    From: jcma@reagan.ai.mit.edu (John C. Mallery)

	Date: Tue, 11 Dec 90 13:30:07 -0800
	From: jbarnett@nrtc.northrop.com (Jeff Barnett)

	We are thinking of replacing several 36xx machines with the new 1200's.
	The 3600 machines each have either 4 or 6 mega-word memories.  From your
	experiences with similar conversions, can anyone estimate the amount of
	memory necessary for the 1200's to get similar performance.  By similar
	performance, I mean how much memory do I need so that I page thrash about
	the same fraction of the time that I do now.  I understand that this
	question is not particularly well-formed but I would appreciate any
	heuristic advise that is available.

Assuming you're running roughly the same software on each machine
(Genera 8.0.x and comparable versions of your application), there
shouldn't be much difference in memory requirements for 3600s versus
XL1200s.  Replace your 4M 3600s with 4M XL1200s, and your 6M 3600s with
8M XL1200s, and you'll get a very significant performance improvement.

Some of the factors involved include:

 - I-machines use some main memory (on the order of a hundred kilowords)
   for the IFEP operating system; the 3600 FEP had dedicated RAM for this
 - I-machines use slightly more main memory to describe their potentially
   much larger virtual memory
 - I-machines have better code density and so tend to use main memory
   more efficiently
 - Embedded I-machines (MacIvories, UX1200s) have a somewhat larger working
   set due to their larger user interface support (integration with MacOS
   or X).  This doesn't affect XL1200s.
 - Color systems generally need more memory than monochrome, to store
   deep bitmaps being pushed around by the user interface

    From my experience, an XL1200 with 4 megawords of memory runs the same natural
    language application 6 times faster with 10% of the page faults (if you belief
    the time macro) as a 3640 with the same amount of memory.  This application is
    memory-intensive because it builds a knowledge base from incoming sentences
    and sentence may refer to prior knowledge.  Some of this performance doubtless
    comes from using areas localize memory.  I suspect somebody fixed up paging
    performance for the I-machine.  Perhaps someone on this list can tell us about
    it. (Note that paging performance is currently one of the major advantages of
    the lisp machine over unix lisps).

The I-machine virtual memory architecture is indeed better than the
L-machine's, which takes more "soft" page faults and generally incurs
more overhead to manage a given set of pages.  Also, in Genera 8.0 the
virtual memory system for both machines is substantially improved over
Genera 7.2, particularly for systems with larger physical or virtual
memories; the major difference is the implementation of a cheaper and
better page replacement policy.