Path: utzoo!utgpu!attcan!uunet!husc6!mailrus!cornell!uw-beaver!uw-june!pardo
From: pardo@june.cs.washington.edu (David Keppel)
Newsgroups: comp.arch
Subject: Scientific Virtual Memory (WAS: Cray & Amdahl Virtual memory debate)
Summary: prediction
Keywords: virtual pyisical memory prediction
Message-ID: <5382@june.cs.washington.edu>
Date: 27 Jul 88 19:56:42 GMT
References: <4232@cbmvax.UUCP> <76700035@p.cs.uiuc.edu> <34254@yale-celray.yale.UUCP> <851@l.cc.purdue.edu> <12407@ames.arc.nasa.gov>
Reply-To: pardo@june.cs.washington.edu (David Keppel)
Organization: U of Washington, Computer Science, Seattle
Lines: 23
Disorganization: U of Washington, Computer Science, Seattle

[ Debate: to VM or not VM on scientific computing ]

The issues seem to be that when you have paging, etc., you spend a lot
of your time waiting for page faults to complete.  With async. I/O,
you can "predict" the next region to "fault" and preload it.  A
crystal ball would be very useful here.

Thought: One scheme that appears to help cache hit/miss ratios
(according to a paper by A.J.Smith, in any event) is a predictive
scheme as follows:

  + On a fault, fetch both the thing you need and the next one after.
  + Mark the next one after as "fetched on prediction".
  + First time you touch a thing marked "fetched on prediction", fetch
    the next one after that, and mark it as "fetched on prediction".

Given that array access patterns are *sometimes* predictable, I can
immagine this scheme working very well.  Given that they are not
*always* predictable, I can immagine this scheme oozing mildew.  In
general, tho' I think it would be a much larger win for scientific
computing than for (say) text editing.  Any insights?

	;-D on  ( Is this revile-lant? )  Pardo