Virtual Memory and Control Theory

[Jerry Leichter]

All that is true but may very well be beside the point:  What't the total
cost of doing virtual memory?  You can't look just at the benefits if you
claim to provide a mathematical justification.

Years back, Gene Amdahl was asked why he didn't put virtual memory support
into the 360 architecture.  His answer:  He had never seen a virtual memory
system that didn't impose a 10-12% performance hit, which he found unacceptable.

Fade 20-odd years later.  I.P. Sharp is a major Canadian timesharing service,
providing APL to large numbers of users.  They use Amdahl 470's.  They run
the Amdahls with virtual addressing turned off.  Guess what:  Sharp's 470's
give about 10% more performance than anyone else's 470's.

(I'm afraid I don't remember the reference for these interesting facts; it
was a pretty authoritative one, though.)
							-- Jerry
					decvax!yale-comix!leichter leichter at yale

BTW, before you get TOO carried away in your analysis:  A non-virtual memory
system with overlays is a closed-loop system.  There are a LOT of possible
approaches.  The Sharp APL implementations presumably keep all the system
code in memory at all times, and move user workspaces in and out as needed;
since the code that controls the movement knows a LOT about the semantics
of the workspaces, it is in a much better position to make good decisions
about what and when to page where than a straight VM system, which has to
guess.
							-- J