[net.arch] Virtual Memory on Supercomputers

lamaster@nike.uucp (Hugh LaMaster) (09/19/86)

If you refer to the hardware reference manuals for the Cray X-MP
and CDC Cyber 205, you will discover that each system claims to
do logical to physical address translation in one clock cycle.
There has been an assumption in some of the recent postings that
Cray address generation takes zero cycles but this is not so:
The logical address in the user's field length must be added to
the user's base address, which takes a cycle.  An associative
register virtual to physical address translation also takes one
cycle.  

For those who don't know, the Cyber 205 has two page sizes which
allow a large amount of data to be referenced without having to
even refer to the page table (current models address 8 MBytes with
"no extra penalty".  Hence, realistically, there is NO EXTRA PENALTY
for virtual address translation on the Cyber 205.

Furthermore, there has been a misunderstanding about "turning off
virtual memory".  User programs on the 205 always run with address
translation active.  It can't be turned off (nor should it) except
in monitor mode.  Users with programs whose references to data
arrays are such that all of the data must reside in physical memory
run with appropriate working sets to do this, so that the data is
not paged.  Perhaps this is what was meant by "turning off virtual
memory".  Even when every user's program requires this, there are
still benefits to having virtual memory available, as I have
mentioned previously.

The real penalty to virtual memory is the added logic required to
implement it; there is no penalty to using it, and there are very
real benefits to having it (better memory utilization, faster switch
times between jobs, etc. etc.)

I leave it Neil Lincoln, Seymour Cray, Steve Chen, et. al. to decide
what architectural features their next CPU's will have.  Virtual
memory in supercomputers is an engineering trade-off which should
not be approached as a moral issue.  It is similar to the RISC debate:
computer architectural design is like packing your backpack to go hiking:

If you need it, take it along.  And constantly keep paring away at 
unnecessary baggage.  (Apologies to Colin Fletcher)




   Hugh LaMaster, m/s 233-9,   UUCP:  {seismo,hplabs}!nike!pioneer!lamaster 
   NASA Ames Research Center   ARPA:  lamaster@ames-pioneer.arpa
   Moffett Field, CA 94035     ARPA:  lamaster%pioneer@ames.arpa
   Phone:  (415)694-6117       ARPA:  lamaster@ames.arc.nasa.gov

"The reasonable man adapts himself to the world, the unreasonable man
adapts the world to himself, therefore, all progress depends on the
unreasonable man." -- George Bernard Shaw

("Any opinions expressed herein are solely the responsibility of the
author and do not represent the opinions of NASA or the U.S. Government")