🚀 go-pugleaf

In article <157@s1.sys.uea.ac.uk> kh@s1.sys.uea.ac.uk (Kevin Hammond CMP) writes:
>Having had some experience of working with VM  operating systems,  I'm
>not sure that they really are that wonderful.   Demand Paging perhaps,
>to avoid writing overlays (you have something a  bit like this  on the
>Mac with the segment loader, I suppose), but  VM?  The problem is that
>if you really need more memory than you've got, you will always thrash
>(we have this problem with  our Sun workstations,  it makes them  very
>slow), and if you don't you don't really have a problem.  Buying  more
>memory (or  writing more efficient  programs) usually seems the better
>option if you have the choice.  Paging from a floppy or slow hard disk
>isn't my idea of fun either.

Here's a circumstance that often comes up in which paging is useful:

Suppose you're running some typical multi-tasking "big" operating systems
with many quiescent daemon processes, such as mail servers, print servers
or network servers.  Individually these programs do not take up enormous
amounts of memory, but together, the total amount of memory can be quite
significant.

But, 99% of the time these programs don't run!!--they just sit there waiting
for some event to happen.  In that case I'd want their memory to get
paged out onto disk so that my ray-tracing software can crunch with
all the memory it can get.  Of course, if on of the daemons should wake
up, there would be some paging delay and slowdown, but most of the time
they don't need all their memory, and they only use only 5ms of time
anyway, so it's OK.

I'd rather have VM so that if I need "just a little bit extra memory" for 
a little while I can get it, instead of getting a crash.


Also, many programs don't use alot of their memory (whether in program
code or data) for significant portions of the time.  By paging out
this temporarily unused memory, you can let other backround processes
get some RAM.

For very large compute-intensive jobs, however, VM cannot replace real
live chips.   BTW I believe that some models of Cray supercomputers do
not have virtual memory.  But I believe all ETA supercompters do.

Matt Kennel
mbkennel@phoenix.princeton.edu