Re: 3.0 wishlist Was: Overview of 2.2.x goals?

Larry McVoy (
Wed, 21 Jan 1998 10:04:07 -0800

: >Actually the Cray Origin 2000 systems do something like this. The pages
: >can migrate between nodes and the OS tries to automatically arrange the
: >physical location of the pages so that memory access to remote nodes is
: >minimized. I'm not sure how well this actually works in practice though.
: I didn't know this, but I seem to be right in the assumption that one can
: not easily find distributed shared memory in more-available systems.

This stuff on the Origin doesn't work very well. I was there when they
put it in, and I believe that they ship the system with it turned off.
The overhead of moving the page was worse than just getting the data
remotely (the Origin has remote shared memory with hardware support,
you can't tell that your memory isn't right here unless you measure
latency and latency to a remote page was typically a 100 ns more than
a a local page).

: >It could be helpful to people who want to run numeric codes that were
: >developed for the T3D/E, Origin 2000, etc. It probably wouldn't be that
: >difficult to build something like the Cray distributed shmem library on
: >top of DIPC.
: Scientific programming is among the ideal domains for DIPC. Linux is more
: available than some of the systems you mentioned above, so I guess porting
: such libraries will be a great help to many people.
: This is not just about performance: People can develop and test their
: applications on cheap multi-computers, built using Linux, and then run it
: on more speedy hardware, if they ever need to.

I agree with the "Linux is a widely available platform part". Scientific
programming has moved to the MPI (message passing interface) model in a
big way. Even on SMP machines. If anyone is serious about supporting
scientific apps, then get an optimized version of MPI up on Linux.
I'll bet you that the Beowolf people are working on this.

DIPC is cute but is likely to be ignored by the big apps.