On Mon, Apr 09, 2001 at 12:07:05AM -0700, Erik E. Fair wrote:
> Don't forget latency. A gigabit per second to a memory on the moon 
> would be very painful to use, despite the amount of bandwidth.

Yes, yes, I know, and I'm not suggesting anyone would want to use
this as a practical replacement for swap any time soon, if ever.
(It's far more probably that we will just have more and more
ludicrous mounts of RAM than that we would want to move backing
store somewhere else... and "mere" gigabit ether strikes me as
slower than even aging disk hardware; myranet or ATM at least would
be necessary to get even marginally decent performance). As I said,
it's a step on the way to in-kernel, networked, shared memory (which
it would be acceptable to have slower than swap if what you want is
seamless IPC in a cluster).

As you point out, bandwidth is really less the issue than latency
(though clogging of the route causes greater latency in either
medium). But all this is more academic than practical right now
anyway. :^>

(Although... what I'm doing *might* beat out the performance of
swap over NFS for diskless machines, which would be pleasant for
some people, I expect.)

       ~ g r @ eclipsed.net