From: Dan Clemmensen (Dan@Clemmensen.ShireNet.com)
Date: Mon Nov 23 1998 - 05:50:51 MST
Ken Meyering wrote:
>
> Dan@Clemmensen.ShireNet.com (Dan Clemmensen):
>
> > We've been building peer-processor datacomms swithces since I designed
> > the architecture for our first one in 1985. I do try to keep up with
> > this stuff. Please beleive me when I say that the networking isn't
> > particularly hard. The problem is in task assignment and task migration.
> >
> > Note that when the computers are all in the same room, there are cheaper
> > ways to build a Beowulf than by using ether-switches. In particular,
> > if the comuters are arranged in a grid, each computer can use two 100BaseT
> > cards instead of one. In this arrangement each computer can act as a crosspoint
> > for its row and column, switching on average only as many packets as it
> > originates or terminates. This gives you massive, incrementally expandble
> > bandwidth for approximately $50/node.
>
> Do you have much knowledge of radio telemetry and digital
> amplification? I'm wondering if it would be possible to build a
> generic wireless computing box that is capable of autonegotiating an
> optimal datacast hierarchy. In other words, using wireless signaling,
> have the boxes self-organize so that some data is being broadcast to
> all other boxes, and some data is being point-cast between boxes.
Sophisticated new radio work is happeninbg mostly in cellular
telephony. They are moving to "software-defined radio" techniques,
using serious amounts of DSP CPU power to manage the spectrum.
Your proposal would require even more CPU power. Why not just
apply the CPU power to the application instead, and use
cheaper datacomms? In general, the RF spectrum is quite limited
by comparison to fiber or even COAX.
>
> As far as "task migration", couldn't you have the tasks
> "automagically" balance by having them transferred between processing
> nodes directly through the air?
>
Task Migreatoin requires bandwidth, Bandwidth requires CPU at both
ends. The problem is determining whether the CPU woudl be better
spent on on the task rather than on migrating it.
> I'm not sure what the state of the art is in high-resolution
> high-speed gain amplification, but it seems that you could have a
> signalling system that doesn't pollute the electromagnetic spectrum,
> but is sufficiently intelligent to carry traffic across large areas.
Sure, but not at high bandwidth.
>
> Would it be possible to have a large supercomputer made up of such
> nodes, such that the same devices could also be a wireless
> telecommunications infrastructure, at the same time as being a
> distributed computing system?
>
> You could make it such that you mass produce the boxes (with
> superscalar RAM), to achieve the lowest possible per unit cost.
>
The comms component will cost a good deal more than the rest of the
computer to achieve perhaps 2Mbps connectivity. Cheaper to spend
about ten bucks per computer for 100baseT ethernet.
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 14:49:50 MST