Wireless Beowulf?

Jim Lux James.P.Lux at jpl.nasa.gov
Mon Nov 25 17:11:53 EST 2002


As a practical matter, the channels overlap quite a lot.. In the US, for 
instance, channel 1 is from 2.401 to 2.423 (center is 2.412), and channel 2 
is from 2.404 to 2.428 (center is 2.417)...
The 11 Mbps connection is going to occupy at least 11 MHz of real bandwidth 
(more like 20 MHz in reality, by the time you count sidelobes, etc.), so if 
you try and run 802.11 links on channel 1 and channel 2 simultaneously, 
they'll work, but interfere with each other (reducing the effective range).

To guarantee no overlap.. Channel 1, Channel 6, Channel 11 is one combination

Contrary to popular belief, spread spectrum does NOT increase the basic 
capacity of a chunk of bandwidth in a given spatial location. It just 
provides another way to divide up the resource.

Also, 11 Mbps is the on-air speed, not the actual over the wire bit rate.. 
If machine A is talking via an accesspoint to machine B (both with wireless 
connections)(i.e. not direct A to B), then each packet has to go from A to 
the AP, then get transmitted from the AP to B.  The over-the-air link is 
one way at a time (a given node doesn't transmit and receive 
simultaneously). I don't know about broadcast packets (i.e. does 802.11b or 
g support one to many transmissions...) You could also do "adhoc" mode 
configurations of 802.11, which is more like shared coax cable networking 
(10Base5...)

Finally, there is a fair amount of overhead in the packets as transmitted 
over the air.  The start of the packet has a header consisting of some 192 
bits at 1 Mbps before the "real data" gets transmitted at 1,2,5.5, or 11 
Mbps.  That's almost 200 microseconds for header, and if you send a 1500 
byte data packet behind that at 11Mbps, the header takes 20% of the total 
packet time. (there is a "short header" defined.. who knows if it's 
actually used by any available gear...)


As John pointed out, wireless is an attractive way to do network management 
or talk to a specific node, but probably not appropriate for the main 
fabric (except in a very embarassingly parallel situation).  one also has 
to consider the cost of the wireless interface itself..

Maybe a wireless access point/interface connected to the main switch, and 
then you use that to talk to the individual nodes.

At 09:30 PM 11/25/2002 +0100, you wrote:
>On Mon, 25 Nov 2002, Michael Worsham wrote:
>
> > Has anyone researched using the wireless 11a (or soon to be released 11g)
> > protocol as an idea for simplistic beowulf design? I know it won't be as
> > fast as fiber, but rather than having 5,000+ wires hanging everywhere,
> > perhaps it could be a viable idea in a closed-room location?
> >
>
>On 802.11b, in Europe, there are 13 radio channels available. Guess for
>a lot of ndes communicating this wouldn't be that hot!
>
>But this is a nice idea!
>
>Me, I think an interesting use of wireless with clusters is for
>out-of-band management and diagnosis.
>I certainly have found it useful to have my laptop with a wireless card in
>handy - in a massive server room this saves running in and out to (say)
>do some pings or ssh sessions, or to check things on the Web.
>Cyclades also have terminal servers with built-in wireless - which I think
>could prove very useful if the main network or a big router goes kaput!
>
>
>_______________________________________________
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit 
>http://www.beowulf.org/mailman/listinfo/beowulf


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list