[Beowulf] Re: Wireless cluster

[Jim Lux]

At 06:35 AM 3/24/2006, Eugen Leitl wrote:

>From: jmmckel at attglobal.net
>
>----- Forwarded message from "John McKelvey jmmckel%x%attglobal.net" 
><owner-chemistry at ccl.net> -----
>
>From: "John McKelvey jmmckel%x%attglobal.net" <owner-chemistry at ccl.net>
>Date: Fri, 24 Mar 2006 09:28:57 -0500
>To: "Leitl, Eugen -id#3h6-" <eugen at leitl.org>
>Subject: CCL: Wireless cluster
>Reply-To: CCL Subscribers <chemistry at ccl.net>
>
>Sent to CCL by: John McKelvey [jmmckel!^!attglobal.net]
>Cheers!
>
>Please don't laugh, or maybe this is so funny [ridiculous?] that you'll
>get a good laugh, and it will make your day!
>
>Anyway, I want to build a small linux cluster [6-8 total processors] but
>don't have a lot of  cooling in one place.  Running wires would not be
>practical, if not impossible  Now, the application is extremely
>coarse-grain, and only a very, very small amount of data gets moved
>about once initialization of a job is done..  Can it be done "wireless?"


Most definitely. As Verne quotes Arne Saknussemm,"Quod feci!" (or, 
depending on which translation: "Hod feci".. these things are important.. 
just like exponentiation as A^B or A**B, or 026 vs 029){and no, I don't 
know how to say it in West Saxon, yet}


I used 802.11a wireless access points (Dlink D7000AP) in 
point-to-multipoint bridging mode on each node.


Be very aware that data rates on wireless links (claims of tens of Mbps 
notwithstanding) are truly low.  AND, there's a huge latency.  Ping times 
for short packets) between nodes (with the nodes side by side) are in the 
100 millisecond range for the first ping, then drop on successive pings to 
tens of milliseconds.

First, the wireless channel is shared among all contenders, and they can 
only transmit one at a time. And there's not any centralized allocation of 
time slots.  THink in terms of old style wired ethernet all on the same coax.
Second, the packets all start at the low speed(2Mbps), and then change to 
the high speed in the middle of the packet, assuming propagation is good 
enough.
Third, there's a fair amount of overhead on each packet (in terms of time) 
for synchronization sequences, settling time for the radios, etc.
Fourth, there is a LOT of non-determinism in the packet timing.  Don't be 
expecting to run NTP and synchronize your clocks to microseconds with wireless.

However, except for speed, it works fairly well.. you can PXE boot across 
the wireless network from a NFS server, for instance.

>Please share your laughs...
>
>Cheers..
>John McKelvey
>

James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875