[Beowulf] PVM on wireless...

Robert G. Brown rgb at phy.duke.edu
Wed Feb 6 13:24:06 PST 2008


On Wed, 6 Feb 2008, Jim Lux wrote:

> You brave man.. trying to do what is trivial in a wired network with wireless 
> stuff.
>
> I would look for timing assumptions that aren't met in the wireless 
> environment. There's a channel capacity issue, of course, but there's also 
> some constraints on round trip messages, particularly if you've got a 
> "infrastructure" network as opposed to "ad-hoc".  A packet from A to B has to 
> go from A to Access Point( AP), which takes some back and forth handshaking 
> and protocol overhead.  Then, it gets sent from AP to B, with more back and 
> forth.  Don't expect 1 ms ping times...
>
> I spent quite a while getting NTP (which I thought would be trivial.. it 
> explicitly handles long delays and intermittent connections) to work in a 
> 802.11a network, complicated by the fact that I was using Access Points (in a 
> "point to multipoint" configuration) as the interfaces, so the computers 
> actually had a wired ethernet connection through a dumb 5 port switch, to the 
> wireless AP. Getting PXE and DHCP to work was trivial by comparison
>
> Lots of weird things happen in these systems because there are hidden 
> assumptions about timing and whether a path exists between two points.

This is what I think that it probably is -- a race condition of some
sort caused by a timing assumption, almost certainly of UDP packets as
TCP should be robust.  I should look at whether PVM can be built on top
of just TCP these days.  It used to be UDP "for efficiency" but that
always means that you have to code your own reliability, packet
reordering and so on into the connection, usually leaving some things
OUT or you'll end up re-implementing TCP anyway, probably badly.  I
could be bitten by something left out that is causing certain packet
sequences to arrive out of (presumed) order and have the master waiting
forever for a packet that already came and was dropped.

But only a look at the raw code (or setting a tcp-only flag at build
time, if there is such a thing) will tell me for sure.

    rgb

>
> Jim
>
>

-- 
Robert G. Brown                            Phone(cell): 1-919-280-8443
Duke University Physics Dept, Box 90305
Durham, N.C. 27708-0305
Web: http://www.phy.duke.edu/~rgb
Book of Lilith Website: http://www.phy.duke.edu/~rgb/Lilith/Lilith.php
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977



More information about the Beowulf mailing list