[Planetlab-users] Re: can't login to nodes on slice..again..and again...and..again

[Robert P Ricci]

Thus spake Anirban Banerjee on Mon, Feb 26, 2007 at 08:35:51PM +0000:
> I think that as an engineer and researcher, it is imperative that we be able
> to come up with  "reproducible" results from our implementations. A small
> example: I implemented a basic DHT on PL to hash files to nodes. Now any
> results that I observe (n/w latency, CPU time..etc) are good for (say) the
> first 2-3 times I run my experiments. The 4th time some nodes/links fail and
> boom I have to replace the failed node (automatically with a script) and
> re-make the DHT/links/neighbor table etc.. I realize that I could I have
> used existing code on PL, but the point is : If I report results from PL,
> they should be reproducible. It seems flaky to report that average latency
> is x sec, till some crappy node /link screws up the curve ... and so on.

PlanetLab is good at providing you real network conditions, and in the
real world, links do change over time. This is one of PlanetLab's basic
strengths. (Another is that you can potentially get real users, since
PlanetLab nodes are all over the place.) For some experiments, this is
exactly what you want: to show that your application performs well "in
the real world," not just "under a controlled environment." For an
example of the type of research PlanetLab is great for, see this paper:
http://www.srhea.net/papers/opendht-worlds05.pdf

But, a consequence of the "real-ness" of PlanetLab is that it is a
fundamentally unreproducable environment. The network your packets
traverse is not controllable by PlanetLab, except at the very edge (ie.
the node's NIC). And you're sharing that network with tons of people,
not just other PlanetLab users. So, when you run your experiment again,
there is no guarantee that network conditions, or even network topology,
are the same as your first run.

If what you really want is an environment in which you can run
repeatable experiments with predictable network and host conditions,
then maybe what you want is an emulator.  You should give Emulab
(www.emulab.net) a look. (Caveat: This is a self-plug - in addition to
being a PlanetLab user, I'm also an Emulab developer.) You could also
check out ModelNet (modelnet.ucsd.edu).

-- 
/-----------------------------------------------------------
| Robert P Ricci <ricci at cs.utah.edu> | <ricci at flux.utah.edu>
| Research Associate, University of Utah Flux Group
| www.flux.utah.edu | www.emulab.net
\-----------------------------------------------------------