[Beowulf] TIPC in a Beowulf?

Mark Hahn hahn at mcmaster.ca
Sun Feb 3 14:42:56 PST 2008


> Has anyone heard of or seen TIPC <tipc.sourceforge.net> used in a
> Beowulf Cluster?

I haven't.  I sat in on tipc meetings at OLS a few times, and have the 
impression that TIPC people are much more into telecom/footprint issues
rather than HPC.  (and yes, I believe these are very different focuses - 
for HPC, the main issue is latency (since bandwidth is not that hard.))

I _think_ I'm not confusing TIPC with SCTP (which also seems to be rather
telecom-oriented.)

here are some kind of shocking performance measures:
http://www.strlen.de/tipc/

no mention of latency there.

> Some folks from Wind River (creators of the protocol I think) came and
> gave a talk about it at my school. They said it can be used over IP,
> or even on it's own through ethernet, and would even work with myrinet
> or infiniband with proper drivers.

well, TIPC is trying to do a lot that TCP isn't.  for instance, I think 
it's trying to do fairly full group membership as well as topology-aware
routing.  I'm not sure these are as critical to HPC-type clustering 
as they would be for HA-type clustering.

I'm also a bit skeptical of a protocol that aims to put everything into
one kernel-resident layer...

> I'm still not very familiar with programming a Beowulf, but Inter
> Process Communication is an equally viable paradigm just like Message
> Passing, right?

TIPC is a form of MP.  don't confuse MP with MPI!  MPI is important and 
widespread, but I don't think many people would say that it's perfect.
MPI-over TCP in particular is kind of a shame, since TCP is really a 
protocol designed for flakey, overloaded, heterogenous WANs, 
not the kind of dedicated, homogenous, flat network you find in an HPC cluster.

I'm looking forward to OpenMX - it's a message-passing layer amenable to 
ethernet, but well-suited for MPI.  any OpenMX people care to comment?

regards, mark hahn.



More information about the Beowulf mailing list