Re: [Etherboot-developers] multicast

[Eric W B. <ebi...@ln...>]

Ronald G Minnich <rmi...@la...> writes:

> FYI, Erik hendriks (here) just REMOVED multicast support from his beoboot
> code. There were just enough problems with switches etc. that in the end
> he decided it was a net loss performance-wise.
> 
> This may or may not affect what you want to do with multicast.

Hmm. I hadn't realized this was on the list.  A little more general
answer then.

For most drivers opening up the nic level filter to accept multicast
packets is about 2 or 3 lines of code.  So it is worth writing the code,
and setting the precedent even if multicast booting does not prove
a general purpose win.  Except for those 2 or 3 lines there is no overhead
for a multicast capable etherboot if you don't compile in the support.

As for how worthwhile the project is.  I coded the support after watching
the unicast traffic for booting several hundred computers simultaneously
be a human visble bottleneck.  Taking something like 15minutes and with atftp-0.3
showing up in top.

Beyond that we will be running a full scale test on MCR a 1000 node
cluster, in a couple of weeks, and I can give a report on how useful
it is.  So far all I have heard is, that it works and is boring.  I
don't plan to reevaluate that until 

To date I have not encoutered any fatal switch problems, and requesting
that switches that work on a large cluster where it is needed is not a
real problem.  When I use multicast I would be perfectly happy to use
broadcast,  the multicast is just a hint to the hardware to not send
the packets to everyone. 

Finally what was Erik doing that made multicast support less
performant than unicast?  Did he have a bad retransmission algorithm?
Even in the case the switch drops 50% of the packets, if you have 3
receivers for the same data you have gained performance.  Unless the
multicast traffic is jamming up your network between nodes doing other
things, in which case multicast is quite definitely not the tool for
the job.

Eric