From: Bjoern Olausson (namdlist_at_googlemail.com)
Date: Thu Jun 16 2011 - 07:40:07 CDT
On Thursday 16 June 2011 13:23:36 Ajasja Ljubetič wrote:
> > hmm... i wouldn't say that. you have to be a bit more creative. ;)
> > for example, you can set up a small 3 node cluster
> > without(!) an expensive switch by simply buying 3 dual
> > port IB cards and 3 cables.
> Wow, that is a nice trick!
> So something like this <http://geizhals.at/a642147.html> and some cables
> and one is in business? Does this also work with four cards and four
> cables (so what I'm really asking is: must every node be directly
> connected to every other?)
> And if one would like to connect just two such nodes, does it make sense to
> buy two cables?
Your aproach remindes me of the old day when the 10BASE2 Network was stat of
the art ;-) But I don't know if it is possible to daisychain Infiniband cards
like that or to connect them in a Token ring manner.
> And since most motherboards have multiple network cards these days, does
> ethernet bonding <http://en.wikipedia.org/wiki/Channel_bonding> make sense
> for NAMD applications if one does not have IB?
Only if you need the increased transfer rate.
The point where IB outperforms any ETHERNET protocol is the latency and this
is what you are out for commonly when using Infiniband instead of ETHERNET.
Using ETHERNET your CPUs will spend a lot of time idle waiting for new data
packages and most likely the delay is not due to traffic jams because of
If I am not totally of, a one million particle system produces ~50MB of data
per frame and I bet your system will be smaller, won't it?
Latency for ETHERNET is >100 microseconds compared to IB with an end to end
latency of less then 4 microsecond and bonding will not reduce the latency
So at last the latency decides how many ns/day you will crunch (Given that you
simulation size will not exhaust the transfer bandwidth)
-- Bjoern Olausson Martin-Luther-Universität Halle-Wittenberg Fachbereich Biochemie/Biotechnologie Kurt-Mothes-Str. 3 06120 Halle/Saale Phone: +49-345-55-24942
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 05:24:06 CST