[Beowulf] which 24 port unmanaged GigE switch?

[Håkon Bugge]

On Apr 6, 2010, at 19:01 , David Mathog wrote:

> What's bothering me though is that both D-link and Netgear claim the
> same top forwarding rate as the more expensive switches, at 48 Gbps. 
> Giving them the benefit of the doubt, that they are not intentionally
> misrepresenting the speed of these units, perhaps there is some other
> variable in play which degrades their performance more than the other
> switches?  

You should check that XON/XOFF (PAUSE protocol) is enabled in _both_ directions by means of ethtool. DIfferent switches have different defaults. Sometimes, enabling XON/XOFF has a dramatic positive effect on packet loss.

> This is another one of those cases where having a benchmark would be
> enormously helpful.  


I tend to use an MPI which can run over TCP/IP and run a naive all-to-all benchmark, changing payload sizes and number of active MPI processes per node. As a side note, Linux has in the order of ten different congestion protocols, but none of them seems useful in an HPC environment, when you're exposed to packet loss.


Håkon