[openib-general] Re: Problem configuring ipath_ether

Michael S. Tsirkin mst at mellanox.co.il
Sun Mar 26 07:23:31 PST 2006


Quoting r. Bryan O'Sullivan <bos at pathscale.com>:
> Subject: Re: [openib-general] Re: Problem configuring ipath_ether
> 
> On Sat, 2006-03-25 at 22:36 +0200, Michael S. Tsirkin wrote:
> 
> > Since IP over IB
> > just puts an IP packet inside a UD packet without overhead, what wire protocol
> > changes can thinkably be done, while still using InfiniBand protocols, to
> > improve performance?
> 
> The ipath_ether device provides an MTU of ~8K, by sending packets back
> to back.  Even though a real 8K MTU isn't possible, this technique gives
> considerably better throughput than IPoIB, at least on hardware that can
> sustain a high messaging rate.

Thanks, sounds interesting. Some questions:
Grepping for MTU I even see 16384 and 32K - this is done with the same
technique, isn't it?
Any ideas on why isn't something like this done in the standard IPoIB then?
Anyone on the list knows whether/when this is a good idea?
I expect it would be relatively easy to hack something like this into existing
IPoIB codebase - I expect no one tried this yet?

-- 
Michael S. Tsirkin
Staff Engineer, Mellanox Technologies



More information about the general mailing list