[openib-general] mvapich-gen2 IA64 compile problem

Sayantan Sur surs at cse.ohio-state.edu
Wed Oct 12 11:33:52 PDT 2005


Hello,

* On Oct,5 Roland Dreier<rolandd at cisco.com> wrote :
>     Peter> If that is the purpose it fails badly since I have both
>     Peter> EM64T machines with PCI-X and AMD64 machines with
>     Peter> PCI-express. Or am I missing something here?
> 
> Nope, the fact that it's messed up is what I was trying to point out.
> >From a quick look at the code, it looks like you should set the
> environment variable VIADEV_DEFAULT_MTU to "MTU1024" on PCI-X systems
> and "MTU2048" on PCI Express systems.
> 
> In general, one wants to use the largest possible MTU to get maximum
> performance.  However, Mellanox PCI-X HCAs have a quirk that causes an
> MTU of 1024 to be faster than an MTU of 2048 for RC transport.  This
> quirk does not exist in PCI Express HCAs, and presumably PathScale and
> IBM HCAs also perform best with the maximum possible MTU.  So Mellanox
> PCI-X HCAs are the only case where we want to use a lower MTU.

Since the MTU size can affect the performance and there are several
different types of HCAs available in the market, we are working towards
making a runtime decision of the MTU instead of compiling it in, or
asking the user to input. This patch should be available sometime soon.

Thanks,
Sayantan.

> 
>  - R.
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general

-- 
http://www.cse.ohio-state.edu/~surs



More information about the general mailing list