[ewg] Re: [ofa-general] Feedback on Developer's Summit

Or Gerlitz ogerlitz at voltaire.com
Tue Nov 6 00:46:28 PST 2007


Dror Goldenberg wrote:
> a) it's an option that is turned off by default and people that are 
> willing to take the risk can turn it on and get better performance
> b) indeed we will need to make it work with the kernel lso, no need to 
> have something specific for ipoib

> Anyway, I am planning to present those things during the IPoIB SO talk. 
> Wouldn't this be good enough ?

Dror,

I have missed this post of yours, sorry for asking again if you can 
explain the approach. With what you say here, it seems that discussion 
on the LRO and non-complaint-checksum-offload patches can continue over 
  the list once they are submitted to review for the upstream kernel.

If possible, I'd like to hear more details on what does the HW actually 
knows to do (eg checksum and LSO offloads - for what IB transports, UDP 
or TCP, IPv6 or IPv4 only where the IP version matters), the slides in 
Sonoma were --very-- short to describe this and reviewing the patches 
without having the complete picture is tough.

Can you shed more light (Roland, indeed its more education then debate 
on open issues) on what you called in the Sonoma slides RCA (Receive 
Core Affinity), I understand this is actually multiple EQs? I'd be happy 
to learn on the motivation and suggested SW usage.

Dror/Roland - how about a 30m session between 11:30-12 (before the SA 
cache)?

Or.




More information about the general mailing list