> > Regarding the memory issue w/o SRQ, do you think there is a way to use
low
> > watermark to release prepost buffer in large connections?
>
> Maybe with UC - with RC you'll get RNR and connection'll get closed
before you
> have time to handle the low watermark. So sure, might be an interesting
idea,
> but isn't low watermark a SRQ feature?
>
> > I think most of the prepost buffers are empty in that case becauseof
the BW.
>
> I don't really get the argument.
>
> --
> MST
That's just some random idea. :) Some other ideas like to share RQ
buffer based on source-destination address, per CPU RQ buffer ...which
might hurt performance too much?
It might be too complicated to have UD/RC mode coexisted?
Maybe it's better to set up a small RQ size for now, and later when
high watermark patch is available we can use it to address RQ overrun?
Thanks
Shirley_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
