Initially we spoke with folks about replacing MDT with a "soft" LSO
implementation, participated in some discussions about it, then decided
not to do it. Instead, we are planning to use extend the IB driver
(called IPoIB) to use the LSO capability that is built into the most
recent generation IB HCAs.
-David
Cathy Zhou wrote:
Also, can you point to a definition of the MDT interface? Are there
any consumers of it already in the tree so I can see how it works?
All this talk of it has made me quite curious.
No. The only consumer I know of for MDT is Cassini. I doubt it will
ever be opened, but I certainly don't speak for the powers that be.
(Otherwise it would long ago have been ported to GLDv3 and open
sourced. :-)
It seems to also be in GLDv2, but only for use of Infiniband
drivers. gld_register() seems to allow on IB to use MDT.
Interesting, I had forgotten about that one. IB should be ported to
GLDv3, if not already done. (I know I code reviewed such a while back.
Not sure if it ever integrated.)
Yes, the ibd driver has been ported to GLDv3. I think there was concern
that there could be performance regression with the GLDv3 ibd driver since
GLDv3 doesn't support MDT, and I am not sure what is the final number. But
since the porting is done, I guess it turned out to be a non issue.
- Cathy
_______________________________________________
networking-discuss mailing list
[email protected]
_______________________________________________
networking-discuss mailing list
[email protected]
