[ofa-general] New proposal for memory management

Woodruff, Robert J robert.j.woodruff at intel.com
Wed Apr 29 15:52:22 PDT 2009


Brian wrote,

>I think there are other problems with the verbs interface that would still
>make MPI implementers twitch (some of which are in the slides Jeff sent out
>to begin this discussion).  But I certainly wouldn't say no to a real set of
>tag matching primitives.  Of course, that opens a whole can of worms that
>I'm not sure OFED is ready to deal with.

>It also may or may not solve the memory registration problem.  If the memory
>in the matching verb still had to be registered, we haven't solved the
>problem that started this discussion.  So the verb would have to also handle
>memory registration, which seems to go against the general "OFA way". 

I think if we did such a thing, we could implement a set of tag-matching 
primitives (similar to MX or PSM) that are kind of a separate library 
from the OFA RDMA verbs, just like PSM for Qlogic is a separate library and
not part of the OFA verbs. Just like with MX and PSM, I think the registration
can be done my the tag-matching driver (like PSM or MX do) and
not require MPI to do it. Think of this as "the MPI tag-matching interface" library 
for OFA. 

However, this would only completely solve your problem and complexity of using the
OFA RDMA verbs if all the hardware vendors implemented tag-matching in their 
NICs. Seems like if they want to better support MPIs, that is what they
would do and then MPIs would only have to use the simple tag-matching
primitives and would not have to worry about things like memory registration
caches and such.

Anyway, I think it is an interesting idea worth perusing with the IHVs as the long
term solution to most of the issues that Jeff raised in Sonoma. 

woody




More information about the general mailing list