[openib-general] putting in dead wood for DAPL and similar abomination

Jon Bauman jon-openib at umich.edu
Wed Mar 2 14:32:22 PST 2005


At 05:05 PM 3/1/2005, Christoph Hellwig wrote:
> Similar hint to the NFS over RDMA folks at CITI -
> if you want your stuff to go in use the openib helper directly below
> the transport switch - differnet RDMA transports are too diverse to
> be sanely abstracted out and DAPL does a horrible job at that.  If
> we need to consolidate code for differnt transports we can put it
> into a library later on.

CITI folk here.

I'm not familiar with the openib helper you refer to, but since you 
mention the transport switch, I'll assume you're referring to the 
client. I'm currently working on the NFS over RPCRDMA server, so this 
isn't of much help to me.

While I'd agree that DAPL has it's shortcomings, it's not finalized 
yet, and I know of no other alternatives. On the other hand, I don't 
agree that the different RDMA transports are necessarily too diverse to 
provide a reasonable API for them. It seems silly to invest a lot of 
effort writing directly for IB, since we couldn't reuse the code for 
other transports. Why create a nonstandard library after the fact when 
so much work has gone into DAPL already? Even if DAPL needs to change, 
we can later make our changes just once at that layer.

We should have basic functionality with NFS atop DAPL in the near 
future that will enable us to plug in different transports without 
changing the ULP code. Would that convince you that DAPL is at least a 
useful starting point?




More information about the general mailing list