[ofa-general] New proposal for memory management

[Tom Talpey]

At 09:07 AM 5/1/2009, Jeff Squyres wrote:
>On Apr 30, 2009, at 6:10 PM, Barrett, Brian W wrote:
>
>> I'm done now.  You don't want to fix your crap, that's fine.  Just  
>> don't be
>> surprised by the continued "why you shouldn't use IB" presentations  
>> from
>> people who have to write applications to it.
>>
>
>
>Let's not forget that Brian is not only an MPI developer (i.e., a  
>network programmer), he's also a customer.
>
>If OpenFabrics only wants the HPC market, you can probably ignore this  
>entire thread.  The OpenFabrics-based MPI's will hobble along like  
>they have been.  If you want larger markets, it's probably pretty safe  
>to assume that Brian's reactions are going to be quite similar to  
>enterprise network programmers.

Completely agree. I will add that enterprise network programmers are
going to reject registration caching as well, because it introduces
vulnerabilities into the data path - silent data corruption. For example,
storage won't tolerate it, databases won't, etc.

The problem is that userspace memory registration is slow. Let's address
that, not address how to make a hack (registration caching) go faster.

We've solved this in the kernel with FRMR, why not take a similar solution
up to user verbs? Wouldn't that address it, by allowing the library to safely
and efficiently manage registration on a per-io basis?

Tom.