[ofa-general] RE: tcp/rdma port unification patch in librdmacm

[Steve Wise]

Or Gerlitz wrote:
> Sean Hefty wrote:
>> I thought about that, but didn't know whether it was in use.  I agree 
>> that
>> existing apps shouldn't break.  (I was thinking more along the lines 
>> of adding a
>> new call that would make this behavior explicit, but haven't taken 
>> the time to
>> really study the details.)
>>   
>
> Yes, lets not break existing apps such as stgt. I'm fine with adding 
> new call.
>


Or maybe map PS_TCP port space to the be the unified space, and have a 
new PS_RDMA or PS_IB port space that doesn't?


>> Does anyone know if a kernel patch to fix this has been accepted 
>> directly into
>> the distros?
>>   
>
> Sorry, but I wasn't sure to follow what you mean by "fix this" ... did 
> you refer to kernel apps that don't use different port numbers for 
> their TCP vs RDMA listeners, or you referred to the rdma_cm patch 
> which is not merged into mainline?
>


I haven't pushed to get a distro to pull in the cma patch.  I assumed 
since it wasn't destined for upstream, that they wouldn't accept it.


>> Also, does anything keep MPI from doing exactly what we're discussing 
>> in the
>> application as part of using the librdmacm?  (Besides having all apps 
>> duplicate this
>>   
> no, nothing prevents an app to open/bound a socket as long as the port 
> is available and its ulimit allow to open the number of sockets it 
> wants. This actually somehow brings me back to square one with regard 
> to the actual problem, so I'll ask you guys a question as a reply to 
> earlier post on this thread...
>

Nothing prevents this, but why force OpenMPI, MVAPICH2, HP MPI, Intel 
MPI, and Scali MPI to all do this?


Steve.