Re: InfiniBand Implementations

Nick Maclaren wrote:
> 
> Well, as I asked about HyperTransport, I shall ask about InfiniBand :-)
> 
> As previous threads demonstrated, the InfiniBand specification attempts
> to be all things to all me, but it was and is very clear that the way
> it is going to be used is that implementors of devices, software and
> perhaps even interfaces will make assumptions about which subsets they
> are going to use and optimise or even support accordingly.  Just as
> happened with SCSI.
> 
> My specific question is about the software designs, and the way that
> the hardware interfaces intend that the software should use them.  In
> particular, that dreaded question of user-space DMA, with its related
> issues of whether the user can provide arbitrary locations for DMA
> use, the scope of pinning of such locations, what support there is for
> memory migration or swapping and so on.
> 
> Has anyone started to document any of that yet?  Because it is obviously
> a prerequisite for writing (or even designing) efficient and robust
> code to use InfiniBand.
> 

You mean something along the lines of http://www.rdmaconsortium.org/home ?
It is a little vague about OS accomadation of the DMA stuff.  But maybe that's
deliberate for vendor lock in. :)

Joe Seigh