[ofa-general] Re: New proposal for memory management

[Jason Gunthorpe]

On Wed, Apr 29, 2009 at 01:33:02PM -0700, Ralph Campbell wrote:
> Correct. This is what I was referring to when I said
> hardware support would be needed to make use of the 2.6.27
> page mapping callback notifications. I don't know of any
> current HCAs which support this sort of dynamic mapping.

You can use the callback notifications to pin the memory and register
it with the HCA when a new VMA is allocated, and do the reverse when a
VMA is removed.

Paging at the HCA level is only necessary to avoid pinning the ram.

MPI may be a special enough case where this is actually doable without
a huge overcommit. It would be very interesting to see /proc/PID/smaps
information for a running MPI job to compute how many unallocated
pages are present in a job.

-- 
Jason Gunthorpe <jgunthorpe at obsidianresearch.com>        (780)4406067x832
Chief Technology Officer, Obsidian Research Corp         Edmonton, Canada