Hi Chip and Robert,

James Osborn has been thinking about how smoothly the MILC code would
sit on the interface and had some questions, which I couldn't answer:

1. It seems the MSGHANDLE objects are multipurpose.  Are they going to
   be diversified?

2. Is there a reason to use MSGHANDLE for broadcasts?  Aren't they
   always blocking operations?

3. What happens on a switch-based machine when you call
   declareSendRelative?

4. Is there a good reason to split the SendRelative and
   general-purpose SendTo calls?  Why not have a common send call and
   let the interface decide whether it can be done via a SendRelative
   or SendTo?  This would be implemented by assigning each node a
   unique number (as is done currently with MPI, etc) and sending to a
   destination node.  If we were running on a grid-based machine, the
   interface would have access to the physical grid displacement
   implied by the send, and could then invoke the appropriate send.
   One could do multiple sends in the same way - simply provide a list
   of destination node numbers, etc, and have the interface figure out
   how to set it up as a multiple send or a sequential send.

   I am assuming that with SendTo's on a grid-based machine, the
   interface is going to figure out how to do that efficiently.

5. Are we assuming that SendRelative is SIMD?  Or could one mix
   SendTo's and SendRelative's?

Regards,
Carleton

PS James heard a rumor from his parents that Jlab security has
   recently made it very difficult for visitors to enter the lab.  Is
   this correct?