Heh, this reminds me of my time at Avici Systems: they made a 3d toroidal mesh-based "Terabit Switch Router" using wormhole routing with virtual channels to avoid head of line blocking (Bill Dally was associated with the company..). The company went public and provided the core routers for AT&T for a number of years. (Superseded by Cisco's HFR (Huge F*ing Router..).
The questions involved how much can you get away with (before deadlock) if you don't have enough virtual channels. Also is it ever worth extra buffering outside of the VCs. Also, how to get more bandwidth if one VC does not provide enough. The TSR used source routing (a list of turns on the head flit I think..), so how do you compute the routing list for arbitrarily incomplete meshes?
It's interesting that Feynman was involved in this field of interconnection networks, although for actual papers I do find his son's writing on the CM-5.
The questions involved how much can you get away with (before deadlock) if you don't have enough virtual channels. Also is it ever worth extra buffering outside of the VCs. Also, how to get more bandwidth if one VC does not provide enough. The TSR used source routing (a list of turns on the head flit I think..), so how do you compute the routing list for arbitrarily incomplete meshes?
It's interesting that Feynman was involved in this field of interconnection networks, although for actual papers I do find his son's writing on the CM-5.