Design Alternatives (2)
2-D Decomposition - On a ÖP ´ ÖP processor grid, messages of size 2N/ÖP ´ Z to 4 neighbors, so
T = tcN2Z/P + ts4 + tw2NZ/ÖP
Good if ts < twNZ(2-1/ÖP)
3-D Decomposition - On a Px ´ Py ´ Pz grid,
T = tcN2Z/P + ts6 + tw2(N2/(PxPy) + (NZ)/(PxPz) +(NZ)/(PyPz) )