Birds Eye View of Interconnection Networks The Power of Topology The Effect of Network The Grace of Solution
Glimpse at Interconnection Networks
Seas for Networks A wide variety of direct interconnection networks have been proposed for, or used in, parallel computers They differ in topological, performance, robustness, and reliability attributes.
Classification: Static vs. Dynamic Static or Direct Ring Mesh Hypercube ……… Dynamic or Indirect: Bus, Crossbar, MIN 000 001 100 110 111 011 010 101 P1 P2 P3 M1 M2 M3
Line, Ring and Fully Connected Linear Array: the simplest topology. Fully Connected: Direct connection between every pair of processors, Highest cost, Similar to crossbar in some properties. Ring: Each node in the ring is connected to only two other nodes.Chordal Ring: A compromise between Ring and Fully Connected Network.
Mesh and Torus Mesh Topology 2D Torus 3D Torus Why mesh or Torus: 1) make full use of boundary link; 2) Node/edge Symmetry. Is there any other wraparound approach? Read my Midimew Papers. What is the diameter of torus, in comparison with meish? How to layout the torus? Or a ring? 2D torus: Meshes with ``wraparound'' connections, e.g. the node at the top of the grid has an ``up'' link that connects to the node at the bottom of the grid (also left to right).
Hypercube From 2-dimentional to 3-dimensional or even higher-dimensional.
Tree and Star Tree and Star
Fat Tree four paths 8 port 4 port PE1 PE2 PE3 PE4 PE5 PE6 PE7 PE8 PE9 PE10 PE11 PE12 PE13 PE14 PE15 PE16 All 8 PEs on the left half(subtree) have a path to the Pes on the right half Note that a message going from PE2 to PE5 may choose any one of four paths from the lower left router to the one at the root. This means that all four PE's attached to the lower left router have a path available to them to reach another node.
Crossbar switch Crossbar switch P0 –> M1 P1 –> M3 P2 –> M2 P3 –> M0 Crossbar switch On the other end of the spectrum is bus topology. Cheapest, but poorest in performance. So again, we have to make compromise/trade off between cost and performance. We therefore have multistage. N*N switches (N processors N memory modules). The switch configures itself dynamically to connect a processor to a memory module. No contention -- Supports N! permutations. Costly, hard to scale, wastes switches for most patterns.
Multistage Network : Omega Network p processors and log2 p stages Each stage consists of a perfect shuffle
Multistage Network : Butterfly ButterFly Network