Download to read offline
BIL406-Chapter-5-Network Structures.ppt
- 1.
- 2. Chapter 5 Network Structures •5.1 Introduction • 5.2 System interconnection architectures • 5.3 Network properties and routing; Node degree and Network diameter; Node degree, Network diameter, Average Distance and Bisection width • 5.4 Data routing functions; Perfect shuffle and exchange, Hypercube routing function, Broadcast and Multicast and Network throughput
- 3. • 5.5 Networkperformance • 5.6 Static networks ; Point to point Networks> Binary Tree, ternary tree and quadtree, Fat tree, Linear arrays, Rings, Complete graph, Grid and Torus, AMP (A Minimum Path Systems and Hexagonal Grid • 5.7 Dynamic networks; Bus networks and Switch networks> Switch modules, Multi-stage networks, Delta networks or Omega networks, Closs networks and Crossbar networks • 5.8 Comparison of networks
- 4. 5.1 Introduction • Everyparallel computers contains a set of processors and one or more memory modules • These functional units must be connected with another by some type of network. • High level communication; shared memory (virtual or shared) or exchange message. • Network structure should have sufficiently high connectivity (without ant intermediate stations (connections)). • There are number of limitations constraining the physical construction of the network. • Number of connection lines per processors is limited.
- 5. • Cost onn Pes – a) Number of connection per PE (production cost, network throughput) – b) Average Distance between the PEs (network diameter, operating cost) • Distance must be low at acceptable level. • The connection structure should be scaleable (from smaller to larger.) • Network structures are divided into three major classes. – 1 Bus Networks – 2 Switching networks – 3 Point to point networks • Another classification is – 1 static networks – 2 Dynamic networks
- 6. 5.2 System interconnection architectures •Static and dynamic networks are used to connect computer sub system of to construct a parallel system. • • connect disk, memory, I/O and etc. • Various topologies are specified • Scalability, • low latency, high data transfer, • communication bandwidth.
- 7. 5.2 Network propertiesand routing • Static networks are formed of point-to-point direct connections (fixed connection) • Dynamic networks are implemented with switched channels (changing connection). • Before analyze various topology, let us define several parameters used to estimate complexity, efficiency, and cost of network.
- 8. Node degree andNetwork diameter • Node degree • The number of edges ( links or channels) incidents on a node called the node degree d. (in degree, out degree)
- 9. Network diameter • Networkdiameter • The Diameter D of a network is the maximum shortest path between any two nodes. ( The length is measured by number of links traversed) • • To construct minimum path is to minimize network diamter • Dmax , diameter of interconnection network, to minimize number of links a message has to travel between any source processor and other destination within the configuration. ( at the expense of the loss of regularity in a system)
- 10. Average Distance • Theaverage inter-processor distance, davg is the number of links on average that message to traverse between a source and a destination processor. • The average distance for the n processors for a configuration of diameter dmax will be determined. • • p=1 n ( d=1 dmax d * Npd)/ (n-1) • Davg = ------------------------------------ • n • • Npd is the number of distance d away from processor p.
- 11. Bisection width • Whena network is cut into two equal halves, the minimum number of edged ( channels) along the cut is called channel bisection. • • Wire bisection width is B = bw. • • Channel with w = B/w
- 12. 5.4 Data routingfunctions • Data routing achieved through message passing. • • There are same primitive routing functions on a network. • • These are shifting, rotation, permutation (one-to-one), Broadcast (one to all), multicast (many to many), personalized communication (one to many), shuffle, exchange, etc.
- 13. Perfect shuffle andexchange • Hwang page 78, fig 2.14
- 14. Hypercube routing functions •Hwang page 79, fig 1.15
- 15. Broadcast and Multicast •Broadcast and Multicast • Can be easily achieved on SIMD computers using broadcast bus. • Broadcast on a message passing system requires same other mechanism ( spanning tree, flooding)
- 16. Network throughput • Isdefined as total number of messages can be handled per unit time. • number of message can be in the network an once. • a hot spot can degrade performance of the entire network by causing congestion. • Low dimensional networks reduce contention because having a few high-bandwidth channels results in more resource sharing.
- 17. 5.5 Network performance •Network performance effected by some factors that summarized below; – 1 Functionality; How network supports data routing, interrupt handling, synchronization, request message/combing, and coherence. – 2 Network latency; This refer worse case time delay for a unit message to be transferred through network. – 3 Bandwidth; This refers to the maximum data transfer rate in terms of Mbytes/s transmitted through network. – 4 Hardware complexity; This refers to implementation cost such those for wires, switches, connectors, arbitration, and interface logic. – 5 Scalability; This refers to the ability of network to be modularly expandable with a scalable performance with increasing machine resource.
- 18. 5.6 Static networks •Static networks use direct links which are fixed once built. • This type of network is suitable where communication pattern are predictable or implementable with static connection.
- 19. Point to pointNetworks • All connections bi-drectional, (if not stated otherwise) • There is a fixed connection between nodes. • • n : Number of PEs in the network • V : connection for each PE • A or D : Maximum distance between PEs.
- 20. Binary Tree, ternarytree and quadtree • Brunnel, Figure 5.14 and 5.15, page 48.
- 23. Fat tree • Brunnel,fig 5.7, page 44.
- 24. Linear arrays • Hwangfig 2.16.a, page 81
- 25. Rings • Brunnel, Figure5.8, page 45.
- 26. Complete graph • Brunnel,Figure 5.9, page 45.
- 27. Grid and Torus •Brunnel, Figure 5.10, page 46.
- 28. AMP (A MinimumPath Systems) • Alan, page 30, fig 21
- 29. Hexagonal Grid • Brunnel,Figure 5.11, page 46. • Cube and hypercube • Brunnel, Figure 5.12 and 5.13, page 47.
- 32. Sytolic arrays • Hwang,fig 2.18.d, page 83
- 33. 5.7 Dynamic networks •For multipurpose and general purpose. • Communication pattern based on program demands. • Bus systems, multi stage inter connections, crossbar switch networks, • Cost depends on wiring, • Performance depend on network bandwith, data transfer rate, network latency and communication pattern supported.
- 34. Bus networks • cheap •simple • optimal (??) • common (networks, LANs) • Bandwidth remains fixed and divided by users. • One transaction at a time. – 1 Contention and, – 2 Time sharing bus.
- 35. • Von Neumannmodel connects all processor and memory modules and other modules. • Brunnel , (fig 5.2, page 39) • Hwang, ( fig 2.22, page 90) • Parallel reading possible same location but writing is not. • Parallel reading and writing to different memory locations are not possible. • Bust controller must control the access of the bus – 1 centralized and – 2 decentralized.
- 39. Switch networks • Dynamicconnections achieved with active elements. • Varying connection patterns can be engaged during run time of a program. • Crossbar switches, delta networks.
- 40. switch modules • A(axb) switch module has a input and b output • Generally a=b=2k for same k>=1 • Hwang table 2.3, page 91
- 41. Multi-stage networks • Hwangfig 2.23 page 92
- 42. Delta networks orOmega networks • Cost reduced from n2 (switch networkte ) to k??*n • Brunnel fig 5.4 and fig 5.5 page 41 • Hwang page 92, fig 2.224
- 45. Closs networks • Brunnel(fig 5.6, page 42,)
- 46. Crossbar networks • EachPE has n cross points. • total network with n PEs has n2 cross point. • any connection permutation can be specified • collision free. • Fully parallel information exchange can be performed. • Highest bandwidth and interconnection capability are provided by crossbar networks • Same module (memory) multiple request at the same time one request satisfied. • Brunnel, Fig 5.2 page 40. • Hwang figure 2.26, page 94.
- 49. 5.8 Comparison ofnetworks • Summary of static networks • Symmetry affect scalability and routing efficiency.It s fair to say that the total network cost increases with d (degree) and l (links). • Hwang, table 2.2, page 88 • The best solution for a particular system is a function of the system’s intended application, size, speed requirements, cost requirements, etc. • global communication are generally interested rather than local.
- 50. Summary of staticnetworks
- 52. Summary of dynamicnetworks • A summary comparison of dynamic networks is given figure below • (Hwang table, 2.4 page 95)