Lecture 16 Interconnection Networks Embedded Computing Systems Mikko
Lecture 16: Interconnection Networks Embedded Computing Systems Mikko Lipasti, adapted from M. Schulte Based on slides and textbook from Wayne Wolf High Performance Embedded Computing © 2007 Elsevier
Topics n n Interconnection Terms and Metrics Interconnection Models Routing and Flow Control Network-on-Chips © 2006 Elsevier
Interconnection networks n n n Client: sender or receiver on a network. Port: connection to a network on a client Link: connection between two clients (full/half duplex) Topology: organization of network links. Network metrics: q q n Throughput. Latency. Energy consumption. Area (silicon or metal). Quality-of-service (Qo. S) is important for multimedia applications. © 2006 Elsevier
Interconnection network models n n n Source <- line -> termination. Throughput T, latency D. Link transmission energy Eb. Physical length L. Total link area A. Traffic models often use Poisson distribution q q n P(X = x) = m xe-m /(x)! E(x) = m, Var(x) = m. x =0, 1, 2, … Streaming data is produced periodically with rate s and burstiness r © 2006 Elsevier
Network topologies n Major choices for network topologies include. q q q Bus: common connection between a set of senders and receivers Crossbar: fully connected network from every input port to every output port. Buffered crossbar: add queues to a crossbar to enable multiple sources to share crossbar input Mesh: network in which every node is connected to all of its neighbors. Application-specific: topology is matched to the characteristics of the application. © 2006 Elsevier
Bus network n Throughput: q q n Advantages: q q q n T 1 = P/(1+C). – single word Tb = P*(n/(n + C)) – n word block Well-understood. Easy to program. Many standards. Disadvantages: q q q Contention. Significant capacitive load. Do not scale well. © 2006 Elsevier
Crossbar network n n Fully connected network Advantages: q q n No contention. Simple design Low latency Broadcast. Disadvantages: q q Expensive Not feasible for large numbers of ports. © 2006 Elsevier
Buffered crossbar network n n Add queues shared by multiple sources Advantages: q q n Disadvantages: q n Smaller than crossbar. Can achieve high utilization. Requires scheduling. Clos networks q Connect multiple crossbars together in stages © 2006 Elsevier Xbar
Mesh network n n Every node connected to all of its neighbors Advantages: q q n Well-understood. Regular architecture. Disadvantages: q q Poor utilization Variable latency. © 2006 Elsevier
Application-specific. network n n Topology is specific for application(s) Advantages: q q n Higher utilization. Lower power. Disadvantages: q q Must be designed. Must carefully allocate data. © 2006 Elsevier
Network topology questions n n What type of network topology would you chose if you were building a 1, 000 node system? Why would you use a buffered crossbar network instead of a regular crossbar network? What advantages and disadvantages does a 2 D mesh have compared to a 3 D mesh? What types of systems would you expect to use application-specific networks? © 2006 Elsevier
Routing and flow control n Routing determines paths followed by packets. q q n Connection-oriented or connectionless. Wormhole routing divides packets into flits and header flit determines route for remaining flits Virtual cut-through ensures entire path is available before starting transmission. Store-and-forward routing stores packets inside network. Flow control allocates links and buffers as packets move through the network. q Virtual channel flow control treats flits in different virtual channels differently. © 2006 Elsevier
Networks-on-chips n Impact characteristics of MPSo. C: q q q n No. Cs may not have to interoperate with other networks. q n Energy. Performance. Cost. No. Cs have to connect to existing IP, which may influence interoperability. Qo. S is an important design goal. © 2006 Elsevier
Nostrum n n n Mesh network---switch connects to four nearest neighbors and local “resource” Each switch has queue at each input. Selection logic determines order in which packets are sent to output links. [Kum 02] © 2002 IEEE Computer Society © 2006 Elsevier
Scalable, Programmable, Integrated Network (SPIN) n Scalable network based on fat-tree. q n Bandwidth of links is larger toward root of tree. All routing nodes use the same routing function. q Message goes up the tree until a common ancestor reached © 2006 Elsevier
Routing nodes in SPIN n Packet consists of 32 -bit words q q q n n One word header Variable word packet One word checksum trailer Network utilizes input queues and partial crossbars Outputs share buffers q Help with contention © 2006 Elsevier
Ye et al. energy model n n n Assume: energy per packet is independent of data or packet address. Histogram captures distribution of path lengths. Energy consumption of a class of packet: q q q M = maximum number of hops. h = number of hops. N(h) = value of hth histogram bucket. L = number of flits per packet. Eflit = energy per flit. © 2006 Elsevier
Ye et al. energy model n n Longer packets corresponds to longer block sizes Larger packets n Decrease cache misses but increase the miss penalty n Decrease number of packets but increase hops per packet n Decrease cache and memory energy, but increase network energy © 2006 Elsevier
Goossens et al. No. C methodology Geared towards n Application-specific So. Cs n Qo. S-intensive apps n Network dimensioning – determine size of network and buffers n No. C topology – determine connection between elements n No. C configuration – set register values that control flow through the network n © 2006 Elsevier
QNo. C n n Designed to support Qo. S. Two-dimensional mesh, wormhole routing. q n Four different types of service. q q q n Fixed x-y routing algorithm. Each service level has its own buffers. Next-buffer-state table records number of slots for each output in each class. Transmissions based on next stage, service levels, and round-robin ordering. Can be customized for specific application. © 2006 Elsevier
QNo. C Design Methodology © 2006 Elsevier
Xpipes and Net. Chip n n Xpipes is a library of soft IP macros for network switches and links. Net. Chip generates custom No. C designs using xpipes components. © 2006 Elsevier
Xu et al. H. 264 network design n Designed No. C for H. 264 decoder. Process -> PE mapping was given. Compared RAW mesh to application-specific networks. [Xu 06] © 2006 ACM Press © 2006 Elsevier
Application-specific network for H. 264 © 2006 Elsevier [Xu 06] © 2006 ACM Press
RAW/application-specific network comparison © 2006 Elsevier [Xu 06] © 2006 ACM Press
- Slides: 25