Trade Offs in the Design of a Router with Both Guaranteed and Best-Effort Services for Networks on Chip

Transcription

1 Trade Offs in the Design of a Router with Both Guaranteed and BestEffort Services for Networks on Chip E. Rijpkema, K. Goossens, A. R dulescu, J. Dielissen, J. van Meerbergen, P. Wielage, and E. Waterlander

2 why NetworksonChip problems observed for SoC design deep sub micron design complexity wire cost increasing # of IP blocks timing closure increasing dynamism decouple computation from communication application presentation session transport network data link physical application demands network independent services network dependent hardware technology IP IP IP R R R router network

3 why NetworksonChip problems observed for SoC design deep sub micron design complexity key: decouple computation from communication application presentation session transport network data link physical application demands network independent services network dependent hardware technology IP IP IP R R R router network 3

4 outline services combined router architecture guaranteed throughput router architecture besteffort router architecture router prototype conclusions 4

5 services I we need a network that is predictable cost effective requirements application demands guarantees services efficiency hardware technology constraints build guarantees on top of guarantees efficient network is efficient at every layer 5

6 services II timeless guarantees guaranteed data integrity guaranteed data delivery guaranteed inorder delivery time related guarantees (over bounded time interval) guaranteed throughput guaranteed latency besteffort service (BE) guaranteed throughput service (GT) 6

7 guarantees vs. besteffort GT requires dimensioning for guaranteed throughput BE requires dimensioning for average throughput r wc r wc 3 4 guaranteed throughput (bounded interval) ravg time time time guaranteed delivery combination is beneficial 7

8 BE & GT combined architecture conceptually, two disjoint routers a router with GT service class a router with BE service class BE router r wc programming GT router 3 4 priority/arbitration to obtain an efficient combination routers must have similar architectures 8

9 buffering strategy output queuing highest cost highest performance N N input queuing lowest cost lowest performance N X N virtual output queuing moderate cost high performance N X N preferred solution 9

10 contention links in network are shared resources contention occurs when multiple data request same link at same time GT and BE resolve contention differently 0

11 guaranteed throughput to guarantee latency or bandwidth over finite interval cannot drop data must bound contention ratebased scheduling has high buffer costs (deep fifos/output queuing) deadlinebased scheduling even higher buffer costs (deep priority queues) contentionfree routing low buffer costs (shallow fifos)

12 contentionfree routing I scheduling packet injection in network to avoid contention in space: disjoint paths as in pure circuit switching in time: timedivision multiplexing as with a statically scheduled bus in time and space: our solution

13 contentionfree routing II divide time in slots block block block 3 time slot a block is amount of data that fits in a slot block entering router in slot n enters next in slot n+ n n+ n+3 n+ matches with input queuing N X 3 N

14 contentionfree routing III routers have tables that store contention resolution & routing information allow distributed programming S S S S small blocks Æ low buffering cost small slots Æ low latency Æ throughput guarantee on smaller period 4

15 besteffort architecture to ensure high resource utilization statistical multiplexing packetswitching but implement BE service class packetswitching network flow control (routing mode) contention resolution 5

16 packets and flits packet = header + payload H payload packet might be transmitted in smaller parts called flits flit flit flit 3 flit 4 flits divide time in iterations and must be scheduled flit flit flit 3 flit 4 time smaller flit size Æ higher scheduling rate Æ lower latency Æ less storage 6

17 network flow control (routing mode) performance/cost network flow control store and forward routing first receive whole packet then transmit whole packet virtual cutthrough routing send flit immediately if next router can receive entire packet wormhole routing send flit immediately if next router can receive that flit per router latency storage packet packet flit packet flit flit 7

18 contention resolution queuing at input Æ set paths from inputs to outputs router has switch bipartite graph matching 3 X 3 algorithm must be fair have low complexity (to schedule at flit rate) approximation of maximal matching 8

19 combining GT and BE links must be shared by GT and BE traffic grain size of interleaving must match block size = flit size smallest value for this is given by implementation minimize scheduler latency Æ L maximize data path speed Æ F flit size = block size = F L 9

20 router prototype snapshot of current prototype router: control input queuing arity 5 3 bits wide words 8 flits deep BE queues 56 slots 0.5 mm CMOS 500 MHz data path 66 MHz control path flit size is 3 words throughput per link: 500MHz 3bits = 6Gb/s 0

21 conclusions for NoCs, guaranteed services are essential demonstrated the useful combination of: BE service class Æ timeless guarantees GT service class Æ BE + time related guarantees made tradeoffs to come to efficient combined router proved feasibility with router prototype

22 router prototype snapshot of current prototype router: 5 input and 5 output ports (arity 5) 0.5 mm CMOS 500 MHz data path, 66 MHz control path flit size of 3 words of 3 bits 500x3 = 6 Gb/s throughput per link 56 slots & 5x flit fifos for guaranteedthroughput traffic 6x8 flit fifos for besteffort traffic

23 control control 3