Concepts for Robust NoC Communication

Transcription

1 oncepts for Robust o ommunication Martin Radetzki Department of mbedded ystems ngineering Institute of omputer Architecture and omputer ngineering Universität tuttgart martin.radetzki@informatik.uni-stuttgart.de 9/21/2009

2 Future hallenge: Reliability ithin a decade we will see 100 billion transistor chips. That is the good news. The bad news is that 20 billion of those transistors will fail in manufacture and a further 10 billion will fail in the first year of operation. [teve Furber, Living ith Failure, Proc. uropean Test ymposium, pp. 4-8, 2006] oncepts for Robust o ommunication 2

3 Reliability Issues ingle vent ffects e.g. ion, neutron impact causing transient faults Variability manufacturing-induced variations, e.g. in wire thickness, transistor parameters causing permanent and intermittent faults that must be tolerated for acceptable yield tress & Ageing e.g.lectromigration, HI, BTI causing permanent faults that appear during system operation nvironmental Impacts make everything worse... e.g. accellerated ageing under high temperatures, high levels of radiation on space missions oncepts for Robust o ommunication 3

4 Failure Fault - rror Terminology Failure mechanism Physical cause Fault (model) Formal description rror ffect on information xamples ingle-event effect Fabrication defect hip ageing Transient fault Permanent fault (e.g. stuck-at) Delay fault oft error Recurring data corruption o: Packet loss oncepts for Robust o ommunication 4

5 teps for Fault-Tolerance Offline Online xamples on various layers Testing rror detection DL: yndrome 0 L: Packet corruption TL: Packet loss Diagnosis (model) Fault location Fault classification orrection Methods depend on the class of fault DL: Analyze syndrome L: Observe the regular traffic TL: Inject test packets DL: Toggle erroneous bits L: Adapt routing to bypass fault TL: Re-send packet (ARQ) oncepts for Robust o ommunication 5

6 Outline Fault Models and Fault Diagnosis Fault-Adaptive Routing xperimental (imulation) Results oncepts for Robust o ommunication 6

7 Distributed rror Detection and Diagnosis e.g. [Grecu, Pande 2006] Parity checks at switch inputs and outputs Output error => diagnosis of switch fault Input error => diagnosis of switch fault Drop and retransmit packet in case of error Does not help against permanent faults oncepts for Robust o ommunication 7

8 Traditional Fault Models witch faults: the complete switch is assumed to be unavailable Link faults: refers to a link between two switches (and/or a directional output of a switch) oncepts for Robust o ommunication 8

9 Packet Throughput in aturated etwork Packets Maximaldurchsatz per [Pakete/Ressourcenknoten/Zyklus] resource and cycle fault duration t=1 t=4 t=16 t=64 t=256 permanent 8x8 mesh witch-fehlerrate [%] witch failure rate [%] witch faults: Throughput breaks down even at low failure rates oncepts for Robust o ommunication 9

10 tructural Fault Model One multiplexer, part of a crossbar switch: oncepts for Robust o ommunication 10

11 Fault Diagnosis of Individual rossbar onnections One approach [Raik, Ubar 2007]: ending directed test patterns Assumption: Links are fault-free Diagnostic access at the network boundaries entral control, not distributed (scalability?) Regular o operation needs to be interrupted (offline) oncepts for Robust o ommunication 11

12 Proposed Diagnostic Infrastructure ostrum switch Input register R Output driver R R Packet format extended by a R checksum: V H x y Payload PayloadR R R R Input check Router Input register R R R Output check R R R Input check bits R check on all inputs and outputs - Input: Detecting Inter-witch-Faults - Output: Dectecting Intra-witch-Faults L L oncepts for Robust o ommunication 12

13 Fault Matrix Model of individual crossbar connection faults Implemented on chip to store online diagnosis results crc V 000 no fault crc V 001 transient crc OK all patterns crc V 010 intermittent Fault matrix testing send test patterns crc 111 permanent Fault state machine oncepts for Robust o ommunication 13

14 Outline Fault Models and Fault Diagnosis Fault-Adaptive Routing xperimental (imulation) Results oncepts for Robust o ommunication 14

15 Deflection Routing Packets are routed with a preferred direction, e.g. following XY route If that is not possible, the packet is deflected into another direction o blocking o buffers Free of deadlocks on-minimal paths possible Livelock avoidance through priority for old packets oncepts for Robust o ommunication 15

16 Implementing Deflection in witches (1) Routing in preferred direction (here: x towards destination) (2) Routing in other direction towards destination (3) Routing in other direction away from destination (4) Reflection to the packet source of the previous hop oncepts for Robust o ommunication 16

17 Routing Adaptation In order to find feasible routes in case of faults, all incoming packets have to be considered as a whole: packet P1 P2 P3 P4 priority max (4) high (3) low (2) min (1) from preferred route,, dest. dest. P1 veer away,, Reflection P4 P2 dest. P1 P3 P2 dest. P3 P4 oncepts for Robust o ommunication 17 17

18 ost Driven Routing (1) Assign a cost to a routing decision for one packet: Reflects the progress of the packet ost is weighted by priority packet P1 P2 P3 P4 priority max (4) high (3) low (2) min (1) from preferred route,, veer away,, reflection c ij if fij > 0012 avl j, 2 elsif i = j, = 0 elsif i j reduces P i.addr, 1 else connection unavailable reflection approach destination veer away oncepts for Robust o ommunication 18

19 ost Driven Routing (2) In order to consider all packets at once, build the sum of weighted costs c π = ri π i c( π ) i good = ( ) () c( π ) bad = 9 Objective: Find routing permutation π opt with minimal cost sum in a combinational way onstruct permutation tree Bottom-up traversal of the permutation tree Use minimum operations to discard bad routes early oncepts for Robust o ommunication 19

20 Routing Implementation xample: input output min min min min min min min min min min min min min min min min min oncepts for Robust o ommunication 20

21 witch Area and Timing Area [μm 2 ] Max. operating frequency [MHz] oncepts for Robust o ommunication 21

22 Outline Fault Models and Fault Diagnosis Fault-Adaptive Routing xperimental (imulation) Results oncepts for Robust o ommunication 22

23 imulation etup 8x8 mesh with uniformly distributed and transposed data traffic Incremental injection of faults according to a fixed, randomly selected fault pattern Measurement of latency and throughput omparison of three models: Diagnosis and adaptation as presented Automated Repeat Request (ARQ) at receiving end Automated Repeat Request (ARQ) on switch level oncepts for Robust o ommunication 23

24 Fault-Free ase lassic ost based Hop count Injection rate [Packets per ore and ycle] Average network latency reduced by ~12% oncepts for Robust o ommunication 24

25 Packet Throughput with ingle rossbar Faults 0.35 Packets Maximaldurchsatz per [Pakete/Ressourcenknoten/Zyklus] resource and cycle t=1 t=4 t=16 t=64 t=256 perm witch-fehlerrate Failure rate [%] 8x8 mesh Minor throughput reduction even at large failure rates Fault duration has no significant impact oncepts for Robust o ommunication 25

26 Alternative Methods orrect packets : injected packets Diagnosis und Adaptation ARQ with Uniform Traffic ARQ with Transposed Traffic umber of faults Routing around faults is better than accepting faults and performing retransmission oncepts for Robust o ommunication 26

27 onclusions Routing method using information on o permanent fault status Distributed implementation (scalable) an be implemented with acceptable cost urrent research: combined with retransmission to handle transient faults Future work: research optimal combination of techniques on different layers oncepts for Robust o ommunication 27

28 Hierarchical Approach to Fault Tolerance Fault Models Monitoring Diagnosis Reconfiguration Transport etwork Distributed Online- Algorithms, Regular tructures Packets rror Detection ongestion Packet Loss Packet-Fault Packet Loss rror rror orrection Re- Routing Data Link Bandwidth Fault Rate Fault ordstream Bandwidth- Request ncoding (Logic) rror Detection Bit-Faults rror orrection Bit-Faults Physical Bitstream Fault Rate lectr. Param. Temperature Defect Hardening- uggestions ncoding (electrical), Drivers, Monitoring rror Detection Delay Faults Voltage-, urrent-, Temperature-Monitoring rror orrection Adjust Voltage, urrent Use multiple lines pare-lines oncepts for Robust o ommunication 28

29 Parameter Interaction Performance Delay faults orrection overhead witching activity Frequency upply voltage rror rate lectromigration BTI, HI,... Temperature L-oupling Radiation (U) oise / R Power oncepts for Robust o ommunication 29