1. Nov. 2007Reconfiguration and VotingSlide 1 Fault-Tolerant Computing Hardware Design Methods

2. Nov. 2007Reconfiguration and VotingSlide 2 About This Presentation EditionReleasedRevised FirstNov. 2006Nov. 2007 This presentation has been prepared for the graduate course ECE 257A (Fault-Tolerant Computing) by Behrooz Parhami, Professor of Electrical and Computer Engineering at University of California, Santa Barbara. The material contained herein can be used freely in classroom teaching or any other educational setting. Unauthorized uses are prohibited. © Behrooz Parhami

3. Nov. 2007Reconfiguration and VotingSlide 3 Reconfiguration and Voting

4. Nov. 2007Reconfiguration and VotingSlide 4

5. Nov. 2007Reconfiguration and VotingSlide 5 Multilevel Model of Dependable Computing ComponentLogicServiceResultInformationSystem Level  Low-Level ImpairedMid-Level ImpairedHigh-Level ImpairedUnimpaired EntryLegend:DeviationRemedyTolerance IdealDefectiveFaultyErroneousMalfunctioningDegradedFailed

6. Nov. 2007Reconfiguration and VotingSlide 6 S D Requirements for Reconfiguration Overcoming the effect of interconnect failures requires the availability of multiple paths from each source to every possible destination A system consists of modular resources (processors, memory banks, disk storage,... ) and interconnects Redundant resources can mitigate the effect of module failures The main challenge of reconfiguration is dealing with interconnects Assumption: Module and interconnect failures are promptly diagnosed In graph-theoretic terms, we need “edge-disjoint” paths Existence of k edge-disjoint paths between two nodes provides the means for tolerating k – 1 link failures This particular interconnection scheme (torus) is 4-connected and tolerates 3 link/node losses without becoming disconnected

7. Nov. 2007Reconfiguration and VotingSlide 7 Reconfiguration via Programmable Connections Interconnection switch with 4 ports (horizontal lines) and 3 channels (vertical lines) Each port can be connected to 2 of the 3 channels 1 1 2 2 If each module port were connected to every channel, the maximum flexibility would result (leads to complex hardware & control, though) The challenge lies in using more limited connections effectively

8. Nov. 2007Reconfiguration and VotingSlide 8 Multiple-Bus Multiprocessors Failed units can be isolated from the buses No single bus failure can isolate a module from the rest of the system The vertical channels may be viewed as buses and the heavy dots as controllable bus connections, making this method applicable to fault-tolerant multiprocessors Read / Write data Write enable Q FF Write path Read path Connection FF Bus If we have extra buses, then failure in the bus connection logic can be tolerated by avoiding the particular bus For reliability analysis, lump the failure rate of reconfiguration logic with that of its associated bus

9. Nov. 2007Reconfiguration and VotingSlide 9 Row/Column Bypassing in 2D Arrays Similar mechanisms needed for northward links in columns and for the eastward and westward links in rows Row i Bypass row i 0 1 Question: What types of mechanisms do we need at the edges of this array to allow the row and column edge connections to be directed to the appropriate (nonbypassed) rows and columns? 0 1 1 0 0 1 1010 0 1 Bypass row i – 1

10. Nov. 2007Reconfiguration and VotingSlide 10 Choosing the Rows/Columns to Bypass In the adjacent diagram, can we choose up to 2 rows and 2 columns so that they contain all the faults? 0 1 2 3 4 5 6 7 0123456701234567 Spare columns Spare rows 0 2 3 6 Rows with faults 0 1 3 5 Columns with faults 7 0 3 1 7 Convert to graph problem (Kuo-Fuchs): Form bipartite graph, with nodes corresponding to faulty rows and columns Find a cover for the bipartite graph (set of nodes that touch every edge) Question: In a large array, with r spare rows and c spare columns, what is the smallest number of faults that cannot be reconfigured around with row/column bypassing?

11. Nov. 2007Reconfiguration and VotingSlide 11 Switch Modules in FPGAs Interconnection switch with 8 ports and four connection choices for each port: 0 – No connection 1 – Straight through 2 – Right turn 3 – Left turn 8 control bits (why?)

12. Nov. 2007Reconfiguration and VotingSlide 12 An Array Reconfiguration Scheme Three-state switch

13. Nov. 2007Reconfiguration and VotingSlide 13 One-Track and Two-Track Switching Schemes One-track switching modelTwo-track switching model Source: S.-Y. Kung, S.-N. Jean, C.-W. Chang, IEEE TC, Vol. 38, pp. 501-514, April 1989

14. Nov. 2007Reconfiguration and VotingSlide 14 Voting and Its Role in Dependable Computing John von Neumann, 1956: “Probabilistic Logic and Synthesis of Reliable Organisms from Unreliable Components” Voting schemes in these three contexts share some properties General schemes have been devised to cover all these instances Hardware voters for multichannel computation High performance (pipelined) Software voters for multiversion programming Imprecise results (approximate voting) Consistency of replicated data Weighted voting and weight assignment

15. Nov. 2007Reconfiguration and VotingSlide 15 What We Already Know About Voting Majority voting: Select the value that appears on at least  n/2  + 1 of the n inputs Number n of inputs is usually odd, but does not have to be Example: vote(1, 2, 3, 2, 2) = 2. x 1 x 2 x n Majority voter y Majority voters can be realized by means of comparators and muxes This design assumes that in the case of 3-way disagreement any one of the inputs can be chosen 2 3 1 x1x1 x2x2 x3x3 y Compare 0101 Disagree Can add logic to this voter so that it signals three-way disagreement

16. Nov. 2007Reconfiguration and VotingSlide 16 There is More to Voting than Simple Majority Plurality voting: Select the value that appears on the largest number of inputs Example: vote(1, 3, 2, 3, 4) = 3 What should we take as the result of vote(1.00, 3.00, 0.99, 3.00, 1.01)?. x 1 x 2 x n Plurality voter y It would be reasonable to take 1.00 as the result, because 3 inputs agree or approximately agree with 1.00, while only 2 agree with 3.00 Will discuss approximate voting and a number of other sophisticated voting schemes under software design topics Median voting: one way to deal with approximate values median(1.00, 3.00, 0.99, 3.00, 1.01) = 1.01 Median value is equal to the majority value when we have majority

17. Nov. 2007Reconfiguration and VotingSlide 17 Threshold Voting and Its Generalizations Simple threshold (m-out-of-n) voting: Output is 1 if at least m of the n inputs are 1s Majority voting is a special case of threshold voting: (  n/2  + 1)-out-of-n voting Agreement or quorum sets {x 1, x 2 }, {x 2, x 3 }, {x 3, x 1 } – same as 2-out-of-3 {x 1, x 2 }, {x 1, x 3, x 4 }, {x 2, x 3, x 4 } Weighted threshold (w-out-of-  v i ) voting: Output is 1 if  v i x i is w or more. x 1 x 2 x n m Threshold gate y. x 1 x 2 x n w Threshold gate y v 1 v 2 v n The 2 nd example above is weighted voting with v 1 = v 2 = 2, v 3 = v 4 = 1, and threshold w = 4 Agreement sets are more general than weighted voting in the sense of some agreement sets not being realizable as weighted voting

18. Nov. 2007Reconfiguration and VotingSlide 18 Usefulness of Weighted Threshold Voting Unequal weights allow us to take different levels of input reliabilities into account Zero weights can be used to disable or purge some of the inputs (combined switch-voter). x 1 x 2 x n w Threshold gate y v 1 v 2 v n Maximum-likelihood voting prob{x 1 correct} = 0.9 prob{x 2 correct} = 0.9 prob{x 3 correct} = 0.8 prob{x 4 correct} = 0.7 prob{x 5 correct} = 0.6 Assume x 1 = x 3 = a, x 2 = x 4 = x 5 = b prob{a correct} = 0.9  0.1  0.8  0.3  0.4 = 0.00864 prob{b correct} = 0.1  0.9  0.2  0.7  0.6 = 0.00756 Max-likelihood voting can be implemented in hardware via table lookup or approximated by weighted threshold voting (otherwise, we need software)

19. Nov. 2007Reconfiguration and VotingSlide 19 Implementing Weighted Threshold Voters Example: Implement a 4-input threshold voter with v 1 = v 2 = 2, v 3 = v 4 = 1, and threshold w = 4 Strategy 1: If weights are small integers, fan-out each input an appropriate number of times and use a simple threshold voter x 1 x 2 x 3 x 4 4 Threshold gate y Strategy 3: Convert the problem to agreement sets (discussed next) Strategy 2: Use table lookup based on comparison results x 1 =x 2 x 1 =x 3 x 2 =x 3 x 3 =x 4 Result 1 x x x x 1 0 0 0 1 Error 0 1 0 1 x 1 0 0 1 1 x 2 0 x x 0 Error Is this table complete? Why, or why not?

20. Nov. 2007Reconfiguration and VotingSlide 20 Implementing Agreement-Set Voters Example: Implement a voter corresponding to the agreement sets {x 1, x 2 }, {x 1, x 3, x 4 }, {x 2, x 3, x 4 } Strategy 1: Implement as weighted threshold voter, if possible Strategy 2: Implement directly Find a minimal set of comparators that determine the agreement set Complete this design by producing the “no agreement” signal x1=x2x1=x2 x1=x3x1=x3 x3=x4x3=x4 x2=x3x2=x3 x2x2 x1x1 0 1 y

21. Nov. 2007Reconfiguration and VotingSlide 21 Implementing Weighted Plurality Voters Inputs: Data, vote-weight pairs Output: Data with maximal vote and its associated vote tally Source: B. Parhami, IEEE Trans. Reliability, Vol. 40, No. 3, pp. 380-394, August 1991 Sort by data Com- bine votes Select max vote  5, 1   4, 2   5, 3   7, 2   4, 1   5, 4   7, 2   5, 1   5, 3   4, 2   4, 1   7, 2   5, 4   5, 0   4, 3   4, 0  Phase 1Phase 2Phase 3 SorterCombinerSelector

22. Nov. 2007Reconfiguration and VotingSlide 22 5-Sorter5-Combiner5-Selector Details of a Sorting-Based Plurality Voter The first two phases (sorting and combining) can be merged, producing a 2-phase design – fewer, more complex cells (lead to tradeoff) Source: B. Parhami, IEEE Trans. Reliability, Vol. 40, No. 3, pp. 380-394, August 1991 Stages of delay 1 2 3 4 5 6 7 8 9 10 11 12 13  5, 1   4, 2   5, 3   7, 2   4, 1   5, 3   7, 2   5, 1   7, 2   4, 2   5, 3  5, 4   5, 0   4, 3   4, 0   5, 4  4, 3 

23. Nov. 2007Reconfiguration and VotingSlide 23 Unification of Terminology for Voting Participant... Fuser OutcomeOpinion (not voter) Weight (not vote) Social choice research Data fusion research Need for unambiguous terminology Opinion or outcome is a subset of the set of alternatives Dependable computing research