Symbolic Synthesis of Masking Fault-Tolerant Distributed Programs

Symbolic Synthesis ofSymbolic Synthesis ofMasking Fault-Tolerant Masking Fault-Tolerant Distributed ProgramsDistributed Programs

Borzoo BonakdarpourBorzoo BonakdarpourWorkshop APRETAFWorkshop APRETAFJanuary 23, 2009January 23, 2009

Joint work with Sandeep Kulkarni

2

Motivation

• The most important goal of formal methods is achieving correctness in computing systems (programs).– Correct-by-verification

• A program is built manually

• The correctness of the program is verified by a model checker or a theorem prover.

– Correct-by-construction• A program is constructed so that its correctness is guaranteed.

Manual Design

Verification

3

Motivation

• Automated synthesis from temporal logic specifications– Pros:

• Ability to start from a null program

• Capability to handle highly expressive specifications

– Cons:• Highly complex decision procedures

• Limited to no reusability

• Automated program revision– An existing program is revised with respect to a property

4

• Question:

– Is it possible to revise the program automatically such that it satisfies the failed property while ensuring satisfaction of existing properties?

• bugs

• incomplete specification

• change of environment

CounterexampleCounterexamplePropertyProperty

ModelCheckerModel

Checker

ProgramProgram

Motivation

5

Revised Revised programprogramPropertyProperty

RevisionRevisionAlgorithmAlgorithmRevisionRevision

AlgorithmAlgorithm

ProgramProgram

Motivation

6

Motivation

A one-lane bridge is controlled by two traffic signals at the two ends of the bridge.

(sig1 = G) (1 ≤ x1 ≤ 10) sig1 := Y;

[]

(1(1))

(2(2))

(sig1 = Y) (1 ≤ y1 ≤ 2) sig1 := R ;

[]

(sig2 = R) (z1 ≤ 1) sig2 := G ;

[]

{y1}

{z1}

{x2}

Controller Program:Controller Program:

((sig1 = G) (x1 ≤ 10)) ((sig1 = Y ) (y1 ≤ 2)) ((sig1 = R) (z2 ≤ 1)) wait;

SPECSPECbtbt = = {({(00, , 11) | ) | sigsig11 ( (11) ≠ ) ≠ RR sigsig22 ( (11) ≠ ) ≠ RR}}

7

true skip; {z2}

Traffic Controller Fault Action: Traffic Controller Fault Action:

(1(1))

(2(2))

1. (sig1 = sig2 = R) (z1 ≤ 1) (z2 > 1)

2. (sig1 = sig2 = R) (z1 ≤ 1) (z2 = 0)

3. (sig1 = G) (sig2 = R) (z1 ≤ 1) (z2 = 0)

4. (sig1 = G) (sig2 =G) (z1 ≤ 1) (z2 = 0)

Motivation

8

19821982 20002000

19861986

19891989

Vardi and Wolper introduce

automata-theoretic verfication and synthesis

Vardi and Wolper introduce

automata-theoretic verfication and synthesis

2005200519921992

Alur and Henzinger propose verification and synthesis of

real-time systems

Alur and Henzinger propose verification and synthesis of

real-time systems

Clarke, Emerson, Sifakis, and Queille invent model checking

Clarke, Emerson, Sifakis, and Queille invent model checking

19811981

Emerson and Clarke propose synthesis from CTL properties

Emerson and Clarke propose synthesis from CTL properties

McMilan et al. intorduce BDD-based model

checking (1020 reachable states) and find bugs in

IEEE futurebus+

McMilan et al. intorduce BDD-based model

checking (1020 reachable states) and find bugs in

IEEE futurebus+

19931993

Wonham and Ramadge introduce

controller synthesis

Wonham and Ramadge introduce

controller synthesis

Intel reports bug in floating point operations in

Pentium processors

Intel reports bug in floating point operations in

Pentium processors

19941994

Clarke and Grumberg introduce counterexample guided abstraction-refinement (CEGAR), 101000 reachable states

Clarke and Grumberg introduce counterexample guided abstraction-refinement (CEGAR), 101000 reachable states

19991999

Kulkarni and Arora introduce automated addition of fault-tolerance to fault-intolerant programs

Kulkarni and Arora introduce automated addition of fault-tolerance to fault-intolerant programs

Biere and Clarke invent SAT-based model

checking (10500 reachable states)

Biere and Clarke invent SAT-based model

checking (10500 reachable states)

Bonakdarpour, Kulkarni, and Ebnenasir, and, Jobstmann and Bloem independently introduce program revision (repair) techniques

Bonakdarpour, Kulkarni, and Ebnenasir, and, Jobstmann and Bloem independently introduce program revision (repair) techniques

20072007

Bonakdarpour and Kulkarni synthesize distributed programs of size 1050

Bonakdarpour and Kulkarni synthesize distributed programs of size 1050

Clarke, Emerson, Sifakis, and Queille invent model checking

Clarke, Emerson, Sifakis, and Queille invent model checking

McMilan et al. intorduce BDD-based model

checking (1020 reachable states) and find bugs in

IEEE futurebus+

McMilan et al. intorduce BDD-based model

checking (1020 reachable states) and find bugs in

IEEE futurebus+

Mohamed Gouda:Mohamed Gouda:

When does your “12 years” end?!When does your “12 years” end?!

Mohamed Gouda:Mohamed Gouda:

When does your “12 years” end?!When does your “12 years” end?!20082008

Motivation

9

InvariantInvariantInvariantInvariant

f f

f

Fault-SpanFault-Span

State spaceState space

p

p

p

p

p

pp

pp

p

f

f

The Synthesis Problem

10

The Issue of Distribution

• Modeling distributed programs:– A program consists of a set of processes. Each process p is

specified by:• A set Vp of variables,

• A set Tp of transitions,

• A set Rp Vp of variables that p is allowed to read,

• A set Wp Rp of variable that p is allowed to write.

• Write restrictionsa = 0b = 1

a = 0b = 1

a = 1b = 1

a = 1b = 1

a Wp

a = 0b = 1

a = 0b = 1

a = 1b = 1

a = 1b = 1

a Wp

Such transitions cannot be executed by process p.

11

• Read restrictions

a = 1 b = 0

a = 1 b = 0

a = 0 b = 0

a = 0 b = 0

b Rp

a = 1 b = 1

a = 1 b = 1

a = 0b = 1

a = 0b = 1

a = 1 b = 0

a = 1 b = 0

a = 0 b = 0

a = 0 b = 0

b Rp

a = 1 b = 1

a = 1 b = 1

a = 0b = 1

a = 0b = 1

– Such set of transitions form a group.– Addition and removal of any transition must

occur along with its entire group.

The Issue of Distribution

12

What Is DifficultAbout Program Revision?

• Space complexity– The state explosion problem

• Time complexity– NP-completeness

• Identifying the complexity hierarchy of the problem• The need for designing efficient heuristics• Proofs are often helpful in identifying bottlenecks of the problem

The combination of the above complexitiesThe combination of the above complexitiesis the worst nightmare!is the worst nightmare!

13

Daniel MosDaniel Moséé::

As that wise man said “bridging theAs that wise man said “bridging the

gap between theory and practice isgap between theory and practice is

easier in theory than in practice!”easier in theory than in practice!”

Daniel MosDaniel Moséé::

As that wise man said “bridging theAs that wise man said “bridging the

gap between theory and practice isgap between theory and practice is

easier in theory than in practice!”easier in theory than in practice!”

What Is DifficultAbout Program Revision?

14

The Byzantine Agreement Problem

Decision

d.g {0, 1}

(d.j = ) ( f.j = false) d.j := d.g

(d.j ) ( f.j = false) f.j := true

d.j

d.k {0, 1, }

d.l

Decision

f.j

f.k {false, true}

f.l

Final?

GENERAL

NON-GENERALS

Program:

15

The Byzantine Agreement Problem

Byzantine?

b.g {false, true}

b.j

b.k {false, true}

b.l

Byzantine?

(b.j , b.k , b.l , b.g = false) b.j := true

(b.j := true) d.j := 0|1Faults:

16

• Experimental results with enumerative (explicit) state space (the tool FTSyn)– Byzantine agreement - 3 processes

• 6912 states

• Time: 10s

– Byzantine agreement - 4 processes • 82944 states

• Time: 15m

– Byzantine agreement - 5 processes• 995328 states

• Out of memory!

What Is DifficultAbout Program Revision?

17

Polynomial -Time Heuristics

Identify the state predicatems from where faults

alone violate the safety;

S := S ms

ms

SPEC

ff

f f

18

Identify the state predicatems from where faults

alone violate the safety;

S := S ms

Polynomial -Time Heuristics

Re-compute theRe-compute thefault-spanfault-span

ms

Inv.Inv.

Fault-Span

f

f

f

f

p

p

f

p

BDD frontier = Invariant;BDD current = mgr -> bddZero();BDD FaultSpan = Invariant;

while (FaultSpan != current){ current = FaultSpan; BDD image = frontier * (P + F); // -FaulSpan frontier = Unprime(image); FaultSpan = current + frontier; }

BDD frontier = Invariant;BDD current = mgr -> bddZero();BDD FaultSpan = Invariant;

while (FaultSpan != current){ current = FaultSpan; BDD image = frontier * (P + F); // -FaulSpan frontier = Unprime(image); FaultSpan = current + frontier; }

19

Polynomial -Time Heuristics

Yes

No

Identify the state predicatems from where faults

alone violate the safety;

S := S ms

Re-compute theRe-compute thefault-spanfault-span

Identify transitions in the fault-intolerant programIdentify transitions in the fault-intolerant programthat may be included in the fault-tolerant programthat may be included in the fault-tolerant program

Fixpoint?Fixpoint?

Resolve deadlock statesResolve deadlock states

InvariantInvariant

Fault-Span

ffpp

p

Re-computing state predicates or transitions predicates do not occur often in model checking, but it does happen quite often during synthesis.

Re-computing state predicates or transitions predicates do not occur often in model checking, but it does happen quite often during synthesis.

s0

s1

20

Experimental Results

• Polynomial-time sound BDD-based heuristics– The tool SYCRAFT (http://www.cse.msu.edu/~borzoo/sycraft)– C++– CuDD (Colorado University Decision Diagram Package)

• Platform– Dedicated PC – 2.2GHz AMD Opteron processor– 1.2GB RAM

21

Experimental Results

• Goal:– Identifying various bottlenecks of our synthesis heuristics

• Fault-span generation

• Deadlock resolution– Adding recovery– State elimination

• Cycle detection and resolution

• Memory usage

• Total synthesis time

22

Experimental Results

23

Experimental Results

24

Performance of synthesizing the

Byzantine Byzantine agreementagreement

program

Experimental Results

25

• Observations

– 1050 reachable states– State elimination (deadlock resolution) is the most serious

bottleneck– We run of time before we run out of space– Size of state space by itself is not a bottleneck

Experimental Results

26

-----------------------------------------------------------------------------------------------------UNCHANGED ACTIONS:UNCHANGED ACTIONS:-----------------------------------------------------------------------------------------------------1- (d.j==2) & !(f.j==1) & !(b.j==1) (d.j:=dg)-----------------------------------------------------------------------------------------------------REVISED ACTIONS:REVISED ACTIONS:-----------------------------------------------------------------------------------------------------2- (b.j==0) & (d.j==1) & (d.k==1) & (f.j==0) (f.j:=1)3- (b.j==0) & (d.j==0) & (d.l==0) & (f.j==0) (f.j:=1)4- (b.j==0) & (d.j==0) & (d.k==0) & (f.j==0) (f.j:=1)5- (b.j==0) & (d.j==1) & (d.l==1) & (f.j==0) (f.j:=1)-----------------------------------------------------------------------------------------------------NEW RECOVERY ACTIONS:NEW RECOVERY ACTIONS:-----------------------------------------------------------------------------------------------------6- (b.j==0) & (d.j==0) & (d.l==1) & (d.k==1) & (f.j==0) (d.j:=1)7- (b.j==0) & (d.j==1) & (d.l==0) & (d.k==0) & (f.j==0) (d.j:=0)8- (b.j==0) & (d.j==0) & (d.l==1) & (d.k==1) & (f.j==0) (d.j:=1), (f.j:=1)9- (b.j==0) & (d.j==1) & (d.l==0) & (d.k==0) & (f.j==0) (d.j:=0), (f.j:=1)------------------------------------------------------------------------------------------

Experimental Results

27

The effect of exploiting

human knowledgehuman knowledge(Each non-general process is allowed to finalize its decision if no two non-generals are undecided.)

Experimental Results

28

Experimental Results

Performance of synthesizing

token ringtoken ring mutual exclusion with

multi-step recovery

29

Experimental Results

Multi-step vs. single-step single-step

recoveryrecovery for synthesizing

token ring mutual exclusion

30November 7, 2008 Doctoral Dissertation Defense

Open Problems

• Exploiting techniques from model checking

– State space generation (e.g., clustering and partitioning)

– Symmetry reduction

– Counter-example guided abstraction-refinement (CEGAR)

– SMT/QBF-based methods

– Distributed/parallel techniques

31November 7, 2008 Doctoral Dissertation Defense

Open Problems

• Multidisciplinary research problems

– Revising hybrid systems

– Synthesizing programs with multiple concerns (e.g., security, communication, real-time, fault-tolerance, distribution) in epistemic logic

– Program synthesis using graph mining and machine learning techniques

– Biologically-inspired revision/synthesis techniques

THANK YOU!