dependability and real-time

Undergraduate course on Real-time SystemsLinköping University

TDDD07 Real-time SystemsLecture 7: Dependability &

Fault tolerance

Simin Nadjm-Tehrani

Real-time Systems Laboratory Department of Computer and Information Science

Linköping university

56 pagesAutumn 2009


2 of 56Autumn 2009

Dependability and real-time

• If a system is to produce results within time constraints, it needs to produce results at all!

• How to justify and measure how well computer systems do their jobs?


3 of 56Autumn 2009

Dependable systems

• How do things go wrong and why?• What can we do about it?

– This lecture: Basic notions of dependability and replication in fault-tolerant systems

– Next lecture: Designing Dependable Real-time systems


4 of 56Autumn 2009

Early computer systems

• 1944: Real-time computer system in the Whirlwind project at MIT, used in a military air traffic control system 1951

• Short life of vaccum tubes gave mean time to failure of 20 minutes


5 of 56Autumn 2009

Engineers: Fool me once, shame on you – fool me twice, shame on me


6 of 56Autumn 2009

Software developers: Fool me N times, who cares, this is complex and anyway no one expects software to work...


7 of 56Autumn 2009

FT - June 16, 2004

• "If you have a problem with your Volkswagen the likelihood that it was a software problem is very high. Software technology is not something that we as car manufacturers feel comfortable with.”

Bernd Pischetsrieder, chief executive of Volkswagen


8 of 56Autumn 2009

October 2005

• “Automaker Toyota announced a recall of 160,000 of its Prius hybrid vehicles following reports of vehicle warning lights illuminating for no reason, and cars' gasoline engines stalling unexpectedly.”Wired 05-11-08

• The problem was found to be an embedded software bug


10 of 56Autumn 2009

Driver support: Volvo cars

1984 ABS Anti-lock Braking System

2004 Blind Spot Information system (BLIS)

1998 Dynamic Stability and Traction Control (DSTC)

2006 Active Bi-Xenon lights

2002 Roll Stability Control (RSC)

2006 Adaptive Cruise Control (ACC)

2003 Intelligent Driver Information System (IDIS)

2006 Collision warning system with brake support


11 of 56Autumn 2009

Early space and avionics

• During 1955, 18 air carrier accidents in the USA (when only 20% of the public was willing to fly!)

• Today’s complexity many times higher


12 of 56Autumn 2009

Airbus 380

• Integrated modular avionics (IMA), with safety-critical digital

components, e.g.– Power-by-wire:

complementing the hydraulic powered flight control surfaces

– Cabin pressure control (implemented with a TTP operated bus)


13 of 56Autumn 2009

Dependability

• How do things go wrong?• Why?• What can we do about it?

• Basic notions in dependable systems and replication for fault tolerance

(sv. Pålitliga datorsystem)


14 of 56Autumn 2009

What is dependability?

Property of a computing system which allows reliance to be justifiably placed on the service it delivers.

[Avizienis et al.]

The ability to avoid service failures that are more frequent or more severe than is acceptable.

(sv. Pålitliga datorsystem)


15 of 56Autumn 2009

Attributes of dependability

IFIP WG 10.4 definitions:• Safety: absence of harm to people and

environment• Availability: the readiness for correct

service• Integrity: absence of improper system

alterations• Reliability: continuity of correct service• Maintainability: ability to undergo

modifications and repairs


16 of 56Autumn 2009

[Sv. Tillförlitlighet]Means that the system (functionally) behaves as specified, and does it continually over measured intervals of time.Typical measure in aerospace: 10-9

Another way of putting it: MTTF - One failure in 109 flight hours.

Reliability


17 of 56Autumn 2009

Faults, Errors & Failures

• Fault: a defect within the system or a situation that can lead to failure

• Error: manifestation (symptom) of the fault - an unexpected behaviour

• Failure: system not performing its intended function


18 of 56Autumn 2009

Examples

• Year 2000 bug• Bit flips in hardware due to cosmic

radiation in space• Loose wire• Air craft retracting its landing gear while

on ground

Effects in time: Permanent/ transient/ intermittent


19 of 56Autumn 2009

Fault Error Failure

• Goal of system verification and validation is to eliminate faults

• Goal of hazard/risk analysis is to focus on important faults

• Goal of fault tolerance is to reduce effects of errors if they appear - eliminate or delay failures

Some will remain…


20 of 56Autumn 2009

More on dependability

Four approaches [IFIP 10.4]:

1. Fault avoidance2. Fault removal3. Fault tolerance4. Fault forecasting

Next lecture


21 of 56Autumn 2009

Google’s 100 min outage

September 2, 2009: A small fraction of Gmail’s servers were

taken offline to perform routine upgrades.

“We had slightly underestimated the load which some recent changes (ironically, some designed to improve service availability) placed on the request routers .”

Fault forecasting?


22 of 56Autumn 2009

More on dependability

Four approaches [IFIP 10.4]:

1. Fault avoidance2. Fault removal3. Fault tolerance4. Fault forecasting


23 of 56Autumn 2009

Fault tolerance

• Means that a system provides a degraded (but acceptable) function – Even in presence of faults– During a period defined by certain

model assumptions

• Foreseen or unforeseen?– Fault model describes the foreseen

faults


24 of 56Autumn 2009

External factors


25 of 56Autumn 2009

Fault models

• Leading to Node failures– Crash – Omission – Timing– Byzantine

• Leading to Channel failures– Crash (and potential partitions)– Message loss– Message delay– Erroneous/arbitrary messages


26 of 56Autumn 2009

On-line fault-management

• Fault detection– By program or its environment

• Fault tolerance using redundancy– software– hardware– Data

• Fault containment by architectural choices


27 of 56Autumn 2009

From D. Lardner: Edinburgh Review, year 1824:

”The most certain and effectual check upon errors which arise in the process of computation is to cause the same computations to be made byseparate and independent computers*; and this check is rendered still more decisive if theircomputations are carried out by different methods.”

* people who compute

Redundancy


28 of 56Autumn 2009

Static Redundancy

Used all the time (whether an error has appeared or not), just in case…

– SW: N-version programming– HW: Voting systems– Data: parity bits, checksums


29 of 56Autumn 2009

Dynamic Redundancy

Used when error appears and specifically aids the treatment

– SW: Recovery methods, Exceptions– HW: Switching to back-up module– Data: Self-correcting codes– Time: Re-computing a result


30 of 56Autumn 2009

Server replication models

• Active replication

X

• Passive replication

: Denotes a replica group


31 of 56Autumn 2009

• Active replication– Group membership: who is up, who is down?

• Passive replication– Primary – backup: bring the secondary server up to

date

• What do we need to implement to support transition to operational mode?– Message ordering– Agreement among replicas

Increasing availability


32 of 56Autumn 2009

Relating states

Needs a notion of time (precedence) in distributed systems

Recall from lecture 4:• If clocks are synchronised

– The rate of processing at each node can be related

– Timeouts can be used to detect faults as opposed to message delays

• Asynchronous systems: – Abstract time via order of events


33 of 56Autumn 2009

• Vector clocks help to synchronise at event level– Consistent snapshots

• But reasoning about response times and fault tolerance needs quantitative bounds

Distributed snapshot


34 of 56Autumn 2009

”Chicken and egg” problem

• Replication is useful in presence of failures if there is a consistent common state among replicas

• To get consistency, processes need to communicate their state via broadcast

• But broadcast algorithms are distributed algorithms that run on every node

• and can be affected by failures...


35 of 56Autumn 2009

Desirable broadcast properties

• Reliable broadcast– all non-crashed processes agree on

messages delivered (agreement)– no spurious messages (integrity) – all messages broadcast by non-crashed processes are delivered

(validity)

All or none!


36 of 56Autumn 2009

How to implement?

• The first step is to separate the underlying network (transport) and the broadcast mechanism

• Distinguish between receipt and delivery of a message


37 of 56Autumn 2009

Application layer

Broadcast protocol

Transport

BroadcastSend

Send

Deliver

Receive


38 of 56Autumn 2009

Desirable broadcast properties

• Reliable broadcast– all non-crashed processes agree on

messages delivered (agreement)– no spurious messages (integrity) – all messages broadcast by non-crashed processes delivered

(validity)

All or none!


39 of 56Autumn 2009

Example implementation

At every node n• Execute broadcast(m) by:

– adding sender(m) and a unique ID as a header to the message m (building m)

– send(m) to all neighbours including itself• When receive(m):

– if previously not executed deliver(m) then • if sender(m) /= n then send(m) to all

neighbours• deliver(m)


40 of 56Autumn 2009

Directly after a receipt?

While relaying?

After sending to some but not all neighbours?

This is where failure models come in...

What if n fails?


41 of 56Autumn 2009

Algorithms for broadcast

• Correctness: prove validity, integrity, agreement, order

Typical assumptions:– no link failures leading to partition– send does not duplicate or change

messages– receive does not ”invent” messages

In presence of a given failure model!


42 of 56Autumn 2009

The consensus problem

• Processes p1,…,pn take part in a decision• Each pi proposes a value vi

– e.g. application state info• All correct processes decide on a

common value v that is equal to one of the proposed values

Assume that we have a reliable broadcast


43 of 56Autumn 2009

Desired properties

• Every non-faulty process eventually decides (Termination)

• No two non-faulty processes decide differently (Agreement)

• If a process decides v then the value v was proposed by some process (Validity)

Algorithms for consensus have to be proven to

have these properties

In presence of given failure model


44 of 56Autumn 2009

Basic impossibility result

[Fischer, Lynch and Paterson 1985]

There is no deterministic algorithm solving the consensus problem in an asynchronous distributed system with a single crash failure


45 of 56Autumn 2009

A way around it

Assume Synchrony:

• Distributed computations proceed in rounds initiated by pulses

• Pulses implemented using local physical clocks, synchronised assuming bounded message delays


46 of 56Autumn 2009

Can I prove any protocol correct?

• NO!

• It depends on the fault model:– Assumptions on

how can the nodes/channels fail


47 of 56Autumn 2009

Problem:• Nodes need to agree on a decision but

some nodes are faulty, act in an arbitrary way (can be malicious)

Byzantine agreement


48 of 56Autumn 2009

Byzantine agreement protocol

• Proposed in 1980 by Pease, Shostak and Lamport

How many faulty nodes can the

algorithm tolerate?


49 of 56Autumn 2009

Result from 1980

• Theorem: There is an upper bound t for the number of byzantine node failures compared to the size of the network N, N 3t+1

• Gives a t+1 round algorithm for solving consensus in a synchronous network

• Here:– We just demonstrate that t nodes

would not be enough!


50 of 56Autumn 2009

Scenario 1

• G and L1 are correct, L2 is faulty

G

L1 L2

1

1

1

0

1

0L2

said

0L2

said

0

L1 said 1L1 said 1

G said 1

G said 0L1 L2

G


51 of 56Autumn 2009

Scenario 2

• G and L2 are correct, L1 is faulty

G

L1 L2

1

0

0

0

1

0L2

said

0L2

said

0

L1 said 1L1 said 1

G said 1

G said 0L1 L2

G


52 of 56Autumn 2009

Scenario 3

• The general is faulty!G

L1 L2

1

1

0

0

1

0L2

said

0L2

said

0

L1 said 1L1 said 1

G said 1

G said 0L1 L2

G


53 of 56Autumn 2009

2-round algorithm

… does not work with t=1, N=3!

• Seen from L1, scenario 1 and 3 are identical, so if L1 decides 1 in scenario 1 it will decide 1 in scenario 3

• Similarly for L2, if it decides 0 in scenario 2 it decides 0 in scenario 3

• L1 and L2 do not agree in scenario 3 !


54 of 56Autumn 2009

Summary

• Dependable systems need to justify why services can be relied upon

• To tolerate faults we normally deploy replication

• Replication needs (some kind of) agreement on state

• Agreement problems are hard to prove correct in distributed systems

• Require assumptions on faults, and (some kind of) synchrony


55 of 56Autumn 2009

• Remember in a TTP bus: – Nodes have synchronised clocks– The communication infrastructure (collection

of CCs and the TTP bus communication) guarantee a bounded time for a new data appearing comm. interface of a node (its CNI)

– Byzantine faults in nodes can be detected by other nodes!

• Exercise: Which faults are detectable in a system connected with a CAN bus?

Real-time communication


56 of 56Autumn 2009

Timing and fault tolerance

• Implementing fault tolerance often relies support for timing guarantees

• Exercise: How do faults affect support for timing?– How do faults affect response times in

nodes?– How do fault tolerance mechanisms

affect timing?

dependability and real-time

Documents