Fault tolerance


Fault tolerance is the property that enables a system to continue operating properly in the event of the failure of some of its components. If its operating quality decreases at all, the decrease is proportional to the severity of the failure, as compared to a naively designed system, in which even a small failure can cause total breakdown. Fault tolerance is particularly sought after in high-availability or life-critical systems. The ability of maintaining functionality when portions of a system break down is referred to as graceful degradation.
A fault-tolerant design enables a system to continue its intended operation, possibly at a reduced level, rather than failing completely, when some part of the system fails. The term is most commonly used to describe computer systems designed to continue more or less fully operational with, perhaps, a reduction in throughput or an increase in response time in the event of some partial failure. That is, the system as a whole is not stopped due to problems either in the hardware or the software. An example in another field is a motor vehicle designed so it will continue to be drivable if one of the tires is punctured, or a structure that is able to retain its integrity in the presence of damage due to causes such as fatigue, corrosion, manufacturing flaws, or impact.
Within the scope of an individual system, fault tolerance can be achieved by anticipating exceptional conditions and building the system to cope with them, and, in general, aiming for self-stabilization so that the system converges towards an error-free state. However, if the consequences of a system failure are catastrophic, or the cost of making it sufficiently reliable is very high, a better solution may be to use some form of duplication. In any case, if the consequence of a system failure is so catastrophic, the system must be able to use reversion to fall back to a safe mode. This is similar to roll-back recovery but can be a human action if humans are present in the loop.

History

The first known fault-tolerant computer was SAPO, built in 1951 in Czechoslovakia by Antonín Svoboda. Its basic design was magnetic drums connected via relays, with a voting method of memory error detection. Several other machines were developed along this line, mostly for military use. Eventually, they separated into three distinct categories: machines that would last a long time without any maintenance, such as the ones used on NASA space probes and satellites; computers that were very dependable but required constant monitoring, such as those used to monitor and control nuclear power plants or supercollider experiments; and finally, computers with a high amount of runtime which would be under heavy use, such as many of the supercomputers used by insurance companies for their probability monitoring.
Most of the development in the so-called LLNM computing was done by NASA during the 1960s, in preparation for Project Apollo and other research aspects. NASA's first machine went into a space observatory, and their second attempt, the JSTAR computer, was used in Voyager. This computer had a backup of memory arrays to use memory recovery methods and thus it was called the JPL Self-Testing-And-Repairing computer. It could detect its own errors and fix them or bring up redundant modules as needed. The computer is still working today.
Hyper-dependable computers were pioneered mostly by aircraft manufacturers, nuclear power companies, and the railroad industry in the USA. These needed computers with massive amounts of uptime that would fail gracefully enough with a fault to allow continued operation while relying on the fact that the computer output would be constantly monitored by humans to detect faults. Again, IBM developed the first computer of this kind for NASA for guidance of Saturn V rockets, but later on BNSF, Unisys, and General Electric built their own.
In the 1970s, much work has happened in the field
. For instance, F14 CADC had built-in self-test and redundancy.
In general, the early efforts at fault-tolerant designs were focused mainly on internal diagnosis, where a fault would indicate something was failing and a worker could replace it. SAPO, for instance, had a method by which faulty memory drums would emit a noise before failure. Later efforts showed that to be fully effective, the system had to be self-repairing and diagnosing – isolating a fault and then implementing a redundant backup while alerting a need for repair. This is known as N-model redundancy, where faults cause automatic fail-safes and a warning to the operator, and it is still the most common form of level one fault-tolerant design in use today.
Voting was another initial method, as discussed above, with multiple redundant backups operating constantly and checking each other's results, with the outcome that if, for example, four components reported an answer of 5 and one component reported an answer of 6, the other four would "vote" that the fifth component was faulty and have it taken out of service. This is called M out of N majority voting.
Historically, the motion has always been to move further from N-model and more to M out of N due to the fact that the complexity of systems and the difficulty of ensuring the transitive state from fault-negative to fault-positive did not disrupt operations.
Tandem and Stratus were among the first companies specializing in the design of fault-tolerant computer systems for online transaction processing..

Terminology

A highly fault-tolerant system might continue at the same level of performance even though one or more components have failed. For example, a building with a backup electrical generator will provide the same voltage to wall outlets even if the grid power fails.
A system that is designed to fail safe, or fail-secure, or fail gracefully, whether it functions at a reduced level or fails completely, does so in a way that protects people, property, or data from injury, damage, intrusion, or disclosure. In computers, a program might fail-safe by executing a graceful exit in order to prevent data corruption after experiencing an error. A similar distinction is made between "failing well" and "failing badly".
Fail-deadly is the opposite strategy, which can be used in weapon systems that are designed to kill or injure targets even if part of the system is damaged or destroyed.
A system that is designed to experience graceful degradation, or to fail soft operates at a reduced level of performance after some component failures. For example, a building may operate lighting at reduced levels and elevators at reduced speeds if grid power fails, rather than either trapping people in the dark completely or continuing to operate at full power. In computing an example of graceful degradation is that if insufficient network bandwidth is available to stream an online video, a lower-resolution version might be streamed in place of the high-resolution version. Progressive enhancement is an example in computing, where web pages are available in a basic functional format for older, small-screen, or limited-capability web browsers, but in an enhanced version for browsers capable of handling additional technologies or that have a larger display available.
In fault-tolerant computer systems, programs that are considered robust are designed to continue operation despite an error, exception, or invalid input, instead of crashing completely. Software brittleness is the opposite of robustness. Resilient networks continue to transmit data despite the failure of some links or nodes; resilient buildings and infrastructure are likewise expected to prevent complete failure in situations like earthquakes, floods, or collisions.
A system with high failure transparency will alert users that a component failure has occurred, even if it continues to operate with full performance, so that failure can be repaired or imminent complete failure anticipated. Likewise, a fail-fast component is designed to report at the first point of failure, rather than allow downstream components to fail and generate reports then. This allows easier diagnosis of the underlying problem, and may prevent improper operation in a broken state.

Single fault condition

A single fault condition is a situation where one means for protection against a hazard is defective. If a single fault condition results unavoidably in another single fault condition, the two failures are considered as one single fault condition. A source offers the following example:

Redundancy

Redundancy is the provision of functional capabilities that would be unnecessary in a fault-free environment.
This can consist of backup components that automatically "kick in" if one component fails. For example, large cargo trucks can lose a tire without any major consequences. They have many tires, and no one tire is critical.
The idea of incorporating redundancy in order to improve the reliability of a system was pioneered by John von Neumann in the 1950s.
Two kinds of redundancy are possible: space redundancy and time redundancy. Space redundancy provides additional components, functions, or data items that are unnecessary for fault-free operation. Space redundancy is further classified into hardware, software and information redundancy, depending on the type of redundant resources added to the system. In time redundancy the computation or data transmission is repeated and the result is compared to a stored copy of the previous result. The current terminology for this kind of testing is referred to as 'In Service Fault Tolerance Testing or ISFTT for short.

Criteria

Providing fault-tolerant design for every component is normally not an option. Associated redundancy brings a number of penalties: increase in weight, size, power consumption, cost, as well as time to design, verify, and test. Therefore, a number of choices have to be examined to determine which components should be fault tolerant:
An example of a component that passes all the tests is a car's occupant restraint system. While we do not normally think of the primary occupant restraint system, it is gravity. If the vehicle rolls over or undergoes severe g-forces, then this primary method of occupant restraint may fail. Restraining the occupants during such an accident is absolutely critical to safety, so we pass the first test. Accidents causing occupant ejection were quite common before seat belts, so we pass the second test. The cost of a redundant restraint method like seat belts is quite low, both economically and in terms of weight and space, so we pass the third test. Therefore, adding seat belts to all vehicles is an excellent idea. Other "supplemental restraint systems", such as airbags, are more expensive and so pass that test by a smaller margin.
Another excellent and long-term example of this principle being put into practice is the braking system: whilst the actual brake mechanisms are critical, they are not particularly prone to sudden failure, and are in any case necessarily duplicated to allow even and balanced application of brake force to all wheels. It would also be prohibitively costly to further double-up the main components and they would add considerable weight. However, the similarly critical systems for actuating the brakes under driver control are inherently less robust, generally using a cable or hydraulic fluid. Thus in most modern cars the footbrake hydraulic brake circuit is diagonally divided to give two smaller points of failure, the loss of either only reducing brake power by 50% and not causing as much dangerous brakeforce imbalance as a straight front-back or left-right split, and should the hydraulic circuit fail completely, there is a failsafe in the form of the cable-actuated parking brake that operates the otherwise relatively weak rear brakes, but can still bring the vehicle to a safe halt in conjunction with transmission/engine braking so long as the demands on it are in line with normal traffic flow. The cumulatively unlikely combination of total foot brake failure with the need for harsh braking in an emergency will likely result in a collision, but still one at lower speed than would otherwise have been the case.
In comparison with the foot pedal activated service brake, the parking brake itself is a less critical item, and unless it is being used as a one-time backup for the footbrake, will not cause immediate danger if it is found to be nonfunctional at the moment of application. Therefore, no redundancy is built into it per se, and it can suffice, if this happens on a hill, to use the footbrake to momentarily hold the vehicle still, before driving off to find a flat piece of road on which to stop. Alternatively, on shallow gradients, the transmission can be shifted into Park, Reverse or First gear, and the transmission lock / engine compression used to hold it stationary, as there is no need for them to include the sophistication to first bring it to a halt.
On motorcycles, a similar level of fail-safety is provided by simpler methods; firstly the front and rear brake systems being entirely separate, regardless of their method of activation, allowing one to fail entirely whilst leaving the other unaffected. Secondly, the rear brake is relatively strong compared to its automotive cousin, even being a powerful disc on sports models, even though the usual intent is for the front system to provide the vast majority of braking force; as the overall vehicle weight is more central, the rear tyre is generally larger and grippier, and the rider can lean back to put more weight on it, therefore allowing more brake force to be applied before the wheel locks up. On cheaper, slower utility-class machines, even if the front wheel should use a hydraulic disc for extra brake force and easier packaging, the rear will usually be a primitive, somewhat inefficient, but exceptionally robust rod-actuated drum, thanks to the ease of connecting the footpedal to the wheel in this way and, more importantly, the near impossibility of catastrophic failure even if the rest of the machine, like a lot of low-priced bikes after their first few years of use, is on the point of collapse from neglected maintenance.

Requirements

The basic characteristics of fault tolerance require:
  1. No single point of failure – If a system experiences a failure, it must continue to operate without interruption during the repair process.
  2. Fault isolation to the failing component – When a failure occurs, the system must be able to isolate the failure to the offending component. This requires the addition of dedicated failure detection mechanisms that exist only for the purpose of fault isolation. Recovery from a fault condition requires classifying the fault or failing component. The National Institute of Standards and Technology categorizes faults based on locality, cause, duration, and effect.
  3. Fault containment to prevent propagation of the failure – Some failure mechanisms can cause a system to fail by propagating the failure to the rest of the system. An example of this kind of failure is the "rogue transmitter" that can swamp legitimate communication in a system and cause overall system failure. Firewalls or other mechanisms that isolate a rogue transmitter or failing component to protect the system are required.
  4. Availability of reversion modes
In addition, fault-tolerant systems are characterized in terms of both planned service outages and unplanned service outages. These are usually measured at the application level and not just at a hardware level. The figure of merit is called availability and is expressed as a percentage. For example, a five nines system would statistically provide 99.999% availability.
Fault-tolerant systems are typically based on the concept of redundancy.

Fault tolerance techniques

Research into the kinds of tolerances needed for critical systems involves a large amount of interdisciplinary work. The more complex the system, the more carefully all possible interactions have to be considered and prepared for. Considering the importance of high-value systems in transport, public utilities and the military, the field of topics that touch on research is very wide: it can include such obvious subjects as software modeling and reliability, or hardware design, to arcane elements such as stochastic models, graph theory, formal or exclusionary logic, parallel processing, remote data transmission, and more.

Replication

Spare components address the first fundamental characteristic of fault tolerance in three ways:
All implementations of RAID, redundant array of independent disks, except RAID 0, are examples of a fault-tolerant storage device that uses data redundancy.
A lockstep fault-tolerant machine uses replicated elements operating in parallel. At any time, all the replications of each element should be in the same state. The same inputs are provided to each replication, and the same outputs are expected. The outputs of the replications are compared using a voting circuit. A machine with two replications of each element is termed dual modular redundant. The voting circuit can then only detect a mismatch and recovery relies on other methods. A machine with three replications of each element is termed triple modular redundant. The voting circuit can determine which replication is in error when a two-to-one vote is observed. In this case, the voting circuit can output the correct result, and discard the erroneous version. After this, the internal state of the erroneous replication is assumed to be different from that of the other two, and the voting circuit can switch to a DMR mode. This model can be applied to any larger number of replications.
Lockstep fault-tolerant machines are most easily made fully synchronous, with each gate of each replication making the same state transition on the same edge of the clock, and the clocks to the replications being exactly in phase. However, it is possible to build lockstep systems without this requirement.
Bringing the replications into synchrony requires making their internal stored states the same. They can be started from a fixed initial state, such as the reset state. Alternatively, the internal state of one replica can be copied to another replica.
One variant of DMR is pair-and-spare. Two replicated elements operate in lockstep as a pair, with a voting circuit that detects any mismatch between their operations and outputs a signal indicating that there is an error. Another pair operates exactly the same way. A final circuit selects the output of the pair that does not proclaim that it is in error. Pair-and-spare requires four replicas rather than the three of TMR, but has been used commercially.

Failure-oblivious computing

Failure-oblivious computing is a technique that enables computer programs to continue executing despite memory errors. The technique handles attempts to read invalid memory by returning a manufactured value to the program, which in turn, makes use of the manufactured value and ignores the former memory value it tried to access. This is a great contrast to typical memory checkers, which inform the program of the error or abort the program. In failure-oblivious computing, no attempt is made to inform the program that an error occurred. Furthermore, it happens that the execution is modified several times in a row, in order to prevent cascading failures.
The approach has performance costs: because the technique rewrites code to insert dynamic checks for address validity, execution time will increase by 80% to 500%.

Recovery shepherding

Recovery shepherding is a lightweight technique to enable software programs to recover from otherwise fatal errors such as null pointer dereference and divide by zero. Comparing to the failure oblivious computing technique, recovery shepherding works on the compiled program binary directly and does not need to recompile to program.
It uses the just-in-time binary instrumentation framework Pin. It attaches to the application process when an error occurs, repairs the execution,
tracks the repair effects as the execution continues, contains the repair effects within the application process, and detaches from the process after all repair effects are flushed from the process state. It does not interfere with the normal execution of the program and therefore incurs negligible overhead. For 17 of 18 systematically collected real world null-dereference and divide-by-zero errors, a prototype implementation enables the application to continue to execute to provide acceptable output and service to its users on the error-triggering inputs.

Circuit breaker

The circuit breaker design pattern is a technique to avoid catastrophic failures in distributed systems.

Disadvantages

Fault-tolerant design's advantages are obvious, while many of its disadvantages are not:
Hardware fault tolerance sometimes requires that broken parts be taken out and replaced with new parts while the system is still operational. Such a system implemented with a single backup is known as single point tolerant and represents the vast majority of fault-tolerant systems. In such systems the mean time between failures should be long enough for the operators to have time to fix the broken devices before the backup also fails. It helps if the time between failures is as long as possible, but this is not specifically required in a fault-tolerant system.
Fault tolerance is notably successful in computer applications. Tandem Computers built their entire business on such machines, which used single-point tolerance to create their NonStop systems with uptimes measured in years.
Fail-safe architectures may encompass also the computer software, for example by process replication.
Data formats may also be designed to degrade gracefully. HTML for example, is designed to be forward compatible, allowing new HTML entities to be ignored by Web browsers that do not understand them without causing the document to be unusable.

Related terms

There is a difference between fault tolerance and systems that rarely have problems. For instance, the Western Electric crossbar systems had failure rates of two hours per forty years, and therefore were highly fault resistant. But when a fault did occur they still stopped operating completely, and therefore were not fault tolerant.