TOWARDS PARSIMONIOUS RESOURCE ALLOCATION IN CONTEXT-AWARE N-VERSION PROGRAMMING

TOWARDS PARSIMONIOUS RESOURCE ALLOCATION
IN CONTEXT-AWARE N-VERSION PROGRAMMING
Jonas Buys, Vincenzo De Florio and Chris Blondia
University of Antwerp, Dept. of Mathematics and Computer Science, Antwerp, Belgium
Interdisciplinary Institute for Broadband Technology, Ghent-Ledeberg, Belgium
{jonas.buys, vincenzo.deflorio, chris.blondia}@ua.ac.be
Keywords: fault-tolerance, n-version programming, contextawareness, measurement techniques, dependability.

integrated within the dependability scheme. Indeed, the use
of replicas of poor reliability can result in a system tolerant of
faults but with poor reliability [5]. It is therefore crucial for
the system to continuously monitor the operational status of
the available resources and avoid the use of resources that do
not significantly contribute to an increase of dependability, or
that may even jeopardise the schemes' overall effectiveness.

Abstract
Adopting classic redundancy-based fault-tolerant schemes in
highly dynamic distributed computing systems does not
necessarily result in the anticipated improvement in
dependability. This primarily stems from statically predefined
redundancy configurations employed within many classic
dependability strategies, which as well known may negatively
impact the schemes' overall effectiveness. In this paper, a
novel dependability strategy is introduced encompassing
advanced redundancy management, aiming to autonomously
tune its internal configuration in function of disturbances
observed. Policies for parsimonious resource allocation are
presented thereafter, intent upon increasing the scheme's cost
effectiveness without breaching its availability objective. Our
experimentation suggests that the suggested solution can
achieve a substantial improvement in availability, compared
to traditional, static redundancy strategies, and that tuning the
adopted degree of redundancy to the actual observed
disturbances allows unnecessary resource expenditure to be
reduced, therefore enhancing cost-effectiveness.

Secondly, redundant implementations imply significantly
higher development costs and increased infrastructural
requirements. A predetermined degree of redundancy is
therefore cost ineffective in that it inhibits to economise on
resource consumption in case the actual number of
disturbances could be successfully overcome by a lesser
amount of redundancy. Reversely, when the foreseen amount
of redundancy is not enough to compensate for the currently
experienced disturbances, the inclusion of additional
resources may prevent further service disruption.
In this paper, a novel, context-aware dependability strategy is
introduced encompassing advanced redundancy management.
Designed to sustain high availability and reliability, this
adaptive fault-tolerant strategy dynamically alters the degree
of redundancy and the employed selection of resources.
The remainder of this paper is structured as follows: We first
present the concept of n-version programming (NVP)
schemata in Sect. 2 and show how software design failures in
the content failure domain may challenge their effectiveness.
A set of ancillary metrics is first set forth in Sect. 3, allowing
to deduce knowledge of the context in which the scheme is
operating and to detect the proximity of hazardous situations
that may require the adjustment of the redundancy
configuration. We then move on to elaborate on the internals
of the proposed adaptive fault-tolerant strategy in Sect. 4.
Policies for parsimonious resource allocation are presented
thereafter, intent upon increasing the scheme's cost
effectiveness without breaching its availability objective.
Sect. 6 reports on the strategy's effectiveness, after which this
paper is concluded.

1 Introduction
Business- and mission-critical distributed applications are
increasingly expected to exhibit highly dependable characteristics, particularly in the areas of reliability and timeliness.
Redundancy-based fault-tolerant strategies have long been
used as a means to avoid a disruption in the service provided
by the system in spite of failures in the underlying software
components. Adopting fault-tolerance strategies in dynamic
distributed computing systems, in which components often
suffer from long response times or temporary unavailability,
does not necessarily result in the anticipated improvement in
dependability. This primarily stems from statically predefined
redundancy configurations employed within many classic
dependability strategies, i.e. a fixed degree of redundancy
and, accordingly, an immutable selection of functionallyequivalent software components, which may negatively
impact the schemes' effectiveness from the following angles:

2 n-Version Programming with Majority Voting
NVP was first introduced in 1985 as a tool to provide
protection against software design faults [5]. The rationale is
that deploying multiple functionally-equivalent independently
implemented software components will hopefully reduce the
probability of a software fault affecting multiple implementations simultaneously, thereby keeping the system operational.

Firstly, a static, context-agnostic configuration may in time
lead to a more rapid exhaustion of the available redundancy
and therefore fail to properly counterbalance any disturbances
affecting the operational status (context) of the components

1

timed asynchronous system model [2, 6]. For such environments, several types of potential disturbances were defined in
the failure manifestation model provided in [2]. We will now
elaborate on a subset of these disturbances that may ensue
from the activation of a software design fault and manifest
within the content failure domain. Such types of disturbances
exhibit a transient behaviour and emerge exclusively from the
activation of a latent software design fault along the execution
path followed whilst version vi is processing some request
‹c, ℓ, i›, i.e. when the request is in the request processing state
(see Fig. 1). One may distinguish two types of failures, each
with different repercussions on the generated partition ℘( c ,ℓ ) .

Figure 1: State transition diagram of voting round (c,ℓ) and
the underlying invocations ‹c, ℓ, i› of versions vi.
An n-version composite constitutes a client-transparent
replication layer in which all n programs, called versions,
receive a copy of the user input and independently perform
their computations in parallel. Let {ℓx}c be a sequence of
monotonically increasing, strictly positive integer indices
ℓx = x in L = ℕ +, such that each voting round, i.e. a single
invocation of an NVP composite c, is uniquely identified. As
shown in Fig. 1, the arrival of a request message at the
composite interface will trigger the initialisation of a new
voting round (c,ℓ) with ℓ the next element in {ℓx}c.
Immediately after, the system is to retrieve the redundancy
configuration to be used throughout the newly initialised
voting round (c,ℓ), i.e. the amount of redundancy used and,
accordingly, a selection of functionally-equivalent versions
(FeV) (transition from state a to b). We define V as the set of
all FeV in the system. For a given round (c,ℓ), the amount of
redundancy used within the NVP scheme is denoted as
n(c,ℓ) ≥ 1, such that the versions employed for round (c,ℓ) are
contained within V(c,ℓ) ⊆ V and n(c,ℓ) = |V(c,ℓ)|. Having acquired
the redundancy configuration, the request message payload is
then replicated and forwarded to each of the selected versions
vi ∈ V(c,ℓ), for i ∈ {1, ..., n(c,ℓ)}. Each such invocation of a
version vi can be uniquely identified by the tuple ‹c, ℓ, i›. As
soon as a result is available for each of the versions involved,
the transition from state b to c fires and a majority voting
(MV) procedure is called so as to adjudicate the result of the
scheme, which is then returned to the client.

Firstly, despite sharing a common specification, functionallyequivalent versions may exhibit discrepancies resulting in
response value failures (RVF) [4]. Such failures will usually
not make the system appear to fail, as the content of the
response returned via the service interface is syntactically
correct, though not compliant to the functional specifications.
ɺ
During voting, a partition ∪ j ∈{1,…,k ( c ,ℓ ) } Pj is constructed for all
versions vi ∈ V(c,ℓ) that returned a syntactically valid response,
allotting versions that reported equivalent results to the same
equivalence classes Pj.
Secondly, the activation of a design fault may cause an abrupt
interruption of the flow of execution due to an exception ‒ a
case of so-called erroneous value failures (EVF). Falling
within the content failure domain, a syntactically invalid
response message will be returned for the affected invocation,
containing the serialised exception thrown. Versions affected
(c,
by an EVF are classified in PF ℓ ). EVF failures may overrule
previously activated RVF failures affecting an invocation [2].
Adhering to the discrete-event model developed in [2] for the
injection of transient disturbances in NVP schemata, these
two disturbance types will be used to assess the performance
of the dependability strategy introduced in Sect. 4.

3. Redundancy configuration effectiveness
Let P(c,ℓ) be the set of largest cardinality in the generated
(c,
partition ℘(c,ℓ ) PF ℓ.) Then c(c,ℓ ) = | P (c,ℓ ) | represents the largest
max
consent found amongst the versions in V(c,ℓ) within the scope
of (c,ℓ). In order for the scheme to adjudicate a result o,
there should be a consensus amongst an absolute majority of
the n(c,ℓ) versions, i.e. c(c,ℓ ) ≥ m(c,ℓ ) , with
max

The effectiveness of a scheme such as NVP/MV is largely
determined by its redundancy configuration and its ability to
counterbalance the disturbances ensuing from the
environment in which it operates and to which it is subject.
For instance, an NVP/MV scheme can mask failures affecting
the availability of up to a minority of its n(c,ℓ) versions.

(c, ℓ )
.
m ( c, ℓ ) =  n

2

The essential part of any voting procedure is the construction
(c, ℓ )
(c,
ɺ
= {PF ℓ ) , ∪ j ∈{1,…, k ( c , ℓ ) } Pj } of the set of
of a partition ℘
(c,ℓ)
versions V . The notation of a disjoint union to represent
the consensus blocks as part of this partition has been taken
from [1]. This partitioning procedure is influenced by the
disturbances that affected any of the requests ‹c, ℓ, i› involved
during the voting round (c,ℓ). Throughout this paper, the
notion of disturbance is used to denote the event of a single
request ‹c, ℓ, i› struck by the some type of failure, resulting in
the perturbation and, consequently, the (temporary) unavailability of the service that version vi is expected to provide.

(1)

Put differently, m(c,ℓ) is indicative of the smallest degree of
consent needed for a consensus block P(c,ℓ) to qualify for the
equivalence class [o]. Consequently, a scheme is resilient to
withstand disturbances affecting at most a minority
n(c,ℓ) – m(c,ℓ) of the n(c,ℓ) versions used throughout round (c,ℓ).
We now introduce a new metric defined as c(c,ℓ ) - m (c,ℓ )
max
that was designed to provide a quantitative estimation of how
closely the current amount and selection of resources matched
the observed disturbances ― by shortcoming or excess. More
specifically, it can be used to deduce a measure of the
proximity of hazardous situations that may necessitate the

The application of NVP schemata in contemporary distributed
computing systems is assumed to exhibit the properties of a

2

4.2. Redundancy dimensioning model

adjustment of the currently employed redundancy configu(c, ℓ )
(c,ℓ )
ration. One can easily see that Ran(c max − m ) lies in
(c, ℓ )
(c,ℓ)
(c,ℓ)
(c,ℓ)
[-m , n – m ], for c max was defined in [0, n(c,ℓ)]. It
provides an indirect estimation of the shortage or abundance
of redundancy with respect to the disturbances affecting
round (c,ℓ): a positive value essentially quantifies how many
versions exist in excess of the mandatory m(c,ℓ) versions that
collectively constitute a majority for round (c,ℓ). For negative values, the absolute value represents the lack of consent
relative to P(c,ℓ) that would be required so as to constitute a
majority. Such a value is interpreted as a symptom that the
currently experienced disturbances cannot be successfully
counterbalanced by the redundancy configuration used, and
the scheme would fail to guarantee the availability of the
service it seeks to provide despite its fault-tolerant nature. A
critically low value 0 represents a situation for which the
majority was attained by only m(c,ℓ) versions: the available
redundancy was completely exhausted to counterbalance the
maximal number of disturbances the scheme could tolerate
and any additional disturbance would have led to failure.

Given the FeV in V, this model is responsible for
autonomously adjusting the degree of redundancy employed
such that it closely follows the evolution of the observed
disturbances. In the absence of disturbances, the scheme
should scale down its use of replicas and avoid unnecessary
resource expenditure. Contrarily, when the foreseen amount
of redundancy is not enough to compensate for the currently
experienced disturbances, it will dynamically revise that
amount and enrol additional resources if available. The
model determines n(c,ℓ) upon initialisation of round (c,ℓ), such
that nmin ≤ n(c,ℓ) ≤ min(nmax, |V|). Ideally, the redundancy
configuration would involve a number n(c,ℓ) of versions
consistently greater than or equal but close to cr(e(c,ℓ)), i.e. the
minimal amount of redundancy required during the
operational time interval for a voting round (c,ℓ) in order to
stay dependable and have the decision algorithm return the
correct result in spite of being challenged by a number of
disturbances e(c,ℓ) , which can be expressed as the function
cr : ℕ 0 → ℕ + such that cr(e(c,ℓ)) = 2 e(c,ℓ) + 1 [2].

4. A Novel Adaptive Fault-Tolerant Strategy
We introduce our adaptive NVP-based fault-tolerant strategy
(A-NVP) and elaborate on the advanced redundancy
management it supports. Our context-aware reformulation of
the classical NVP approach encompasses two complementary
parameterised models that jointly determine the optimal
redundancy configuration to be used throughout a newly
initialised voting round (c,ℓ) in view of disturbances and how
these were perceived to have affected the system context.
The redundancy configuration is retrieved whilst preparing to
fire the transition from state a to b, as shown in Fig. 1.

Figure 2: View on window-based data structure before n(c,ℓ)
is set for use within a newly initialised round (c,ℓ).
4.2.1. Window of Context Information
Let {yz}c be a monotonically increasing sequence of strictly
positive integer indices yz = z in Y = ℕ + such that each
consecutive completion of some voting round (c,ℓ)
originating from an invocation of c is uniquely identified by
the next element y in {yz}c. Observe how z denotes the
number of completed voting rounds appertaining to c. The
bijection bc : Y → L defines the correlation between the terms
in either of the sequences {yz}c and {ℓx}c. The order in which
rounds complete, i.e. the sequence {yz}c, is not necessarily the
order in which these rounds have been initialised, represented
by the sequence {ℓx}c. Indeed, in large-scale distributed
computing environments, one may expect a significant
amount of variability in the response time of an invocation on
the NVP composite, due e.g. to its dynamically changing
redundancy configuration used for different voting rounds.
More information on timeliness issues can be found in [2].

The first model is responsible for determining the appropriate
degree of redundancy n(c,ℓ) to be used and will allow economising on resource expenditures whenever it can argued safe to
do. Intent upon increasing the scheme’s cost effectiveness
without breaching its dependability objective, policies for
parsimonious resource allocation presented in Sect. 5. Next,
the replica selection model will establish an appropriate set of
resources V(c,ℓ) maximising the objectives expressed by means
of a set of application-agnostic properties [3].
4.1. Application-specific requirements
The optimal redundancy configuration is not only determined
by the context property introduced in Sect. 3, but also by the
characteristics of the application itself, or the environment in
which it operates. Some applications may operate in a
resource-constrained environment. The A-NVP algorithm was
conceived to take these application-specific intricacies into
account, in that both models can be configured by means of a
set of user-defined parameters. Firstly, a parameter nmax ≥ 1
can be used to set an upper bound on the number of replicas
to be used. Contrariwise, parameter 2e + 1 = nmin ≥ 1 can be
used to set a lower bound on the degree of redundancy to be
used, such that the scheme is guaranteed to tolerate at least e
disturbances. Lastly, parameter ninit will set the default degree
of redundancy to be initially used, i.e. n(c,ℓ) = ninit for ℓ = 1.

In order to make an informed decision on the amount of
redundancy n(c,ℓ) to be used, the model will consider the
course of the amount of redundancy used for previously
completed voting rounds and whether or not the selected
redundancy proved to be sufficient to guarantee the scheme's
availability. As such, the context manager within the NVP
composite will maintain a window-based data structure to
store this contextual information for each y in the
subsequence {y min(1,z − r d +1) , … , y z } of {yz}c. More specifically,
for each completed voting round y, the context properties of

3

interest are: the corresponding identifier ℓx = bc(y), the
amount of redundancy n (c,ℓ x ) that was employed throughout its
execution, and the extent to which this redundancy was found
to be capable of masking disturbances ― cf. Sect. 3.

notion of window semantics Sc to epitomise the specific
conditionalities and correlational techniques that enable the
redundancy dimensioning model to deduce the optimum
degree of redundancy matching the scheme's operational
context from the stored information. In this capacity, Sc
defines two ancillary functions describing a relation L → ℕ +.
More specifically, the upscaling function fu(c,ℓ) is responsible
for determining if and to what extent the current level of
redundancy n(c,ℓ-1) should increase, whereas the downscaling
function fd(c,ℓ) quantifies the extent by which n(c,ℓ-1) should
be lowered. The final degree of redundancy n(c,ℓ) to be used
for the continuation of (c,ℓ) is then resolved as follows:

Let rd, ru and rf be numbers in ℕ + such that 1 ≤ rf ≤ ru ≤ rd.
Number rd represents the minimum number of consecutive
successful voting round completions before contemplating
scaling down the current level of redundancy. In line with
Sect. 2, a given voting round (c,ℓ) is observed to have
completed successfully if a sufficiently large degree of
consent could be found between the responses acquired from
the subordinate invocations ‹c, ℓ, i› of the involved versions
vi ∈ V(c,ℓ) such that a majority could be found,
i.e. c(c,ℓ ) − m (c,ℓ ) ≥ csm, with a discretionary safety margin csm
max
defined as an integer in [0, n(c,ℓ) – m(c,ℓ)] (SC). As shown in
Fig. 2, rd imposes an upper bound on the maximum window
capacity maintained by the data structure. It is assumed that
shorter window lengths may result in incautious downscaling
of the redundancy, which in itself might lead to failure of the
voting scheme in subsequent voting rounds. Contrariwise,
one may reasonably expect that the redundancy scheme is less
likely to fail due to the downscaling of the employed degree
of redundancy for larger values of rd, at the expense of
postponing the relinquishment of excess redundancy [3].

n

(c,ℓ)

min(min(nmax, |V|), n(c,ℓ-1) + fu(c,ℓ)) | fu(c,ℓ) > 0 (2a)
max(nmin, n(c,ℓ-1) – fd(c,ℓ))
| fd(c,ℓ) > 0 (2b)
n(c,ℓ-1)
| else
(2c)

The above Eq. (2a) and (2b) formalise the upscaling, resp.
downscaling procedure for ℓ > 1. Observe how the adjustment of the redundancy is constrained by the applicationspecific parameters nmin and nmax, as well as by the amount |V|
of resources available. Any specific Sc should implement
these two functions such that a value is returned only if the
window contains information abiding the success criterion
(SC) and the definitions given for rd, ru and rf hereabove.
The optional safety margin csm expresses the amount of
consent supplementary to the mandatory m(c,ℓ) required for the
successful adjudication of a result. It serves as a parameter to
the redundancy dimensioning model, primarily aiming to
reduce the likelihood that the downscaling procedure itself
would result in failure of the scheme in the first few
subsequent voting rounds. Moreover, such safety margin
could anticipate a shortfall in redundancy when the
effectiveness of the employed redundancy is observed to
exhibit a decreasing trend and proactively trigger the
upscaling procedure. Either way, the underlying rationale for
maintaining a slightly higher degree of redundancy stems
from the assumption that the environment behaves unpredictably and the number of disturbances e(c,ℓ) it brings about
affecting ongoing voting rounds ℓ, therefore, may vary
considerably. In contrast, the dimensioning model was
designed to gradually adjust the degree of redundancy
downwards, targeting cr(e(c,ℓ)), in line with the trend
perceived from the data held within the observation window.
The safety margin can therefore intuitively be seen as the
maximum aberration in terms of additional disturbances that
the scheme can tolerate as compared with the observed trend.

The number ru expresses the maximum number of successive
voting round completions that failed to meet the criterion of
success as defined hereabove, before responding by
considering the use of additional redundancy. Such scenario
includes (1) voting rounds for which a result o could be
(c, ℓ )
adjuducated, yet the cardinality c max of the corresponding
consensus block [o] was insufficient to math the agreed safety
margin csm, and (2) rounds for which the decision algorithm
failed to adjudicate a result ― a case the model will endure at
most rf times within the observation window in spite of undershooting the required amount of redundancy. At risk of
prolonging the scheme's unavailability, temporarily refraining
from increasing the employed degree of redundancy after
observing a potentially hazardous situation might allow the
replica selection model to regain the scheme's intended
dependability by substituting poorly performing versions by
more reliable, idling versions. Smaller values ru would enable
a more rapid detection of potentially hazardous situations,
resulting in system resources to be allocated rather lavishly.
The data structure shown in Fig. 2 is updated upon each
subsequent completion y of a voting round, as represented by
the corresponding element in {yz}c. Each column refers to a
single voting round ℓ = bc(y) and holds information regarding
the redundancy n(c,ℓ) employed and its effectiveness to
counterbalance the disturbances to which it was perceived to
be subject ― information that can be deduced from the
( C ,ℓ )
generated partition ℘ . Information pertaining to round
bc( y z − r d ) will be discarded for values z > rd (see left in Fig. 2).

4.3. Replica Selection Model
Having established the degree of redundancy n(c,ℓ) to be
employed throughout round (c,ℓ), the replica selection model
will then determine an adequate selection of versions V(c,ℓ) to
be used by the redundancy scheme. The proposed model has
been designed so as to achieve an optimal trade-off between
dependability as well as performance-related objectives such
as load balancing and timeliness. To do so, a score is
computed for all versions v ∈ V depending on the contextual
information accrued during previously completed voting
rounds. For more information on this procedure, we refer to

4.2.2. Window Semantics
We will now elaborate on the procedure used to determine the
amount of redundancy to be used throughout a newly
initialised voting round (c,ℓ). In doing so, we use the abstract

4

its initial announcement in [3], which also introduces a
valuable metric named normalised dissent for approximating
the reliability of versions v ∈ V. Moreover, it aims to mitigate
the adverse effects of employing inapt resources that
consistently perform poorly and that, consequentially, may
threaten the effectiveness of the overall redundancy scheme.
If the degree of redundancy n(c,ℓ) to be utilised follows a
constant or decreasing trend, then, depending on the availability of eligible versions that can be used as a substitute, the
model will be successful in excluding such inapt replicas.

discrete-event simulation toolbox and its failure injection
mechanisms that were developed in [2]. Even though the
toolbox provides a multitude of failure injection mechanisms,
a simplistic, trend-based injection mechanism was chosen for
visualisation purposes, as illustrated in the upper graph in
Fig. 3. The total number of randomly chosen versions v ∈ V
for which the system will inject failures in the course of the
corresponding voting round (c,ℓ) is represented by the blue
marks, whereas the green marks indicate how many of those
affected versions were selected for participation in V(c,ℓ) ⊆ V.
Injected failures are set to materialise as EVF and RVF
content failures with a 20, respectively 80% probability. The
discrepant response values returned for invocations affected
by RVF failures are sampled from a uniform distribution [3].
Replicas are selected using the model referred to in Sect. 4.3
which has been configured to target sustained availability.

5. Parsimonious Resource Allocation
Building on the extensible and abstract concept of window
semantics, we will now briefly discuss two specific
implementations that will be used to validate the effectiveness
of the redundancy dimensioning model in Sect. 6.
A first strategy was originally published in [7]. Even though it
was applied on redundant data structures, it can readily be
reused for dynamically determining the redundancy level. It
assumes nmax = 9, and will only report an odd degree of
redundancy. Moreover, the redundancy scheme is initialised
such that it is capable of tolerating up to one failure, hence
ninit = nmin = 3. If the voting scheme failed to find consensus
amongst a majority of the replicas involved during the last
completed voting round, the model will increase the number
of redundant replicas to be used in the next voting round, to
the extent that fu(c,ℓ) = 2. Conversely, when the scheme was
able to produce an outcome with a given amount of redundancy for a certain amount rd of consecutive voting round
completions, a lower degree of redundancy shall be used for
the next voting round, with fd(c,ℓ) = 2. Note how this strategy
restricts the success criterion (SC) by requiring the same
amount of redundancy to be used in the observation interval.

Figure 3: Number of injected failures (above); evolution of
degree of redundancy and its effectiveness (below).
We now compare the effectiveness of four redundancy
dimensioning strategies. Each of these strategies is assumed
to be deployed within an A-NVP/MV scheme operating in an
environment in which the same set V of 20 versions had been
deployed and that exhibits identical failure behaviour as
modelled by the blue trend line in Fig. 3. At any time, no
more than four failures are assumed to affect the versions in
V(c,ℓ), a scenario that could successfully be overcome by a
static redundancy configuration with n = 9. Apart from such
classic strategy, we will evaluate the two strategies for
parsimonious resource allocation introduced in Sect. 5: the
former will be analysed without any safety margin applied,
the latter considering two distinct values csm:

A second strategy is a plain implementation of the mechanism
defined in Sect. 4.2.1. When the degree of redundancy n(c,ℓ-1)
is found to be overabundant, as can be perceived from the
success criterion (SC), the downscaling function fd(c,ℓ) will
try to adjust downwards by an amount of (c(c,ℓ x ) − m (c,ℓ x ) ) ‒csm,
max
with ℓx = bc(yz) the last completed round observed. Undershooting the safety margin will cause fu(c,ℓ) to return
csm ‒ | c(c,ℓ x ) − m (c,ℓ x ) | , with ℓx the round in the observation
max
window delimited by ru exhibiting the largest deviation from
the imposed security margin. The upscaling function fu(c,ℓ)
(c,ℓ )
(c, ℓ )
will try to inflate n(c,ℓ-1) by | c maxx − m x | + csm in case of
redundancy undershooting, with ℓx the eligible round with the
smallest degree of consent.

strategy
A: Classic
B: First
C: Second
D: Second

6. Effectiveness Analysis

csm
0
1
2

rd
20
20
20

#ℓF
0
5
4
0

ttf
∞
47
47
∞

∑cr(c,ℓ)
724
614
664
716

∑n(c,ℓ)
1800
1358
1482
1699

ratio
2.486
2.211
2.232
2.373

Table 1: Overall resource consumption and ineffectiveness.
Number of voting rounds with scheme failure #ℓF.

Given the availability of spare system resources, our redundancy dimensioning model is indubitably capable of scaling
up the employed degree of redundancy, either in response to a
failure of the scheme, or as a precautionary measure if the
effectiveness of the employed redundancy is observed to
deteriorate ― cf. Sect. 4.2.2. In this section, we will analyse
whether the proposed model can effectively and safely reduce
the employed redundancy. In doing so, we have used the

So as to make a fair comparison, the application-specific
parameters are configured with ninit = nmax = 9 and nmin = 3.
For the sake of simplicity, it is assumed that the order in
which voting rounds are initialised is the order in which they
complete, i.e. z = x for all ℓx = bc(yz), and that all context
information is available by the time the next round is
initialised. This assumption improves comprehension and

5

allows the reader to grasp the intuition behind the harvested
contextual information c(c,ℓ ) − m (c,ℓ ) (represented by the circles
max
in the lower graph in Fig. 3) driving the adjustment of the
employed redundancy. The value rd = 20 has been chosen for
demonstration purposes so as to enable the model’s
effectiveness to be concisely captured (e.g. strategy C in
Fig. 3). At the expense of an aggressive allocation strategy,
we set ru = 1 such that the availability of the scheme can
swiftly be regained in case of redundancy undershooting.

technology for implementing fault containment units. A
prototypical service-oriented implementation of the replica
selection algorithm showed how WS-* specifications can be
leveraged to sustain adaptive redundancy management,
broadening the applicability of NVP schemata by increased
interoperability [3]. This A-NVP/MV prototype was extended
to include the proposed redundancy dimensioning model and
policies defined in Sect. 4 and 5.

7. Conclusion

For each voting round ℓ in the sequence {ℓx}c, we will
consider e(c,ℓ), that is, the number of versions vi ∈ V(c,ℓ) that
are affected by some type of content failure. Recall that the
green crosses in the upper graph of Fig. 3 essentially quantify
e(c,ℓ). We can then determine the contextual redundancy
cr(e(c,ℓ)) and compare this value with n(c,ℓ) to observe if the
amount of redundancy n(c,ℓ) actually used was overabundant
or insufficient ― cf. Sect. 4.2. As shown in Table 1, the
relevant cumulative (contextual) redundancy during the
operational life span of the scheme provides vital information
regarding the effectiveness of a redundancy dimensioning
model in estimating the extent to which it can economise on
resource expenditure. One cannot merely judge a dependability strategy in terms of cumulative contextual redundancy;
though sharing identical failure behaviour (blue trend), the
actual number of injected disturbances (green marks) for any
specific voting round is eventually determined by the amount
of redundancy that is actually used ― hence the variation in
the values reported. Instead, one had better considered the
ratio of total cumulative redundancy over cumulative contextual redundancy, for which the actual number of disturbances
is indirectly accounted for. Though smaller, positive values
are indicative of an increased level of parsimony, one should
consider whether this reduction in resource allocation did not
violate the dependability objective. In this respect, strategy B
clearly performs worse than A, C and D despite the smallest
degree of resource consumption, as the scheme experienced a
failure in 5 of the 200 simulated voting rounds. This
behaviour could have been anticipated, as strategy A responds
to actual failures rather than proactively adjusting the
redundancy upwards were the agreed safety margin violated.
Observe the significant improvement with respect to the
number of scheme failures resulting from a modest increase
of the imposed safety margin ― cf. strategies C and D.

In this paper, we have enhanced the redundancy management
facilities of our A-NVP/MV dependability strategy, which
had initially appeared in [3]. The proposed redundancy
dimensioning model aims to autonomously tune the employed
degree of redundancy in view of encountered disturbances,
and is fitted with accompanying policies intent upon
increasing the scheme’s cost effectiveness by parsimoniously
allocating system resources. Based on the discrete-event
simulation models presented in [2] with special attention paid
to the impact of disturbances that manifest as transient
failures in the content domain, these policies haven proven to
be effective in identifying situations necessitating an
adjustment of the redundancy level. It is apparent from our
experimentation that the suggested solution can effectively
achieve a substantial improvement in dependability,
compared to traditional, static redundancy strategies. Furthermore, tuning the adopted degree of redundancy to the actual
observed disturbances allows unnecessary resource expenditure to be reduced, therefore enhancing cost-effectiveness.
As future work, we intend to analyse the impact of the
defined parameters and investigate whether the system could
tolerate graver whimsicality of its deployment environment
by dynamically altering the parameters rd, ru and rf, and how
this may affect the effectiveness of the presented policies.
Moreover, we will renounce the simplification of in-order
voting round completions and analyse the effects of variations
in the individual response times of versions.

References
[1] M. Aigner. “Combinatorial Theory”, Springer (1997).
[2] J. Buys et al. “A Discrete-Event Model for Failure
Injection in Distributed NVP”, technical report (2012).
[3] J. Buys et al. “Towards Context-Aware Adaptive Fault
Tolerance in SOA Applications”, Proceedings of the 5th
ACM International Conference on Distributed EventBased Systems, pp. 63–74 (2011).
[4] F. Cristian. “Understanding Fault-Tolerant Distributed
Systems”, Communications of the ACM, 34(2),
pp. 56-78 (1991).
[5] B. W. Johnson. “Design and analysis of fault tolerant
digital systems”, Addison-Wesley Publishing (1989).
[6] V. De Florio. “Application-layer Fault-tolerance”, IGI
Global (2009).
[7] V. De Florio. “Software Assumptions Failure Tolerance:
Role, Strategies and Visions”, Architecting Dependable
Systems, LNCS 6420, pp. 249–272 (2011).

In general, one may expect that a properly chosen value csm
matching the variability of the environment and the ensuing
disturbances aids in intercepting the trend and may lead to a
reduction of scheme failures due to more efficient proactive
upscaling. As can be seen from Fig. 3, a modest value csm = 1
enables the model to adjust the employed degree of
redundancy such that it follows the trend of failures ― cf. the
blue trend in the graph above with the employed redundancy
represented by the black marks in the lower graph.

7. Service-Oriented Prototype
The use of stateless web services and the confinement of any
communication to take place via explicitly defined service
interfaces appear to suggest that web services are an adequate

6

TOWARDS PARSIMONIOUS RESOURCE ALLOCATION IN CONTEXT-AWARE N-VERSION PROGRAMMING

More Related Content

Similar to TOWARDS PARSIMONIOUS RESOURCE ALLOCATION IN CONTEXT-AWARE N-VERSION PROGRAMMING (20)

More from Vincenzo De Florio (20)

Recently uploaded (20)

TOWARDS PARSIMONIOUS RESOURCE ALLOCATION IN CONTEXT-AWARE N-VERSION PROGRAMMING