An Epidemic Model for Highly Volatile Workloads:

Identification and Numerical Validation

Shubhabrata Roy, Thomas Begin and Paulo Goncalves
Inria, ENS Lyon, UCB Lyon 1, UMR 5668
46 Allee dItalie, 69364 Lyon Cedex 07, France
Email: {shubhabrata.roy, thomas.begin, paulo.goncalves}@ens-lyon.fr

I. Introduction
In recent trend of data-intensive applications with payas-you-go execution in a cloud environment, there are new
challenges in system management and design to optimize the
resource utilization. Types of the application, deployed in a
cloud, can be very diverse. Some applications exhibit highly
varying demand in resources. In this paper we consider a Video
on Demand (VoD) system as a relevant example of a dataintensive application where bandwidth usage varies rapidly
over time.
A VoD service delivers video contents to consumers on
request. According to Internet usage trends, users are increasingly getting more involved in the VoD and this enthusiasm
is likely to grow. According to 2010 statistics a popular
VoD provider like Netflix accounts for around 30 percent
of the peak downstream traffic in the North America and
is the largest source of Internet traffic overall [1]. Since
VoD has stringent streaming rate requirements, each VoD
provider needs to reserve a sufficient amount of server outgoing bandwidth to sustain continuous media delivery (we are not
considering IP multicast here). However, resource reservation
is very challenging in a situation, when a video becomes
popular very quickly leading to a flood of user requests on
the VoD servers. This situation, also known as a buzz,
demands an adaptive resource allocation strategy to cope with
the sudden (and significant) variation of workload. Following
is one example of buzz (see Figure 1) where interest over
a video Star Wars Kid [2] grew very quickly within a
short timespan. According to [3] it was viewed more than

900 millions times within a short interval of time making it

one of the top viral videos. Such bandwidth volatility creates
Number of downloads / day

AbstractIn a previous contribution, we proposed an

epidemic-based stochastic model particularly sound for reproducing volatile workloads time series. We used a Video on Demand
(VoD) service as a paradigm for applications that undergo
highly variable demand. In the present article, we proceed one
step further in the study of this model, and derive a heuristic
identification procedure to calibrate all the model parameters.
We evaluate the performance of our estimator on synthetic time
series. We also show how good can our model fit to real workload
traces with respect to the stationary case in terms of steady-state
probability and autocorrelation structure. Finally, we elaborate
on the way our work may contribute to capture the underlying
dynamic characteristics of an application, providing thus valuable
inputs to a resource management scheme.
Index TermsEpidemic Model; Statistical Estimation; Workload Volatility; Video on Demand; Buzz.

Time (days)
Fig. 1. Video server workload: time series displaying a characteristic

pattern of flash crowd (buzz effect). Trace obtained from [2].

significant challenges to meet, namely, both the desired QoS

and efficient resource allocation. A sensible approach to this
problem is to help the providers in better understanding and
capturing the underlying characteristics of their applications.
For example, if the information diffusion process follows a
gossip- (or epidemic-) behavior, the rate at which a viewer
keeps gossiping about a video and for how long (in average).
In [4], we followed a constructive approach to propose a
stochastic epidemic workload generator for a VoD system
based on a Markov model. We showed that it succeeds to
reproduce the traffic volatility, as exhibited in a real trace,
including the buzz occurrence. But the principal interest of
our model is that it verifies a Large Deviation Principle (LDP)
that gives a probabilistic description of the mean workload of
the system over different time scales. It thus adequately allows
for statistically characterizing extreme rare events such as the
ones produced by buzz transients. Our ultimate objective is
to exploit this large deviation information as an input of an
probabilistic resource management scheme. However, in order
the proposed model to conform with this objective, it needs
to be identifiable and easily calibrated on real data. The
corresponding estimation procedure may not be trivial, since
the VoD model is a non-parsimonious model and accounts for
complex dynamics.
In this paper we advance our work in [4] by a thorough study

of the proposed model to reproduce workload variability of

a data intensive application like a buzz. Our contributions
include:
A heuristic identification procedure for the proposed
workload model,
A numerical evaluation of the estimator for each parameter of the model,
A real case study to assess the adequacy of our model to
fit video workload traces.
Moreover, since we followed a constructive approach, each
parameter of the model accounts for a specific component of
the system, and so, its estimated value also permits to quantify
the importance of the corresponding dynamic effect.
Rest of the paper is organized as follows. In Section II we
discuss the related works. We briefly describe our model
and further analyze it in Section III. Section IV outlines the
parameter estimation procedure and validates the procedure
against synthetic workload traces. In Section V we validate
both our model and the estimation procedure against the real
workload traces. Finally we conclude and discuss future works
in Section VI.
II. Related Work
Information dissemination in a VoD system has been an
active area of research. In [5], it has been already demonstrated
that the epidemic algorithms can be used as an effective
solution for information dissemination in a VoD like P2P
systems. However, in this model an individual process must
have a precise idea about the total number of processes in
the system. Scalability is also another challenge that the
authors addressed in this work. The authors of [6] studied
random epidemic strategies like the random peer, latest useful
chunk algorithm to achieve optimal information dissemination.
But main objective of this work is to demonstrate ways to
achieve performance trade-offs using unstructured, epidemic
live streaming systems. However, it does not bring any information about the underlying dynamics of the streaming
system. Authors of [7] similarly discussed an analytical framework for gossip protocols based on the pairwise information
exchange between interacting nodes. However, this model
only provides an analysis of networks with lossy channels.
Another relevant work to our study is derived in [8] where
the authors proposed an approach to predict workload for
cloud clients. They considered an auto-scaling approach for
resource provisioning and validated the result with real-world
cloud client application traces. However, this work depends
on similar past occurrences of the current short-term workload
history and is not appropriate to deal with sudden and short
large variations of workload, as the ones produced by buzz
effects. Authors of [9] show a statistical study of streaming
traffic. They analyzed VoD session characteristics, amount and
types of media delivered, popularity and daily access profile
in order to develop a workload generator. However, the model
does not involve the dynamics of the process itself, ergo it
is not naturally adapted to infer dynamic resource allocation
strategies. Authors of [10], [11] and [12] also develop user

activity models to describe the usage of system resources. Limitation of these models are that they only give average results.
However, dealing with mean workloads might not be sufficient
to clearly describe applications because of their potential
volatility. In [13] authors proposed a maximum likelihood
method for fitting a Markov arrival process (MAP) to the web
traffic measurements, collected in commonly available HTTP
web server traces. This method achieves reasonable accuracy
in predictive models for web workloads but lacks intuitive
nature to describe users behavior like a gossip based method.
In [14] the authors statistically model traffic volatility in large
scale VoD systems using GARCH (generalized autoregressive
conditional heteroscedasticity) process. Amazon Cloud-Watch
follows this approach and provides a free resource monitoring service to Amazon Web Service customers for a given
frequency. Based on such estimates of future demand, each
VoD provider can individually reserve a sufficient amount of
bandwidth to satisfy in average its random future demand
within a reasonable confidence. However, according to the
authors, this technique only models and forecasts the mean
demand, or the expected demand whereas the real demand
might vary around this predicted mean. They suggested to
provision an additional risk premium to the service providers
for tolerating the demand fluctuation. In another workload
model the authors of [15] [16] proposed a Markov Modulated
Poisson Process (MMPP) based approach for buzz modeling
and then parameter estimation using the index of dispersion.
However, the MMPP model includes only short-term memory
in the system and the obtained statistics is not physically
interpretable to draw inference about the system dynamics.
The model we derive in [4] has the following advantages:
It follows a constructive approach, based on a Markov
model,
It is identifiable and succeeds to capture workload volatility,
It satisfies the large deviation properties, that can be
exploited to frame dynamic resource allocation strategies.
III. Proposed Video on Demand (VoD) Model
Following the lines of related works, we rely on epidemic
models to represent the way information spreads among the
users (gossip-like phenomenon) in a VoD system. Epidemic
spreading models commonly subdivide a population into several compartments: susceptible (noted S) to designate the
persons who can get infected, and contagious (noted C) for
the persons who have contracted the disease. This contagious
class can further be categorized into two parts: the infected
subclass (I) corresponding to the persons who are currently
suffering from the disease and can spread it, and the recovered
class (R) for those who got cured and do not spread the
disease anymore [17]. In these models S (t)t0 , I(t)t0 and
R(t)t0 are stochastic processes representing the time evolution
of susceptible, infected and recovered populations respectively.
In the case of a VoD system, infected I refers to the people
who are currently watching the video and can pass the
information along. In our setting, I(t) directly represents

the current workload which is the current aggregated video

requests from the users. Here, we consider the workload as
the total number of current viewers, but it can also refer to
total bandwidth requested at the moment. The class R refers to
the past viewers. In contrast to the classical epidemic case, one
originality of our approach stems from introducing a memory
effect in our model, assuming that the R compartment can still
propagate the gossip during a certain random latency period.
Then, within a small time interval dt, the probability for a
susceptible individual to turn into an active viewer reads:
PS C = (l + (I(t) + R(t)) )dt + o(dt)

(1)

where > 0 is the rate of information dissemination per unit

of time and l > 0 fixes the ingress rate of spontaneous viewers.
At time t, the instantaneous rate of newly active viewers in
the system:
(t) = l + (I(t) + R(t)),
(2)
corresponds to the arrival rate of a non-homogeneous (statedependant) Poisson process and it varies linearly with I(t) and
R(t). When  l, the arrival process induced by peer-topeer contamination dominates the workload increase, whereas
it is the spontaneous viewers arrival process that prevails when
l  .
Regarding the sojourn time in the (I) compartment, we
assume that the watching time of a video is an exponentially
distributed random variable with mean value 1 , meaning that
viewers leave to the (R) class at rate . As already mentioned,
it also deems reasonable to consider that a past viewer will not
keep propagating the gossip about a video indefinitely. Instead,
users remain active only for a latency random period that we
also assume exponentially distributed with mean parameter
1 , after which they leave the system (at rate ). In general
watching time of a video is much smaller compared to the
memory persistence, leading to  .
Another innovation of our model lies in the buzz generating
mechanism: we resort to Hidden State Markov Chains to
randomly switch between the nominal (buzz-free) and the buzz
regimes. While = 1 in the normal state, the propagation rate
(gossip) jumps to = 2  1 in buzz regime, triggering
thus a sudden and steep increase of the current demand.
Transitions between these two hidden and memoryless Markov
states occur with rates a1 and a2 respectively and characterize
the buzz in terms of frequency and duration (supposedly,
a1  a2 ). Conceptually, the model easily generalizes to
several states to reproduce different types of buzzes, but for
the sake of identifiability, we restrict ourselves to a two
hidden states model only. With these assumptions, and posing
(I(t) = i, R(t) = r) the current state, Figure 2 shows the
transition diagram of the corresponding Hidden State Markov
Chain.
A closed-form expression for the steady state distribution of
the workload (i) of this model is not trivial to derive. However,
we could easily express the analytic mean workload of the
system solving the flow balance equation, i.e. equaling the
incoming and outgoing flow rates in steady regime. For ease,

(i+r)+l

i, r

= 1

i+1, r
r

a1

= 2

i-1, r+1

i, r-1

a2

Fig. 2. Hidden State Markov Chain diagram representing the possible

transitions of the number of current (i) and past active (r) viewers.

we start with = 1 = 2 and generalize the result to 1 , 2

thereafter. We get:
E(i) =

l
,

(3)

which, to be a positive and finite quantity, yields the stability

criterion in buzz-free regime:
1 > 1 + 1 .

(4)

We now extend these results to the case where the model may
exhibit a buzz activity. As alternates between the hidden
states = 1 and = 2 , with respective state probabilities
a2 /(a1 + a2 ) and a1 /(a1 + a2 ), one can simply replace in Eq.
(3) and (4) with the equivalent average value:
1 a2
2 a1
=
+
.
a1 + a2 a1 + a2

(5)

In order to illustrate the flexibility of our workload model

and to validate Eq. (3), we generate three synthetic traces
corresponding to the different sets of parameters verifying
the stability condition of relation (4) and reported in Table
I. Particular realizations of these processes generated over 221
points are displayed in Figure 3. While the synthetic traces
TABLE I
Parameters value used in the workload model to generate the three traces
plotted in Fig. 3. The last two rows correspond to the theoretical mean
workload of Eq. (3) and to the sample mean value estimated from the traces.
1
2

l
a1
a2
E(i)
Emp. mean hii

case (a)
4.762 104
0.0032
0.0111
5 104
104
107
0.0667

case (b)
3.225 105
0.0032
0.0020
3.289 105
104
107
0.0667

case (c)
2.439 105
0.0032
0.0011
2.5 105
104
107
0.0667

1.92
1.74

15.68
16.72

44.72
45.23

corresponding to cases (b) and (c) reproduce distinct and easily

identifiable buzz regimes, the parameter set of case (a) leads
to a workload variation distinct from the typical shape of
Figure 1. Nonetheless, for all 3 configurations, the empirical
means estimated from the 221 samples of the traces are in good
agreement with the expected values of Eq. (3).
Finally, let us notice that even though we consider exponentially distributed random variables in our model, any other
distributions could be used, which, according to the same

balance principle, would lead to a mean workload and to a

stability condition of the same kind as (3) and (4). However,
the estimation procedure we derive in the next section strongly
relies on the exponential assumption and it would need to be
thoroughly reworked to adapt to different hypotheses.
IV. Model Calibration
In this section, we address the identifiability of our model
and design a calibration algorithm to fit workload data. We
start constructing empirical estimators for each parameter of
the model and we numerically evaluate their performance on
synthetic traces.
A. Parameters estimation
Considering a standard epidemic process X with propagation rate , the maximum likelihood estimate b
MLE is derived
in [17], [18] and reads:
b
MLE = n

!1

X(t) dt

(6)

where n is the number of contaminations (i.e. number of

increments of X) occurring within the time interval T .
Very often, maximum likelihood approach yields optimal
results (in terms of estimate variance and or bias) but it
is not always possible to get a closed-form expression for
the estimated parameters. This can either be due to the
likelihood function that is impossible to derive analytically, or
to missing data that preclude straightforward application of the
maximum likelihood principle. Nonetheless, solutions, such
as the Expectation-Maximization (EM) or the Monte Carlo
Markov Chain (MCMC) algorithms exist, which in some cases
can approximate maximum likelihood estimators.
Returning to our model depicted in Figure 2, each
parameter needs to be empirically estimated, assuming that
the instantaneous workload time series is the only available
observation.
Watching parameter . As is the departure rate of users that
leave the infected state after they finished watching a video,
it can directly be inferred from the number n of decrements
of the observable process I(t). Therefore, the MLE of Eq. (6)
straightforwardly applies and leads to:
b
MLE = n

!1

I(t) dt

(7)

Memory parameter . This rate at which past viewers leave the

recovery compartment and stop propagating the virus (gossip),
relates to the decrement density of the non-observed process
R(t). It is thus impossible to simply apply the MLE of Eq.
b to the missing
(6) unless we first construct a substitute R(t)
data from the observable data set I(t). Let us recall that in our
model, all current viewers turn and remain contagious for a
mean period of time 1 + 1 . Then, in first approximation,

we can consider that R(t) derives from the finite memory

cumulative process:
Z t
b
R(t) =
I(u) du,
(8)
t(1 +1 )

which itself, depends on the parameter to be estimated . We

propose an estimation procedure based on the inherent exponential property of the model. From the Poisson assumption,
the inter-arrival time w between the consecutive arrivals of two
new viewers is an exponentially distributed random variable
such that E (w| I(t) + R(t) = x) = ( x + l)1 . Its means that,
e = w/E(w|x) is
for x fixed, the normalized random variable w
exponentially distributed with unitary parameter and becomes
independent of x. Ideally then, for each value of R(t)+I(t) = x,
all the sub-series w x = {wn : R(tn ) + I(tn ) = x}, after
normalization by their own empirical mean, yield independent
and identically distributed realizations of a unitary exponential
random variable. In practice though, as R(t) is not observable,
b is accurately estimated, should this unitary exonly if R(t)
ponential i.i.d. assumption hold true. From there, we propose
the following algorithm: for different values of spanning a
b (t) estimated from Eq. (8) to
reasonable interval, we use R
e . A statistical test applied to
build the normalized series w
e allows for assessing the exponential i.i.d. hypothesis
each w
and then to select the value of that yield the best score.
e = (e
wn )n=1,...,N the statistiMore concretely, we apply to w
cal exponentially test derived in [19]: Form the normalized


e(n1) ) n=1,...,N where
spacings v = v(n) = (N n + 1)(e
w(n) w


e rearranged in ascending order. Let
e(n) n=1,...,N stands for w
w
e
F and G denote the cumulative distribution functions of w
and v respectively, and compute the classical KolmogorovSmirnov distance:
r
1
sup |F(k) G(k)|.
(9)
T =
N 1kN
As F and G are identical for an exponentially i.i.d. random
series, we then expect T to reach its minimum for the value
b (t) of R(t):
of that gives the best estimate R

= argmin T

b
(10)

R
b= R
bb .
Plots of Figure 4 show the evolution of the KolmogorovSmirnov distance corresponding to the traces displayed in
Figure 3. In the 3 cases, T clearly attains its minimum
bound for b
close to to the actual value. The corresponding
b derived from Eq. (10) match fairly
estimated processes R(t)
well the real evolution of the (R) class in our model (see
Figure 5).
Propagation parameters and l. According to our model, the
arrival rate (t) of new viewers is given by Eq. (2). It linearly
depends on the current number of active and past viewers. So,
b
from the observation I(t) and the reconstructed process R(t)
of Eq. (10), we could formally apply the maximum likelihood
Eq. (6) to estimate . In practice however, we have to bear
in mind that: (i) the arrival process of rate (t) comprises a

case (a)

case (b)

case (c)

60

350

80
60

50

300

40

250
200

30
40

150
20
100

20
10
0
500

1000

1500

2000

2500

50

0
1498

1500

1502

1504

0
9326

9328

9330

9332

Fig. 3. Illustration of our model ability at generating different dynamics of workload I(t). See Table I for the parameter values corresponding

to each of these three cases. The Xaxis corresponds to time (in hours unit) while the Yaxis indicates the number of active viewers.

0.03
0.02

conditional empirical mean:

X
1
b n ) = x}. (11)
(x) =
wn : I(x) = {tn : I(tn ) + R(t
|I(x)| t I(x)

case (a)
case (b)
case (c)

0.01
0
6

10

10

10

(logarithmic scale)

10

Evolution of the exponential test statistics (9) applied to the

traces of Figure 3. Dotted vertical lines locate the actual value of
for each case; dot markers on each curve indicate the estimated value
b
corresponding to the minimum point of the statistical test T .
Fig. 4.

b
R(t), R(t)

900
850
800
750
700
1498

1499

1500

1501

1502

1503

1504

time (hrs)
Evolution of the number of active past viewers. Comparison
of the actual (non observable) process R(t) (blue curve) with the
b (red curve) derived from expression (8).
estimated process R(t)
Fig. 5.

spontaneous viewers ingress that is governed by parameter l

and which is independent of the current state of the system;
(ii) depending on the current hidden state of the model (buzzfree versus buzz state), it is alternately = 1 and = 2
that fix the propagation rate in Eq. (2). We designed an
estimation procedure based on a weighted linear regression,
that simultaneously addresses these two issues. We decompose
our rationale in two steps: First, let us consider the buzz-free
state only and = 1 . As discussed in the estimation of the
inter-arrival time w between the consecutive arrivals of two
new viewers is an exponentially distributed random variable
such that E (w| I(t) + R(t) = x) = ( x + l)1 . Concretely then,
b
for different values of the sum I(t) + R(t),
we calculate the

The linear regression of ((x))1 against x yields at one go,

both parameters estimation b
(slope) and b
l (intercept).
Let us now return to the general form of our model with
alternation of buzz and buzz-free periods. In the buzz-free
case, = 1 corresponds to a normal workload activity,
b takes on rather moderate
meaning that the sum I(t) + R(t)
values. Conversely, when the system undergoes a buzz, = 2
b suddenly increases to reach
and the population I(t) + R(t)
significantly larger values. Yet, in both cases, the quantity 1
defined in Eq. (11) remains linear with x but with two different
b = x.
regimes (slopes) depending on the amplitude of I(t)+ R(t)
As a result, it is possible to reduce the bias that 2 causes on
the estimation of 1 , using a weighted linear regression of 1
vs x where the weights p(x) are proportional to the cardinal
of the indicator sets I(x). Indeed, |I(x)| should be smaller for
larger values of x because buzz episodes are expected to be
less frequent than nominal activity periods. Figure 6 confirms
the claim: the plots (x, 1 ) show a manifest linear trend with
higher variability at xs large, meaning a fewer terms entered
the sum of Eq. (11).
Formally, we can apply the exact same procedure to
estimate 2 , but considering opposite weights to favor the
large values of xs. However, due to the large fluctuations of
((x))1 in the corresponding region, the slope b2 is subject
to a very poor estimation variance. Instead, we propose
to apply the ML estimator described in Eq. (6) on the
restriction of I(t) to the buzz periods only. Strictly speaking,
b as well, but since a buzz event
we should consider R(t)
normally occurs on very small interval of time, we assume
b (resp. R(t)) remains constant in the meanwhile
that R(t)
(flash crowd viewers will enter in R compartment only
after the visualization time). In practice, to automatically
identify the buzz periods, we threshold I(t) and consider only
the persistent increasing parts that remain above the threshold.
Transition rates a1 and a2 . As we already said, at time
t, the inter-arrival time w separating to new incomers is a
random variable drawn from an exponential law of parameter

0.45
case (a)
case (b)
case (c)

0.4
0.35

((x))1

0.3
0.25
0.2
0.15
0.1
0.05
0
0

500

1000

1500

b
x = I(t) + R(t)

2000

2500

Weighted linear regression of 1 vs x corresponding to the

three traces of Figure 3. Superimposed are the linear trends fitted on
the respective data.
Fig. 6.

= (i + r) + l, where I(t) + R(t) = i + r and is either equal to

1 or to 2 . We denote f1 (w) and f2 (w) the corresponding
b and the
densities built upon the reconstructed process R(t)
estimated parameters (b1 , b
l) and (b2 , b
l) respectively. For a given
inter-arrival time w = wn observed at time tn , we form
the likelihood ratio f2 (wn )/ f1 (wn ) to determine whether the
system is in buzz or in buzz-free state. Moreover, in order to
avoid non-significant state transitions we resort to a restoration
method inspired by the Viterbi algorithm [20]. Once we have
identified the hidden states of the process, we estimate the
transitions rates ab1 and ab2 from the average times spent in
each state.
B. Numerical Validation
To evaluate the statistical performance of our estimation
procedure, we resort to numerical experiments to empirically
get the first and the second order moments of each parameter
estimator. Owing to the versatility of our model, we must
ensure that the proposed calibration algorithm performs well
for a variety of workload dynamics. To this end, we systematically reproduce the experiments considering the 3 sets of
parameters reported in Table I. For each configuration, we
generate 10 independent realizations of processes similar to the
ones depicted in Figure 3, and use these to derive descriptive
statistics.
The box-and-whisker plots of Figure 7 indicate for each
estimated parameter (centered and normalized by the corresponding actual value) the sample median (red line), the
inter-quartile range (blue box height) along with the extreme
samples (whiskers) obtained from time series of length 221
points each. As expected (owing to the maximum likelihood
procedure), estimation of shows to be the most accurate, both
in terms of bias and variance. But more surprisingly though,
although the estimation b1 derives from a heuristic procedure
b of Eq. (8),
that itself depends on the raw approximation R(t)
the resulting performance is remarkably good: bias is always

negligible (less than 5% in the worst case (c)) and the variance
always confines to 10% of the actual value of 1 . Notice also
that the estimation of 1 goes from a slight underestimation in
case (a) to a slight overestimation in case (c), as the buzz
effect, i.e. the value of 2 , grows from traces (a) to (b).
Compared to b1 , the estimation of 2 behaves more poorly
and proves to be the most difficult parameter to estimate. But
we have to keep in mind that this latter is only based on buzz
periods which represent only a small fraction of the entire
time series. Regarding the parameter , its estimation remains
within a 20% inter-quartile range but cases (a) and (c) show
a systematic bias (median hits the lower quartile bound). Let
us then recall that the procedure, described by Eq. (10) to
determine b
selects within some discretized interval, the value
of that yields the best T score. It is then very likely that
the true value does not coincide with any sampled point of
the interval and therefore, the procedure picks the closest one
that systematically lies beneath or above. Finally, estimation
of the transition parameters a1 and a2 between the two hidden
states relies on all other parameters estimation, cumulating so
all relative inaccuracies. Nonetheless and despite a systematic
underestimating trend, precision remains within a very acceptable confidence interval. Convergence rate of the empirical
estimators is another important feature that binds the estimate
precision to the amount of available data. Using the same data
set, the bar plots of Figure 11 depicts the evolution of the mean
square error MSE(b
) = E{(b
)2 } where generic stands
for any parameter of the model with the length N of the
observable time series I(t). As our purpose is to stress the
rate of convergence of these quantities towards zero, to ease
the comparison, we normalize the MSE of each parameter by
its particular value at maximum data length (i.e 221 points
here). Then, the estimator rate of convergence corresponds
to the decaying slope of the MSE with respect to N in a loglog plot, i.e. MSE(b
) O(N ). For the different parameters
of our model we obtain convergence rates that lie between
1 = 0.9 and a2 = 0.2, leading each time to sub-optimal
convergence ( < 1). It is worth noticing that, despite its
relatively ad hoc construction, the estimator of 1 has an
almost optimal convergence rate, which proves the rationality
of our approach.
V. Model Fitting to Real Workload Data
We now apply the calibration procedure detailed in the
previous section, to fit our model on real data and to assess
the data-model adequacy in the context of highly volatile
workloads. As a paradigm for variable demand applications,
we use a VoD trace, released by the Greek Research and
Technology Network (GRNET) VoD servers [21]. Since the
trace shows modest activity with a workload level that is
not sufficient to stress-test our system, we scale up the data,
shrinking all inter-arrival times by a factor of 10. The resulting
workload time series displayed in Figure 9, clearly shows two
distinct periods of steady activity before and after the time
index t = 200. We consider the sub-series on both sides of this
cutoff time, as two individual workload processes, referred to

case (a)

case (b)

case (c)
0.3

0.3
0.2

0.2

0.2
0.1

0.1

0.1

0
0

0.1
0.1

0.2

0.1
0.2

0.3
0.2
0.3

a1

a2

0.4

a1

a2

Box-and-Whisker plots for relative estimation errors of the model parameters for the three different sets of prescribed parameters
reported in Table I. For each case (a)-(c), statistics are computed over 10 independent realizations of time series of length 221 points.

Fig. 7.

10
9

l
a1

log2 MSE(b
)

7
6
5

a2

4
3
2
1
0

12

15

18

21

log2 (N)
Evolution of the Mean Square Error versus the data length
N in a log-log plot. For the sake of conciseness, we only show here
the results corresponding to the case (b) of Table I. Corresponding
figures for the other two cases are deported in the Appendix section.
Fig. 8.

as trace I and trace II respectively, and we calibrate our model

on each of them separately.
Results of the parameters estimation are reported in Table
II, and we verified that in both cases the stability condition of
Eq. (4) was satisfied. In the same vein, we also compared the
empirical mean workload of each trace with its corresponding
theoretical value given by the formula (3). We obtain for
trace I a relative difference of 12% (E(i) = 5.59 compared to
hii = 4.99), and of 12.5% for trace II (E(i) = 0.621 compared
to hii = 0.71). Naturally, the correspondence here is not as
striking as it is with the synthetic traces of Section III. But
we must bear in mind that first, ab initio nothing guarantees
that the underlying system matches our model dynamics and,
second, traces I and II can possibly encompass short scale
non-stationary periods (e.g. day versus night activity) which
are not accounted for in our model. Notwithstanding this, the
match we observe is quite satisfactory and we now focus on
higher order statistics to further characterize the data model
adequacy. As we do not have a closed-form for the steady

state distribution of our Markov process model, nor we have

an analytic expression for its autocorrelation function, we use
the two sets of estimated parameters of Table II to synthesize
two time series that we compare to the real workload traces
I and II. We refer to those synthetic traces as to the fitted
traces I and II. The plots in Figure 10 show the empirical
steady state densities and the sample autocorrelation functions
of both the real and the fitted traces. The superimposition of
the different curves is a clear evidence of our model ability
at catching the statistical distribution of the number of current
viewers along time. But also, and perhaps more importantly,
it demonstrates that the dynamical mechanism underlying our
constructive model is able to perfectly reproduce the temporal
structure of the real traces, by imposing the correct statistical
dependencies between distant samples I(t) and I(t + ) of the
process.
In addition to serve as a mean to evaluate the goodness-offit of our model, the estimated parameters bring on their own,
a valuable insight about the system itself. For instance let us
compare the propagation rates b1 and b
l estimated from traces
I and II, successively. In the first case, b1 < l, meaning that
arrival of new viewers is dominated by spontaneous incomers
and is not so much due to information propagation through
gossip. Conversely, b1 > b
l for the second workload regime,
indicating that the spontaneous attraction of the server has
severely dropped whereas the peer-to-peer diffusion component significantly increased but not sufficiently to sustain the
mean workload activity. At the same time, the index b
tripled,
meaning that the mean memory period for propagation shrank
by a factor of 3. This parameter could then be used as an
indicator of the content interest delivered by the server, and
of its lifetime in users mind.
TABLE II
Estimated Parameters from traces I and II separately.

I
II

b1
0.0013
0.0049

b2
0.0084
0.0183

0.0039
0.0118

0.0028
0.0095

b
l
0.0032
0.0005

ab1
3.13.104
1.32.105

ab2
0.022
0.041

Steady-state distribution

Autocorrelation function

10

10

10

10

10

10

10

10

10

10

10

10

10

10

50

100

150

0.5

1.5

10

10

10

10

10

10

10

10

10

10

10

10

10

10

15

20

0.1

0.2

0.3

0.4

0.5

time lag (hrs)

Number of current viewers

Fig. 10. Comparison of the empirical steady-state distribution and of the autocorrelation function of the real (blue curves) and the fitted (red

curves) traces. Top two plots correspond to trace I and bottom plots correspond two trace II.

VI. Conclusion

Number of current viewers

140
120

Trace I

Trace II

100
80
60
40
20
0
0

50

100

150

200

250

300

time (hrs)
Fig. 9. Real workload time series corresponding to a VoD server
demand from [21]. Initial trace was scaled up by a factor of 10
to increase the mean workload. Trace is chopped into two separate
processes (Trace I and II) corresponding to different activity levels.

Many applications deployed on a cloud infrastructure, such

as a Video on Demand service, are well known for undergoing
highly volatile demand, making their workload hard to qualitatively and quantitatively characterize. With this objective,
and based on an existing constructive model, we devised an
heuristic identification procedure that aims at estimating the
parameters values of the model from a single collected trace.
First, we numerically evaluated the accuracy of this procedure using several synthetic traces. Our experiments show that
the procedure introduces little bias and typically recovers the
actual parameters value with a relative error of about 10%.
Second, we apply this same procedure against two real video
server workload traces. Obtained results demonstrate that,
once the model has been calibrated, it succeeds to reproduce
the statistical behavior of the real trace (in terms of both
the steady-state probabilities and the autocorrelations for the
workload time series).
Moreover, owing to the constructive nature of our model, the
estimated values of the parameters provide valuable insight on
the application that it would be difficult, or even impossible,
to infer from the raw traces. The captured information may

answer questions of practical interest to cloud oriented service

providers, like: is the application workload mostly driven by
spontaneous behaviors, or is it rather subject to a gossip
phenomenon?
In our future work we aim to exploit this procedure to
capture the large deviation properties of the corresponding
processes with final objective to devise probabilistic resource
management strategies. As a final remark, let us stress that the
model we propose is by no mean restricted to VoD workload
and can transpose to any volatile applications involving a
dissemination procedure.

[19] S.R. Jammalamadaka and E. Taufer, Testing exponentiality by comparing the empirical distribution function of the normalized spacings with
that of the original data, J. Nonparametric Statistics, vol. 15, pp. 719
729, 2003.
[20] J. Kleinberg, Bursty and hierarchical structure in streams, in Proc. the
8th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, 2002.
[21] GRNET, Video traces obtained from grnet, 2011, http://vod.grnet.gr/.

Appendix
10

Acknowledgments
This work has been supported by the EU FP7 project SAIL.
log2 MSE(b
)

[1] Sandvine, Sandvines spring 2011 global internet phenomena report

reveals new internet trends, May 2011, http://www.sandvine.com/news/
pr detail.asp?ID=312/.
[2] B. Andy, Star kids the data dump, http://waxy.org/2008/05/star
wars kid the data dump/.
[3] BBC, Star wars kid is top viral video, November 2006, http://news.
bbc.co.uk/2/hi/entertainment/6187554.stm/.
[4] P. Goncalves, S. Roy, T. Begin, and P. Loiseau, Dynamic resource
management in clouds: A probabilistic approach., IEICE Transactions
on Communications, special session on Networking Technologies for
Cloud Services, vol. E95-B(08), August 2012.
[5] P.T Eugster, R. Guerraoui, A. Kermarrec, and L. Massoulie, Epidemic
information dissemination in distributed systems, IEEE Computer
Society, vol. 37, no. 5, pp. 6067, May 2004.
[6] T. Bonald, L. Massoulie, F. Mathieu, D. Perino, and A. Twigg, Epidemic live streaming: Optimal performance trade-offs, ACM SIGMETRICS Performance Evaluation Review - SIGMETRICS 08, vol. 36, no.
1, pp. 325336, June 2008.
[7] R. Bakhshi, D. Gavidia, W. Fokkink, and M. van Steen, A modeling
framework for gossip-based information spread, in Proc. 8th International Conference on Quantitative Evaluation of SysTems, September
2011.
[8] E. Caron, F. Desprez, and A. Muresan, Pattern matching based forecast
of non-periodic repetitive behavior for cloud clients, Journal of Grid
Computing, vol. 9, no. 1, pp. 4964, March 2011.
[9] R. Garcia, X. Paneda, V. Garcia, D. Melendi, and M. Vilas, Statistical
characterization of a real video on demand service: User behaviour and
streaming-media workload analysis, Simulation Modelling Practice and
Theory, doi:10.1016/j.simpat.2007.02.004, vol. 15, pp. 672689, 2007.
[10] V.O.K. Li, W. Lao, X. Qiu, and E. W. M. Wong, Performance model of
interactive video-on-demand systems, IEEE JOURNAL ON SELECTED
AREAS IN COMMUNICATION, vol. 14, pp. 10991108, 1996.
[11] D. Melendi, R. Garcia, X. G. Paneda, and V. Garcia, Multivariate
distributions for workload generation in video on demand systems,
IEEE COMMUNICATIONS LETTERS, vol. 13, pp. 348350, 2009.
[12] S. Kanrar, Analysis and implementation of the large scale video-ondemand system, International Journal of Applied Information Systems,
vol. 2, no. 2, February 2012.
[13] S. Pacheco-Sanchez, G. Casale, B. Scotney, S. McClean, G. Parr, and
S. Dawson, Markovian workload characterization for qos prediction in
the cloud, in Proc. IEEE Cloud, June 2011.
[14] D. Niu, B. Li, and S. Zhao, Understanding demand volatility in large
vod systems, in Proc. NOSSDAV, June 2011.
[15] D. Perez-Palacin, J. Merseguer, and R. Mirandola, Analysis of bursty
workload-aware self-adaptive systems, in Proc ICPE, April 2012.
[16] R. Gusella, Characterizing the variability of arrival processes with
indexes of dispersion, IEEE JOURNAL ON SELECTED AREAS IN
COMMUNICATIONS, vol. 9, no. 2, February 1991.
[17] A. Barrat, M. Barthelemy, and A. Vespignani, Dynamical Processes on
Complex Networks, Cambridge University Press, 1st edition, November
2008.
[18] H. Andersson and T. Britton, Stochastic Epidemic Models and Their
Statistical Analysis, vol. 151, Springer, Lecture Notes in Statistics, 2000.

l
a1

7
6

a2
5
4
3
2
1
0

12

15

18

21

log2 (N)
10

l
a1

log2 MSE(b
)

References

a2
5
4
3
2
1
0

12

15

18

21

log2 (N)
Evolution of the Mean Square Error versus the data length
N in a log-log plot. Top: case (a) (top) ; Bottom: case (c) of Table I.

Fig. 11.