Informatica 40 (2016) 197–206 197
Protected Elastic-tree Topology for Survivable and Energy-efficient Data
Center
Dieu-Linh Truong, Elena Ouro and Thanh-Chung Nguyen
School of Information and Communication Technology
Hanoi University of Science and Technology, Vietnam
E-mail: linhtd@soict.hust.edu.vn, elenaouro@gmail.com, ntchung92@gmail.com
Keywords: data center, energy efficiency, survivability, elastic tree
Received: January 1, 2016
Recently, using third party data centers has become a popular choice for storing enterprise data and de-
ploying services. The fast growth of data centers in size and in number makes them become huge energy
consumption points. Up to 70% of energy consumption is due to server running and cooling. In order to
reduce energy consumption, recent researches have proposed turning off certain switches in data centers
with little traffic flow. However, when those switches are turned off, the data center becomes vulnerable to
failures due to low connectivity between servers. In order to overcome this weakness, this paper proposes
to use path protection to ensure that all connections in the data center retain survivability upon any single
failure. The paper also proposes an algorithm to calculate a tailored topology for the data center so that
unnecessary switches can still be turned off. The simulation results show that the proposed solution makes
data centers survivable while still saving energy significantly, mostly in big size data centers.
Povzetek: Nov algoritem poišče primerne varne rešitve v podatkovnih centrih pri varčevanju z energijo na
osnovi elastične topologije.
1 Introduction
Cloud computing is currently the common choice for end
users and enterprises to store their data and process their
services by using third party data centers. End users and
enterprises can now focus on their business issues with-
out concerning themselves with the building and maintain-
ing of their own storage servers or network devices since
this infrastructure is hosted in data centers. This advantage
leads to a quick growth of data centers in size and in num-
ber. However, these data centers consume a huge amount of
energy for running servers, network devices in order to pro-
cess user requests and transmit data between servers within
data centers. According to US Federal Energy Manage-
ment Program report in [1], data centers in 2013 accounted
for 2.7% of the 3831 billion kWh used in the US. Federal
data centers used about 5 billion kWh in 2013, or nearly
10% of federal electricity use. These numbers show the
importance of utilising energy efficiently in data centers.
It is stated in [2] that cooling infrastructure of data center
facilities can require one to two times the energy used to
power the IT equipment itself. Therefore, many researches
look for efficient energy utilisation solutions for data cen-
ters either by locating data centres close to green energy
resources, using energy efficient devices [3], or by limiting
the number of running devices in data center.
Servers in a data center interconnect through a system of
switches. Commonly, these switches are organised in hier-
archical topology with 3 layers: a core layer, an aggrega-
tion layer and an edge layer. The switches in the edge layer
link to servers (hosts). In upstream direction, traffic from
source servers arrives at edge switches then is regrouped
and routed to the aggregation layer switches. In turn, the
aggregation layer switches regroup traffic to forward to a
core switch. Traffic follows the downstream direction sim-
ilarly to get to the destination servers.
Fat-tree topology [4] was proposed for data centers for
the first time in 2008 [5]. A k-ary Fat-tree contains
k modular data centers called PODs (Performance Opti-
mized Data center) that links together through core layer
switches. Each POD contains two layers of aggregation
and edge switches where each switch of one layer links to
all switches of the other layer. Figure 1 shows an exam-
ple of 4-ary Fat-tree data center. Fat-tree structure has been
developed to reduce the over-subscription ratio and to re-
move the single point of failures of the hierarchical archi-
tecture in the data center network through multiple-linking
of a switch to other layers.
Fat-tree data centers tend to consume a lot of energy,
however. Data centers in general, and Fat-tree data centers
in particular, are usually designed with sufficiently high ca-
pacity to tolerate access traffic in peak hours, meanwhile
the access traffic varies significantly within a day, a week,
or month. As a result, data centers often work under ca-
pacity during normal hours. It is wasteful if all devices in a
data center are kept working during this time.
With the objective to save energy a solution, Elastic tree,
has been proposed in which the main idea is to turn off
some devices in Fat-tree [6] when they are not needed. In
198 Informatica 40 (2016) 197–206 D.-L. Truong et al.
Figure 1: A Fat-tree data center with k = 4.
Elastic tree, the traffic is compressed to some switches and
links while the free switches and ports are turned off. The
topology undergoes dynamic changes in concert with the
change of traffic. That is why the topology is called Elastic
tree. Elastic tree certainly saves more energy than Fat-tree.
The idea of Elastic tree has also been developed further in
RA-TAH [7] where besides turning off some devices, some
others can also be put into sleep mode or into a reduced
working rate for more efficient use of electricity. However,
both Elastic tree and RA-TAH make data center vulnerable
to failures due to low connectivities. For example, in Figure
2, after turning off 3 core switches and some aggregation
switches, the data center becomes a tree with a single root,
then there is a single path between any pair of servers. If
a failure occurs on any link between the core and aggrega-
tion layers, a whole POD will be disconnected. If a failure
occurs between the aggregation layer and the edge layer,
an edge switch with all its associated servers will be dis-
connected from the data center. Although it is possible to
wake up or to turn on the devices for taking over the traffic
from the failed switches/links, the recovery delay remains
important due to the slow starting/waking up process and
long traffic re-direction process.
In this paper, we propose to add the protection capability
to the Elastic tree so that even if some devices are turned
off, the data center remains survivable upon any single fail-
ure. The proposed architecture is called Protected Elastic
tree. We focus on protecting the communication between
servers within the data center only. The main contributions
of the paper are: i) proposition for the use of a path pro-
tection model in Elastic tree, ii) calculation of the mini-
mum energy consumption topology for the Protected Elas-
tic tree, and iii) analysis of the impact of different factors
on the energy saving capability of the Protected Elastic tree
topology.
A preliminary result of this research was presented in
[8]. In that paper, due to limitations in the implementation,
we were only able to perform the experiments on a single
data center size with low traffic loads. In this paper, with
the new implementation, we perform more extensive exper-
iments and study further the impacts of network load and
data center sizes on the energy saving level of the proposed
Protected Elastic tree topology.
The remainder of this paper is organised as follows. The
Figure 2: An Elastic-tree data center with k = 4 when
3 core switches and some aggregation switches are turned
off.
next section presents briefly the idea of conventional Elas-
tic tree. Section 3 presents the protection model that we
propose to use to reinforce Elastic tree data centers. Section
4 explains how to find the topology of the Protected Elastic
tree. Section 5 shows how much energy can be saved with
Protected Elastic tree and analyse the impact of traffic load
and data center size on energy saving. Finally, Section 6
concludes the paper.
2 Elastic tree
Fat-tree term usually refers to the hierarchical topology
where links closer to the root have greater capacities, al-
though the Fat-tree concept here should be understood dif-
ferently. It is a type of multistage circuit switching net-
work. A k-ary Fat-tree contains k PODs. Each POD con-
tains two layers of k/2 aggregation switches and k/2 edge
switches. Each edge switch links to k/2 servers (or hosts)
and k/2 aggregation switches of the upper layer. Each core
switch links to k PODs by k aggregation switches. All
switches are identical with k ports. Figure 1 shows the
connection path between two servers through 3 layers of
switches in dash line in a 4-ary Fat-tree data center. Fat-
tree has a robust structure but it still consumes a significant
amount of energy.
Elastic tree as proposed in [6] is a system for dynam-
ically adapting the energy consumption of a Fat-tree data
center network. Elastic tree is controlled by three logical
modules Optimiser, Routing, and Power control as shown
in Figure 3. The role of the Optimiser module is to find the
minimum power network subset, which satisfies current (or
statistical) traffic conditions. Its input consists of the topol-
ogy, a traffic matrix, and the power model of each switch.
The Optimiser outputs a set of active components to both
the Power control and Routing modules. The Power control
then turns on or off ports, linecards, or entire switches ac-
cording to this output, while the Routing module chooses
routes for all flows, then pushes routes to the switches in
the network.
Elastic tree refers to the switch power consumption
model proposed in [9]. Power consumption of a switch
consists of a fixed component (chassis power and power
Protected Elastic-tree Topology for. . . Informatica 40 (2016) 197–206 199
Figure 3: Control components of Elastic tree system (figure
from reference [6]).
consumed by each linecard) and a variable component that
depends on the number of active ports and the capacity and
utilisation of each port.
Pswitch = Pchassis + nlinecard × Plinecard (1)
+
∑
i
npri × Pri × Fu
Where Pswitch, Pchassis, Plinecard and Pri are respec-
tively, the power consumed by a switch, a chassis, a
linecard without active ports and an active port running at
rate ri. npri is the number of active ports running at rate
ri. Fu is an utilisation scaling factor for each port. For
simplicity, this factor is considered identical for all switch
ports in Elastic tree.
It is clear that an active switch without traffic processing
still consumes energy. Consequently, a minimum power
network subset is sought for carrying the required network
load for a data center. This network contains only switches
and links that must be active in each layer in order to sat-
isfy the network load. With the help of Software Defined
Network technologies such as OpenFlow [10], unnecessary
switches, line cards or ports are turned off to conserve en-
ergy. The network traffic is then routed so that it flows only
over the active switches. Network traffic may change and
the minimum power subset of active switches must be re-
computed.
Some algorithms have been proposed to identify the
minimum power network subset for a data center given a
required network load. The algorithms include a multi-
commodity flow formulation which is a mixed integer
linear program, a greedy bin-packing algorithm and a
Topology-aware Heuristic (TAH) [6]. Amongst these al-
gorithms, TAH is the fastest with the least computational
effort. TAH does not identify exactly which switches are
a part of the minimum power network subset but instead
identifies the number of active switches in each layer. From
these numbers the total power consumed by the Elastic tree
can be estimated.
The idea of TAH is as follows; based on the statistic of
traffic between layers, TAH identifies the number of links
required between an edge switch and the aggregation layer
to support the up and down traffic from and to that edge
switch. Assuming that the traffic is perfectly divisible, this
number of links is equal to the traffic bandwidth divided by
the link rate. The number of active switches in the aggre-
gation layer is then equal to the number of links required
to support traffic of the most active source from the upper
layer or lower layer. Similar observation holds between the
the aggregation layer and the core. Detailed computation
can be found in [6]. In this paper, TAH is used to estimate
the energy consumption by an Elastic tree.
In [6], the authors have also discussed adding a redun-
dancy level to the Elastic tree by adding k parallel Min-
imum Spanning Trees (MSTs) that overlap at the edge
switches. However, the study does not investigate further
on how many MSTs would be sufficient.
3 Protection model for Elastic
tree Data center
As seen in Figure 2, an Elastic tree is sensitive to failures.
Its connections need to be protected. In this study we focus
solely on single failure scenario. The single failure sce-
nario assumes that there is, at most, one failure in the data
center and this failure is repaired before another one may
occur. The single failure scenario is a typical assumption
in the research and also in practice since the frequency of
failure is low [11] making the possibility of having multiple
failures negligible.
There are several well-known topological protection
schemes: path protection, link protection, and ring protec-
tion [12]. In the path protection model, the connection to
be protected is called working path. Another path, called
backup path, that shares the same end nodes with the work-
ing path is used to replace the working path when the latter
fails due to link or node failures. When no failure is present
the backup path is idle. Since the backup path should not be
affected by any failure on the working path, the two paths
should be disjoint. In link protection, each link in the work-
ing path is protected separately by a backup segment going
from one end of the link to the other end. When a link
fails, only one backup segment is used to replace the link.
Intuitively, the link protection tends to require more backup
resource than path protection since many backup segments
are involved for protecting a single path. Ring protection
is used in networks with ring topology. In that network, a
connection follows a part of the ring and is protected by a
backup connection following the remaining part of the ring
in inverse direction. Due to the tree form topology of Fat-
tree data centers, we focus on the path protection scheme.
A server-to-server connection in data center may travel
through a suite of devices in the following order: source
server - edge switch - aggregation switch - core switch -
aggregation switch - edge switch - destination server. In
Protected Elastic tree, we propose to use the path protection
model to protect the part between the edge switches. The
part between the source server - edge switch and the part
between the edge switch - destination server will be left
unprotected.
200 Informatica 40 (2016) 197–206 D.-L. Truong et al.
According to path protection model, the part between
two edge switches will be protected by a disjoint path be-
tween the same two edge switches. From this principle,
we apply three different protection configurations for Near,
Middle and Far traffic. The notions of Near, Middle and
Far traffic are defined as in [6] and [7].
– Near traffic refers to a flow between source and desti-
nation servers linked to the same edge switch. For this
kind of flow, no protection is offered.
– Middle traffic refers to a flow between source and des-
tination servers linked to different edge switches of
the same POD. This kind of flow passes through an
intermediate aggregate switch then returns to the edge
layer within the POD. The backup flow simply uses
another aggregate switch in order to guarantee that the
working flow is disjointed. See Figure 4 for an exam-
ple.
– Far traffic refers to a flow between source and des-
tination servers linked to different PODs. A flow in
this traffic passes through an aggregation switch of the
source POD, to a core switch, then to an aggregation
switch of the destination POD, and finally to the desti-
nation edge switch. The backup flow needs to use dif-
ferent aggregation and core switches than the working
flow in order to guarantee the disjointedness. Readers
are referred to Figure 5 for an example of a working
flow and its backup flow of the far traffic.
Looking at the topological aspect only (without link capac-
ity consideration), the protection configuration for Middle
traffic is always possible since all aggregation switches and
edge switches of the same POD link to each other. Pro-
tection configuration for Far traffic is also always possible.
Since each core switch links with all PODs then with k ≥ 4
there are always at least four core switches linking to both
the source and destination PODs. Amongst these four core
switches there exists at least two core switches linking to
two different pairs of aggregation switches; one pair in the
source POD, and the other in the destination POD. There-
fore, there always exists two disjoint paths between two
edge switches through these two different pairs of aggre-
gation switches. If one path is serving as the working path
then the other can be utilised as the backup path.
Now consider the bandwidth capacity aspect. In non-
protection mode, a Fat-tree data center with unified link
capacity is fully loaded when all servers saturate their links
to edge layer. In protection mode links between edge, ag-
gregation and core layers have to carry not only the work-
ing flows but also the backup flows, therefore the total
consumed bandwidth is doubled in comparison with non-
protection case. Consequently, under high traffic load, it
is still possible that the data center will not have enough
capacity to allocate backup flows for all working flows.
Figure 4: Example of a backup configuration for Middle
traffic. The working flow is in dash line and its backup
flow is in red line.
Figure 5: Example of a backup configuration for Far traffic.
The working flow is in dash line and its backup flow is in
red line.
4 Minimum active topology with
path protection
Similar to the conventional Elastic tree, the Optimiser in
Protected Elastic tree must find the minimum active topol-
ogy, however in this instance the path protection mech-
anism is integrated. The minimum active topology with
path protection is a minimal network subset that can ac-
commodate not only all working flows for a given traffic
matrix but also a backup flow for each working flow. The
traffic matrix considered here contains the demanded flows
between pairs of edge switches. Since we offer the pro-
tection between edge switches, we are interested in flows
between edge switches only. A flow of traffic between a
pair of edge switches is in fact an aggregation of multiple
flows between servers. Here we assume that a rough traffic
matrix between edge switches is known in advance. This
assumption is quite practical since this rough traffic matrix
can be estimated easily by observing the traffic statistics
over time.
In order to find such a minimal network subset, we will
try to accommodate working and backup paths for all flows
in the traffic matrix. During this accommodation process,
switches and ports are activated gradually. We assume
that all switches and links in the data center are identical.
Switches have k ports and link rate is r. The switches in
data center are linked in k-ary Fat-tree topology. We also
assume that flows are perfectly divisible. Working flows
Protected Elastic-tree Topology for. . . Informatica 40 (2016) 197–206 201
between edge switches are allocated one by one, followed
by the backup flows. The flows are aggregated maximally
to active links, switches and ports until those active devices
are fully used before activating new devices. The following
steps describe how to build the minimum active topology
with path protection.
– We start with an initial active topology as the set of
switches forming a Minimum Spanning Tree (MST)
with a core switch as root and all servers are leaves.
– For allocating the working flow from a source edge
switch es to a destination edge switch ed, we browse
all active paths between es and ed to find the one with
largest residual bandwidth. In the case that all paths
are fully used, we browse from all remaining possi-
ble paths between es and ed to find the one that re-
quires least additional power consumption for activat-
ing. The power consumption of an newly activated
path includes the power for activating new switches,
ports and line cards as estimated by (1).
– Once all working flows are routed their backup paths
will be accommodated. In order to find a backup flow
for a working flow between es and ed, we first exclude
the aggregate switches of its working flow from the
list of usable switches for the backup path. This ex-
clusion ensures that the backup path is disjoint from
the working path. Then, the backup flow is sought in
similar manner as when we allocate a working flow.
If the algorithm fails to find backup paths for some of
the working flows, the Optimiser ends with a false status.
Once the algorithm terminates successfully each connec-
tion request in the traffic matrix will have a backup path
which is disjoint with the working path. Therefore, when a
single failure occurs in the network the working path or the
backup path will not be affected, thus one of them is avail-
able for carrying the traffic. Consequently, the data center
is 100% available to carry the traffic in the traffic matrix
under any single failure.
Although in the above procedure we have to browse all
the possible paths between two edge switches, the special
structure of Fat-tree limits the number of paths available to
browse. From an edge switch there are k/2 choices of ag-
gregation switches to go up. For each aggregation switch,
there are again k/2 choices of core switches. From a core
switch there is only a single choice of route to an edge
switch. Therefore, there are at most k2/4 possible paths
between two edges switches in Far traffic. Similarly, there
are at most k/2 possible paths between two edge switches
in Middle traffic.
Once the Optimiser has calculated the minimum active
topology and the working and backup paths for flows in
the traffic matrix, it gives the Routing module this informa-
tion, e.g., the list of active switches and the list of working
and backup paths for flows between edge switches. Later
on, when the Routing module receives a connection request
between two servers, it can easily identify the two edge
switches associated with the two servers thanks to child-
parent relationship of these devices. The Routing module
will route the connection request over the working path reg-
istered for the flow between the two switches while the cor-
responding backup path is used for protection.
5 Evaluation of energy consumption
of Protected Elastic tree data
centers
The advantage of path protection scheme integrated in Pro-
tected Elastic tree data center is obvious as the data center
is 100% survivable upon any single failure. However, with
the presence of a backup path in parallel with a working
path for each flow between edge switches, the data cen-
ter consumes more energy than the conventional Elastic
tree data center. In order to evaluate how much of energy
the protection scheme imposes into the data center, the en-
ergy consumption of the Protected Elastic tree is compared
against the fully active Fat-tree data center and against the
conventional Elastic tree data center without protection.
The best topology for the Protected Elastic tree is iden-
tified by “Minimal active topology with path protection”
algorithm proposed in Section 4. In the previous work in
[8], the algorithm was implemented in a data center em-
ulation platform: Ecodane [13] which did not allow tests
with a high traffic load due to the large computation effort
that they involve. In this paper, the algorithm has been im-
plemented in Scilab [14], an open source numerical com-
putational package. With the new implementation, we can
perform more extensive experiments. The minimal topol-
ogy for the conventional Elastic tree is calculated using the
TAH algorithm which has also been implemented in Scilab.
Let λij be the total requested bandwidth from server i to
server j. Let #server be the number of servers in the data
center. The total bandwidth capacity provided by links be-
tween servers and edge switches is #servers× r. Since a
flow between server i and server j uses two links between
server layer and edge layer, the maximum total acceptable
load between servers in the data center is 0.5#servers×r.
From this observation, we define the network utilisation in-
dex as the ratio between the total requested bandwidth be-
tween servers in data center and the maximum total accept-
able load of the data center. It is calculated by:
u =
∑
ij λij
0.5#servers× r
× 100% (2)
In case on non-protection, when u = 100% the data cen-
ter is maximally loaded. In case of protection, for each pair
of servers, a flow for working path needs another flow for
backup path. Therefore, the maximum network utilisation
for protected data center is u = 50%.
In all simulations, switches are wired in Fat-tree topol-
ogy. Data center sizes vary with k = 4, 6, 8. All links
202 Informatica 40 (2016) 197–206 D.-L. Truong et al.
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
o
w
e
r 
s
a
v
in
g
 (
%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(a) k=4.
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
o
w
e
r 
s
a
v
in
g
 (
%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(b) k=6.
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(c) k=8.
Figure 6: Power saving level with Middle traffic
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(a) k=4.
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(b) k=6.
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(c) k=8.
Figure 7: Power saving level with Far traffic
are bi-directional with bandwidth capacity of r = 10
Gbps. Traffic requests are generated for 3 traffic models:
Far traffic, Middle traffic and Mixed traffic. Mixed traf-
fic does not include Near traffic since we are not inter-
ested in protecting Near traffic in this research. In order
to reserve enough spare capacity for protection, traffic is
generated with increasing network utilisation until it ap-
proaches u = 50% but without saturating the data center.
Each server-to-server connection requests a bandwidth uni-
formly distributed in range [0-1] Gbps.
The total energy usage by Fat-tree and conventional
Elastic tree are evaluated based to their number of active
devices. According to (1), the power consumption of a
switch consists in a fixed part including power consump-
tion of chassis and line cards, and a dynamic part including
power consumption of ports. Let us denote the fixed part
of the power consumption by P . Let us also consider that
the power consumption of a port is constant regardless of
its working rate and is denoted it by p.
It is easy to prove that a k-ary Fat-tree data center has
a total of 5k2/4 switches, 5k3/4 ports and 3k3/4 links.
Since they are all active, the power consuming by a Fat-
tree, regardless of traffic matrix, is
PFat =
5
4
k2P +
5
4
k3p (3)
In Elastic tree, all edge switches must be active in order
to be ready to receive data from the servers, therefore, the
number of active edge switches is always k2/2. Assume
that a conventional Elastic tree topology, identified by TAH
for a given traffic matrix, uses x aggregation switches (x ≤
k2/2), y core switches (y ≤ k2/4) and has np active ports,
then the total energy consumed by the Elastic tree is
PElastic =
k2
2
P + xP + yP + npp (4)
Protected Elastic-tree Topology for. . . Informatica 40 (2016) 197–206 203
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(a) k=4.
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(b) k=6
 0
 20
 40
 60
 80
 100
 0  10  20  30  40  50
P
ow
er
 s
av
in
g 
(%
)
Network utility (%)
Protected Elastic tree
Elastic tree
(c) k=8
Figure 8: Power saving level with Mixed traffic
Th Energy saving level of the Elastic tree over the Fat-
tree is defined as:(
1− PElastic
PFat
)
× 100% (5)
We denote the energy consumed by Protected Elastic tree
by PProtected.Elastic. The energy consumed by Protected
Elastic tree is also evaluated by using (4) where x, y and
np are the number of active aggregation switches, the num-
ber of active core switches and the number of active ports
in Protected Elastic tree. Those numbers are obtained from
the calculations in the algorithm “Minimal active topology
with path protection”. The energy saving level of the Pro-
tected Elastic tree over Fat-tree is defined as:(
1− PProtected.Elastic
PFat
)
× 100% (6)
In all tests, the power consumed by a switch chassis in-
cluding line cards is set P = 146 watts, the power con-
sumed by a port is set p = 0.9 watts. These numbers have
been chosen after analysing the study of switch power con-
sumption in [9].
Figure 6, 7 and 8 show the power saving level of the Pro-
tected Elastic tree and Elastic tree with k = 4, 6, 8 for Mid-
dle, Far and Mixed traffic. The detailed results are shown
in Table 1, 2 and 3 where the columns show the following
information:
– Network utility (in percentage)
– Power consumed by Fat-tree topology PFat
– Number of active switches, number of active ports,
total power consumption and energy saving level (in
percentage) of the Protected Elastic tree
– Number of active switches, number of active ports,
total power consumption and energy saving level of
the conventional Elastic tree
It is clear that the Protected Elastic tree consumes more
energy than the conventional Elastic tree. The observed
gap between energy saving level of Protected Elastic tree
and Elastic tree varies roughly between 10% and 25%.
In the following sessions the impact of network utility,
traffic model and data center size on the energy saving per-
formance of the proposed Protected Elastic tree topology
are analysed.
5.1 Impact of network utility on energy
saving
We can observe in Figures 6, 7, 8 that, except in Figure 6a,
the power saving levels of both conventional Elastic tree
and Protected Elastic tree decrease when the network util-
ity increases. The obvious reason is that when the network
utility increases, both the conventional Elastic tree and Pro-
tected Elastic tree have to use more switches in order to
accommodate the increase in traffic load.
In the case of the data center with size k = 4 for Middle
traffic in Figure 6a, the power saving levels of both Elas-
tic tree and Protected Elastic tree are constant regardless of
the network utility. This special case can be explained as
follows; with the Middle traffic, all flows (including work-
ing and backup) travel only inside a POD and do not in-
volve core switches. In a non-protected Elastic tree data
center, all edge switches are active in order to be ready to
receive traffic from the servers. At the aggregation layer,
one aggregation switch per POD must be used to carry
all traffic from the POD’s edge switches. With the net-
work utility at 50% or under, one aggregation switch with
two aggregation-edge links is sufficient to carry the traf-
fic from/to the four servers of the POD. Therefore the total
number of active switches is always: 8 (edge) + 4 (aggrega-
tion) =12 switches. Now let us consider the Protected Elas-
tic tree data center, again 8 edge switches must be active.
Both aggregation switches inside a POD must be active in
204 Informatica 40 (2016) 197–206 D.-L. Truong et al.
Protected Elastic tree Elastic tree
Utility PFat Nb. swt. Nb. ports Power Saving Nb. swt. Nb. ports Power Saving
Far traffic
8.16 2992 18 46 2669.4 10.78 13 24 1919.6 35.84
16 2992 18 48 2671.2 10.72 13 24 1919.6 35.84
24.49 2992 19 52 2820.8 5.72 14 28 2069.2 30.84
30.05 2992 19 52 2820.8 5.72 14 28 2069.2 30.84
36.59 2992 19 52 2820.8 5.72 14 28 2069.2 30.84
41.95 2992 19 52 2820.8 5.72 14 28 2069.2 30.84
Middle traffic
6.4 2992 17 40 2518 15.84 12 16 1766.4 40.96
11.98 2992 17 40 2518 15.84 12 16 1766.4 40.96
18.81 2992 17 40 2518 15.84 12 16 1766.4 40.96
23.91 2992 17 40 2518 15.84 12 16 1766.4 40.96
31.16 2992 17 40 2518 15.84 12 16 1766.4 40.96
38.29 2992 17 40 2518 15.84 12 16 1766.4 40.96
Mixed traffic
7.65 2992 18 46 2669.4 10.78 13 24 1919.6 35.84
13.08 2992 18 46 2669.4 10.78 13 24 1919.6 35.84
21.11 2992 18 48 2671.2 10.72 13 24 1919.6 35.84
27.91 2992 19 52 2820.8 5.72 14 28 2069.2 30.84
35.45 2992 19 52 2820.8 5.72 14 28 2069.2 30.84
42.73 2992 19 56 2824.4 5.6 14 28 2069.2 30.84
Table 1: Test results with k=4
Protected Elastic tree Elastic tree
Utility PFat Nb. swt. Nb. ports. Power Saving Nb. swt. Nb. ports Power Saving
Far traffic
9.22 6813 32 96 4758.4 30.16 25 48 3693.2 45.79
13.75 6813 33 100 4908 27.96 26 52 3842.8 43.6
18.21 6813 34 104 5057.6 25.77 27 56 3992.4 41.4
22.13 6813 34 104 5057.6 25.77 27 56 3992.4 41.4
26.04 6813 34 104 5057.6 25.77 27 56 3992.4 41.4
30.6 6813 39 132 5812.8 14.68 32 74 4738.6 30.45
35.47 6813 41 156 6126.4 10.08 33 78 4888.2 28.25
Middle traffic
9.28 6813 31 84 4601.6 32.46 24 36 3536.4 48.09
13.41 6813 33 92 4900.8 28.07 24 36 3536.4 48.09
16.79 6813 36 106 5351.4 21.45 24 36 3536.4 48.09
22.22 6813 36 110 5355 21.4 25 40 3686 45.9
26.99 6813 37 114 5504.6 19.2 26 44 3835.6 43.7
31.23 6813 37 116 5506.4 19.18 29 58 4286.2 37.09
35.92 6813 37 118 5508.2 19.15 29 62 4289.8 37.04
Mixed traffic
9.16 6813 32 96 4758.4 30.16 25 48 3693.2 45.79
13.87 6813 33 100 4908 27.96 26 52 3842.8 43.6
17.7 6813 33 100 4908 27.96 26 52 3842.8 43.6
22.09 6813 34 104 5057.6 25.77 27 56 3992.4 41.4
27.59 6813 35 108 5207.2 23.57 27 56 3992.4 41.4
33.51 6813 40 142 5967.8 12.41 32 76 4740.4 30.42
38.26 6813 42 162 6277.8 7.86 35 86 5187.4 23.86
Table 2: Test results with k=6
Protected Elastic-tree Topology for. . . Informatica 40 (2016) 197–206 205
Protected Elastic tree Elastic tree
Utility PFat Nb. swt. Nb. ports Power Saving Nb. swt. Nb. ports Power Saving
Far traffic
8.7 12256 51 164 7593.6 38.04 42 84 6207.6 49.35
16.7 12256 53 172 7892.8 35.6 44 92 6506.8 46.91
24.12 12256 61 246 9127.4 25.53 52 122 7701.8 37.16
32.64 12256 65 288 9749.2 20.45 56 152 8312.8 32.17
40.04 12256 71 300 10636 13.22 62 176 9210.4 24.85
Middle traffic
7.32 12256 51 154 7584.6 38.12 40 64 5897.6 51.88
14.94 12256 57 202 8503.8 30.62 40 64 5897.6 51.88
22.13 12256 62 232 9260.8 24.44 46 92 6798.8 44.53
29.95 12256 65 266 9729.4 20.62 48 116 7112.4 41.97
37.41 12256 65 272 9734.8 20.57 52 144 7721.6 37
Mixed traffic
7.67 12256 51 164 7593.6 38.04 42 84 6207.6 49.35
15.19 12256 52 168 7743.2 36.82 43 88 6357.2 48.13
22.99 12256 61 238 9120.2 25.59 44 92 6506.8 46.91
30.09 12256 63 268 9439.2 22.98 54 134 8004.6 34.69
38.36 12256 65 282 9743.8 20.5 55 142 8157.8 33.44
Table 3: Test results with k=8
order to provide disjoint working and backup paths for each
connection in the POD. Therefore, 8 aggregation switches
will be involved. Since Protected Elastic tree topology is
built from a MST with a core switch as root then one core
switch will also be involved. Consequently, the total num-
ber of active switches is always 8 (edge) + 8 (aggregation)
+ 1 (core)=17 switches regardless of network load. This
phenomena of constant power saving does not happen with
Middle traffic for bigger data center sizes.
5.2 Impact of traffic model on energy saving
Since Middle traffic does not go through core switches as it
happens in Mixed and Far traffic, those core switches can
be turned off. This characteristic leads to several advan-
tages in Middle traffic:
– The Protected Elastic tree with Middle traffic saves, in
general, slightly more energy than Far and Mixed traf-
fic for the same data center and with the same network
utility (see Tables 1, 2, 3).
– Protected Elastic tree data centers under Middle traf-
fic always save a certain amount of energy regardless
of the network utility. This amount is equivalent to
the energy consumed by core switches. That is the
reason why the energy saving level lines of Middle
traffic become stable as the network utility increases,
meanwhile, those for Far and Mixed traffic continue
to decrease. We can observe that even for networks
with high utility (close to 40%) the saving ratios are
still considerable, i.e., 15.84%, 19.15% and 20.57%
when k = 4, 6 and 8 respectively.
We can also notice that the power saving level lines for
Far and Mixed traffic have similar shapes. This similarity
demonstrates that the Far traffic has a stronger impact on
power saving than Middle traffic. The reason is that Far
traffic involves more links and switches due to their long
connection paths and Middle traffic can profit from those
links and switches for its connections without activating
additional elements.
5.3 Impact of data center size on energy
saving
The performance of Protected Elastic tree over different
data center sizes can also be observed in Figures 6, 7, 8.
We have two remarks:
– Power saving lines of both Elastic and Protected Elas-
tic tree shift up slightly from k = 4 to k = 6, and the
same observation can be made from k = 6 to k = 8,
for all traffic patterns. This means that both the Elastic
tree and the Protected Elastic tree become more en-
ergy efficient as the size of the data center increases.
– The power saving lines of the Protected Elastic tree for
Far and Mixed traffic are closer to those of the Elastic
tree when size of the data center increases. The con-
clusion to be drawn from this is that the extra energy
consumption for protection becomes less important in
large data centers. The main reason of this phenom-
ena is that larger data centers have a higher connectiv-
ity and, therefore, more possibilities to route backup
flows over switches and ports that are already active
to carry working flows.
206 Informatica 40 (2016) 197–206 D.-L. Truong et al.
These two observations show that the Protected Elastic
tree topology is an option worthy of consideration for large
size data centers.
6 Conclusions
The Fat-tree topology has been proposed as a topology for
data centers characterised by a low over-subscription and
a high availability. Elastic tree has been proposed in order
to reduce the energy consumption of Fat-tree data centers
by deactivating unnecessary switches and links. Although
Elastic tree is a highly energy efficient topology, it severely
damages the availability of the Fat-tree since the network
connectivity is reduced remarkably. In this paper, we pro-
posed the possibility of adding protection capabilities to
the conventional Elastic tree by allocating backup paths
for each aggregated flow between edge switches. Many
backup paths can profit from existing active switches and
ports while some others require the activation of additional
elements in the network. This results in a Protected Elas-
tic tree topology where 100% of connections are survivable
for any single failure. The simulation results show that the
path protection scheme generates a decrease in the energy
saving ratio of 10%-25% for the proposed Protected Elastic
tree in comparison with the conventional Elastic tree. How-
ever, the Protected Elastic tree data center saves a notable
amount of energy if it is compared with the Fat-tree, mostly
for the Middle traffic model. Moreover, the proposed solu-
tion becomes more energy efficient as the data center size
increases.
Acknowledgement
The authors would like to thank Assoc. Professor Nguyen
Huu Thanh for his encouragement to perform this research.
References
[1] FEPM, “FEMP Data Center Program Overview,”
Tech. Rep. DOE/EE-1191, Federal Energy Manage-
ment Program, March 2015.
[2] Anthesis, “Data Center Efficiency Assessment,”
Tech. Rep. IP:14-08-A, US Natural Resources De-
fense Council, August 2014.
[3] K. K. Nguyen, M. Cheriet, M. Lemay, V. Reijs,
A. Mackarel, and A. Pastrama, “Environmental-
aware virtual data center network,” Computer Net-
works, vol. 56, no. 10, pp. 2538 – 2550, 2012. Green
communication networks.
[4] C. Leiserson, “Fat-trees: Universal networks for
hardware-efficient supercomputing,” Computers,
IEEE Transactions on, vol. C-34, pp. 892–901, Oct
1985.
[5] M. Al-Fares, A. Loukissas, and A. Vahdat, “A scal-
able, commodity data center network architecture,”
in Proceedings of the ACM SIGCOMM 2008 Confer-
ence on Data Communication, SIGCOMM ’08, (New
York, NY, USA), pp. 63–74, ACM, 2008.
[6] B. Heller, S. Seetharaman, P. Mahadevan, Y. Yi-
akoumis, P. Sharma, S. Banerjee, and N. McKe-
own, “Elastictree: Saving energy in data center net-
works,” in Proceedings of the 7th USENIX Confer-
ence on Networked Systems Design and Implemen-
tation, NSDI’10, (Berkeley, CA, USA), pp. 17–17,
USENIX Association, 2010.
[7] T. N. Huu, N. P. Ngoc, H. T. Thu, T. T. Ngoc,
D. N. Minh, V. G. Nguyen, H. N. Tai, T. N. Quynh,
D. Hock, and C. Schwartz, “Modeling and exper-
imenting combined smart sleep and power scaling
algorithms in energy-aware data center networks,”
Simulation Modelling Practice and Theory, vol. 39,
pp. 20 – 40, 2013. S.I.Energy efficiency in grids and
clouds.
[8] D. L. Truong and T. C. Nguyen, “Protected elastic-
tree topology for data center,” in Proceedings of the
Sixth International Symposium on Information and
Communication Technology, Hue City, Vietnam, De-
cember 3-4, 2015, pp. 129–134, 2015.
[9] P. Mahadevan, P. Sharma, S. Banerjee, and P. Ran-
ganathan, “A power benchmarking framework for
network devices,” in Proceedings of the 8th Inter-
national IFIP-TC 6 Networking Conference, NET-
WORKING ’09, (Berlin, Heidelberg), pp. 795–808,
Springer-Verlag, 2009.
[10] “Openflow.” https://www.
opennetworking.org/sdn-resources/
openflow/.
[11] M. To and P. Neusy, “Unavailability analysis of long-
haul networks,” IEEE Journal on Selected Areas in
Communications, vol. 12, pp. 100–109, Jan. 1994.
[12] J. Zhang and B. Mukheriee, “A review of fault man-
agement in WDM mesh networks: basic concepts
and research challenges,” IEEE Network: Magazine
of Global Internetworking, vol. 18, pp. 41–48, Mar.
2004.
[13] H. T. Nguyen, D. C. Bui, D. T. To, N. N. Pham,
Q. T. Ngo, T. H. Truong, and M. N. Tran, “Ecodane:
A customizable hybrid testbed for green data cen-
ter networks,” in Advanced Technologies for Commu-
nications (ATC), 2013 International Conference on,
pp. 312–317, Oct 2013.
[14] “Scilab.” http://www.scilab.org.