QAMO: QoS Aware Multipath-TCP Over Optical Burst Switching in
Data Centers
Sana Tariq and Mostafa Bassiouni
Department of Elec. Eng. & Computer Science, University of Central Florida, Orlando, U.S.A.
Keywords:
MPTCP, All-optical Network, Data Center, TCP, OBS, QoS.
Abstract:
The rapid advancement in cloud computing is leading to a promising future for shared data centers hosting
diverse applications. These applications constitute a complex mix of workloads from multiple organizations.
Some workloads require small predictable latency while others require large sustained throughput. Such
shared data-centers are expected to provide potential service differentiation to client’s individual flows. This
paper addresses two important issues in shared data centers: bandwidth efficiency and service differentiation
based on QoS (Quality of Service). We first evaluate the multipath-TCP (MPTCP) protocol over an
OBS (Optical burst switching) network for improved bandwidth utilization in dense interconnect datacenter
networks. We next present a simple and efficient ‘QoS aware MPTCP over OBS’ (QAMO) algorithm in
datacenters. Our experimental results show that Multipath-TCP improves throughput over an OBS network
while the QAMO algorithm achieves tangible service differentiation without impacting the throughput of the
system.
1 INTRODUCTION
Many internet applications today are powered by
data centers equipped with hundreds of thousands
of servers. The concept of shared datacenters also
became popular with the widespread adaptation of
cloud. There is a growing interest in introducing QoS
(Quality-of-Service) differentiation in datacenters,
motivated by the need to improve the quality of
service for time sensitive datacenter applications and
to provide clients with a range of service-quality
levels at different prices. Over the past decade,
considerable attention has been given to different
areas of cloud computing e.g., efficient sharing of
computational resources, virtualization, scalability
and security. However, less attention has been
paid to network management and QoS (Quality-of-
Service) provisioning in datacenters. The inability
of today’s cloud technologies to provide dependable
and predictable services is a major showstopper
for the widespread adoption of the cloud paradigm
(Rygielski and Kounev, 2013).
The type of applications hosted by datacenters
are diverse in nature ranging from back-end services
such as search indexing, data replication, MapReduce
jobs to front end services triggered by clients such as
web search, online gaming and live video streaming
(Chen et al., 2011). The background traffic contains
longer flows and is throughput sensitive while the
interactive front end traffic is composed of shorter
messages and is delay sensitive. The traffic belonging
to the same class can also have differences in relative
priority levels and performance objectives (Ghosh
et al., 2013).
In this paper, we first evaluate the performance
of MPTCP over OBS for datacenter networks and
compare the performance of TCP with MPTCP
under different network loads and topologies. We
next present and evaluate a QoS provisioning
algorithm called QAMO, ‘QoS aware MPTCP over
OBS’. To our knowledge, this is the first research
report that provides QoS provisioning algorithm for
service differentiation using MPTCP over OBS in
datacenters.
The rest of the paper is organized as follows. In
section 2, we review previous work. In section 3,
we describe our networking model that uses MPTCP
protocol over an optical burst switching network for
data centers. In section 4, we present ‘QoS-aware
MPTCP over OBS’ (QAMO) scheme. Simulation
details are discussed in section 5 and the performance
analysis and simulation results are given in Section 6.
We conclude the paper in Section 7.
56
Tariq S. and Bassiouni M..
QAMO: QoS Aware Multipath-TCP Over Optical Burst Switching in Data Centers.
DOI: 10.5220/0005047800560063
In Proceedings of the 5th International Conference on Optical Communication Systems (OPTICS-2014), pages 56-63
ISBN: 978-989-758-044-4
Copyright
c
 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2 PREVIOUS WORK
Future data center consumers will require quality
of service QoS as a fundamental feature. There
have been some research studies on traffic modeling,
network resource management and QoS provisioning
in data centers (Chen et al., 2011), (Benson et al.,
2010a) and (Ranjan et al., 2002). Ranjan, et.
al., studied the problem of QoS guarantees in
data-center environments in (Ranjan et al., 2002).
However, this work is not suitable for highly
loaded shared data-centers with computationally
intensive applications due to the two sided nature
of communication. The work in (Song et al.,
2009) proposed a resource scheduling scheme which
automatically provides on-demand capacities to the
hosted services, preferentially ensuring performance
of some critical services while degrading others when
resource competition arises. A flow scheduling
protocol called Preemptive Distributed Quick (PDQ)
in (Hong et al., 2012) was, designed to complete flows
quickly by emulating a shortest job first algorithm and
giving priority to the short flows. Similarly research
in (Liu et al., 2013) proposed taxonomy to categorize
existing works based on three main techniques:
reducing queue length, prioritizing mice flows, and
exploiting multi-path. DeTail (Zats et al., 2012) is an
idea of cross-layer network stack aiming to improve
the tail of completion time for delay-sensitive flows.
A deadline-aware control protocol was presented in
(Wilson et al., 2011), named D3, which controlled
the transmission rate of network flows according to
their deadline requirements. D3 gave priority to
mice flows and improved the transmission capacity
of datacenter networks. The techniques of giving
priority to mice flows and exploiting multiple of
newer datacenter topologies have proved to be
effective and efficient means of achieving service
differentiation in datacenters.
However none of the presented QoS provisioning
schemes explored the combined approach of using
multiple paths of newly emerging transport protocols
such as Multi-path TCP (MPTCP) with wavelength
reservation at network layer in optical domain to
achieve service differentiation. Our proposed scheme
combines the flexibility of bandwidth reservation at
two levels to achieve QoS in datacenters that marks
its novelty in approach. There is rich research
on QoS schemes in optical burst switching for wide
area networks (Chen et al., 2001), (Zhang et al.,
2003), (Yoo et al., 2000) and (Akar et al., 2010).
OBS has been considered as the best compromise
between optical circuit switching (OCS) and optical
packet switching (OPS) due to its granularity and
bandwidth flexibility, and would be suitable for data
centers eventually as optical switching technology
gets mature (Peng et al., 2012). TCP is the most
dominant transport layer protocol in internet and TCP
over OBS has been extensively studied (Lazzez et al.,
2008), (Shihada et al., 2009) and (Zhang et al.,
2005). Multipath-TCP (MPTCP) has been shown to
provide significant improvement in throughput and
reliability in electronic packet switched networks in
data centers (Raiciu et al., 2011) and (Raiciu et al.,
2010). However, MPTCP has not been studied in the
context of OBS networks before. In this paper we will
develop a QoS provisioning scheme for data center
networks using MPTCP over OBS and evaluate its
performance.
3 NETWORK MODEL
With the popularity of new data center topologies
such as Fat Tree and VL2 and the multitude of
available network paths, it becomes natural to switch
to multi path transport protocol such as MPTCP to
seek performance gains. MPTCP provides significant
improvement in bandwidth, throughput and fairness.
We have used MPTCP over OBS in our proposed
network architecture. In an OBS network, the control
information is sent over a reserved optical channel,
called the control channel, ahead of the data burst
in order to reserve the wavelengths across all OXCs
(Optical cross connects). The control information
is electronically processed at each optical router
while the payload is transmitted all-optically with full
transparency through the lightpath. The wavelength
reservation protocol plays a crucial role in the burst
transmission and we have used just-in-time (JIT)
(Wei and McFarland, 2000) for its simplicity. The
necessary hardware level modifications of optical
switches for supporting OBS in data centers have
been discussed in (Sowailem et al., 2011), and will
not be repeated in this paper.
4 QoS AWARE MPTCP OVER OBS
ALGORITHM
Our proposed algorithm QoS aware MPTCP over
OBS called QAMO combines the multiple paths of
MPTCP and resource reservation in OBS to develop
an adaptive and efficient QoS-aware mechanism.
Data centers handle a diverse range of traffic
generated from different applications. The traffic
generated from real time applications e.g., web
search, retail advertising, and recommendation
QAMO:QoSAwareMultipath-TCPOverOpticalBurstSwitchinginDataCenters
57

systems consists of shorter flows and requires faster
response. These shorter flows (foreground traffic)
are coupled with bandwidth intensive longer flow
(background traffic) carrying out bulk transfers.
The bottleneck created by heavy background
traffic impacts the performance of latency sensitive
foreground traffic. It is extremely important to
provide a preferential treatment to time sensitive
shorter flows to achieve an expected performance for
data center applications. QoS technologies should
be able to prioritize traffic belonging to more critical
applications. Our proposed algorithm provides
priority to latency-sensitive flows at two levels, i)
MPTCP path selection stage and ii) OBS wavelength
reservation stage. We propose that larger bandwidth
be dynamically allocated to high priority flows, in
order to minimize latency and reduce their drop
probability. QAMO algorithm does that as follows:
Let W be the maximum number of wavelengths
per fiber, and K be the number of paths that exist
between a given souce-destination pair. We will
introduce a new term, the priority factor P for a burst
priority defined as the ratio of P
curr
(priority level of
the current burst) to P
max
(maximum priority levels)
i.e., P = P
curr
/P
max
. Priorities of individual bursts are
represented in ascending order as P
1
,P
2
,P
3
,...,P
max
while P
max
is the highest priority level in the bursts.
We next define the number of allocated paths k
curr
for
the burst of a particular priority level as follows in
equation 1.
k
curr
= ⌈K · P⌉ (1)
At path allocation stage a larger number of paths
is allocated for a high priority burst thus reducing its
latency. For example, if P
curr
= P
max
, then P = 1. This
will result in k
curr
= K paths whereas if P
curr
= 0.5·
P
max
, then P = 0.5 and the number of allocated paths
is reduced to half the set of K paths. This will give
the low priority burst, half the number of paths. We
now define the size of the wavelength search space
controlled by equation 2
Wavelengthsearchsize= ⌈W · P⌉ (2)
At wavelength reservation stage in OBS,
equation 2 allocates a larger subset of wavelength
search space for a burst with higher priority level
thereby allowing it a greater chance to get through
and reduce its blocking probability.
In algorithm 1, the priority factor P is used to
adjust the number of allocated paths for concurrent
transmission and the size of the wavelength search
space based on the priority level of the burst.
For high priority bursts, more concurrent MPTCP
paths result in larger bandwidth, and more OBS
Algorithm 1 : QAMO (QoS Aware MPTCP over
OBS) Algorithm.
1: Input:
2: P = P
curr
/P
max
3: K = maximum number of paths
4: W = maximum number of wavelengths
5: w
curr
= current wavelength reserved for current
burst
6: N
k
= vector of all nodes on path k
7: k
curr
= paths allocated to the current burst
8: burst
curr
= current burst
9: Algorithm:
10: for each k in ⌈K · P⌉ do
11: generate k lightpaths by making k concurrent
function calls to generateLightPath()
12: end for
13: generateLightpath(path k)
14: Initialize w
curr
15: for each n in N
k
do
16: if n is the destination node then
break
17: end if
18: if n is the source node then
19: for each w in ⌈W · P⌉ do
20: if w is free then
21: reserve w for burst
curr
at n
22: w
curr
= w break
23: end if
24: end for
25: else
26: if w
curr
is free at n then
27: reserve w
curr
for burst
curr
at n continue
28: end if
29: for each w in ⌈W · P⌉ do
30: if w is free then
31: reserve w for burst
curr
at n
32: w
curr
= w break
33: end if
34: end for
35: end if
36: if no free wavelength at n then
37: return (error) Search failed at node n
38: end if
39: end for
40: return (success)
network wavelengths reduce dropping probability.
The parameter P
max
can be flexible to accommodate
changes in network statistics over time as bursts of
different priority levels are encountered. QAMO
algorithm should only be active when the network is
congested and the received traffic exhibits difference
OPTICS2014-InternationalConferenceonOpticalCommunicationSystems
58

between relative priority levels.
P2
Application
Layer
P1
P3
Applications generating bursts of
different priority levels
Transport Layer
(Multi-
path TCP)
QAMO: K Paths assignment (MPTCP)
Priority
level
information
Priority level information
Link/physical
Layer (Optical
cross connects)
Network Layer
(Optical Burst
Switching)
K=1
K=2
K=N
QAMO: OBS wavelength
reservation for K paths
Priority
level
information
Control packet carries priority information
from OXC to
OXC required for wavelength
reservation
Figure 1: QAMO’s cross-layer design: Changes to the
protocol stack and the burst priority level information flow.
As shown in figure 1, we assume that QAMO
algorithm has access to available information about
QoS requirements of different bursts to process them
correctly. AT MPTCP layer this capability may be
implemented using a specific interface such as the
Implicit Packet Meta Header (IPMH) promoted in
(Exposito et al., 2009). Because of MPTCP IPMH
interface (Diop et al., 2012) and (Diop et al., 2011),
it is possible to assign priority levels for different
flows and gather priority information for each type
of flow at a particular end host. Under QAMO
scheme, when the burst priority level information is
received at MPTCP layer, QAMO initiates k lightpath
requests to the OBS network layer. This information
can be passed on to the OBS network during burst
segmentation process from MPTCP layer. At OBS
network, the current burst priority P
curr
, or the ratio
P = P
curr
/P
max
, can easily be passed on to lower
layer and then from one one OXC to the next via the
control packet and does not demand any significant
resources in the OXC’s. Implementing the reduced
(adjustable) search as in the case of QAMO, to find
a free wavelength requires minor modification to
the standard JIT channel allocation scheme. The
adjustable search in a smaller space of ⌈W · P⌉ for
wavelengths actually leads to a smaller average search
time.
The QAMO scheme has been extensively tested
on the simulation testbed using data center network
topologies FatTree and BCube and is shown
to provide tangible QoS differentiation without
negatively impacting the overall throughput of the
system.
5 SIMULATION DETAILS
The simulation testbed has been developed using
C++. A source-destination pair amongst host nodes is
randomly chosen for each originated burst. For TCP,
to establish the static lightpath, simulation calculates
the shortest path between these nodes using Dijkstra’s
algorithm. In case of MPTCP, it uses K shortest paths
algorithm (derived from Dijkstra’s algorithm) to find
K paths between the source-destination pair. The
wavelength assignment heuristic is first-fit as done in
(Tariq and Bassiouni, 2013) and (Tariq et al., 2013).
Recent research studies on traffic characteristics of
data centers have shown that the traffic in data
centers follows the lognormal distribution with ON-
OFF pattern (Benson et al., 2010a) and (Benson et al.,
2010b). The lognormal distribution is also considered
to be the most fitted distribution for modeling various
categories of internet traffic including TCP (Pustisek
et al., 2008). We have used lognormal arrival with
an ON-OFF behavior in our simulation. The network
nodes are assumed to be equipped with wavelength
converters. We assume that MPTCP is running at end
hosts. Based on the priority of the burst, K control
packets originate from the source node to establish
K lightpaths. Each control packet acquires an initial
free wavelength at the source node, then travels
to the destination node and reserves wavelengths
following QAMO algorithm. If at any node, the same
wavelength as the one reserved on the previous node
is not available then it tries wavelength conversion.
The process continues until the control packet either
reaches the destination node or gets blocked due to
the unavailability of free wavelength at any hop along
the path. Thus, number of lightpaths established = K
number of control packets blocked. The source node
waits for a predetermined time depending on the hop
distance to the destination called offset time before
transmitting the optical burst message. The traffic
used in our simulation is uniformly distributed, i.e.,
any host node can be a source or a destination (Tariq
and Bassiouni, 2013) and (Gao and Bassiouni, 2009).
The simulation clock is divided into time units,
where each simulation time unit corresponds to 1
microsecond (µs). Each node has a control packet
processing time of 20 microseconds and a cut through
time of 1 microsecond as proposed for OBS networks
in data centers (Saha et al., 2012). Each node
QAMO:QoSAwareMultipath-TCPOverOpticalBurstSwitchinginDataCenters
59

can have a certain maximum number W of allowed
wavelengths. Arrival rate/µs denotes the average
arrival rate of the lognormal ON-OFF traffic.
In data center environment a complex mix of
short and long flows is generated. The shorter
flows are usually latency-critical and represent the
largest proportion of flows in data centers (Benson
et al., 2010a). The medium sized and longer flows
constitute background traffic and may belong to
different priority levels (Alizadeh et al., 2011). To
represent these scenarios of data center mixed traffic,
we have used variable burst sizes in different ranges
with uniform distribution within each range (Alizadeh
et al., 2011).
Short burst sizes: S
min
= 5 Kbits to S
max
= 20 Kbits
Medium burst sizes: S
min
= 200 Kbits to S
max
= 1 Mbits
Long burst sizes: S
min
= 20 Mbits to S
max
= 100 Mbits
Our traffic model is based on the findings on data
center traffic characteristics in (Chen et al., 2011),
(Benson et al., 2010a), (Benson et al., 2010b) and
(Alizadeh et al., 2011). To model our traffic we
assume dynamically changing traffic with an average
of 70 - 80% of bursts generated in short burst range
belonging to latency sensitive applications, 10 - 15%
in medium burst sizes while 5 - 10% of bursts belongs
to large burst size range. In order to assign the
priorities we use dynamically changing priority levels
and relative percentages of various priority classes
with an average of 95% short burst messages having
the randomly assigned priorities from the highest
priority range [P5 - P6]; the remaining 5% can have
any priority level. Similarly, 95% of medium and
large burst sizes are randomly assigned priorities from
sets [P3 P4] and [P1 P2] respectively. The remaining
5% from these ranges are assigned random priorities
from set [P1 P6].
6 RESULTS AND DISCUSSION
The topologies used in our simulation tests are
FatTree with 36 nodes and BCube with 24 nodes as
shown in Figure 2 and Figure 3. In FatTree topology
the root level nodes are called high level aggregators
(HLAs), the next layer of nodes are medium level
aggregators (MLAs). The longest lightpath in the
36-node FatTree network has the diameter of 6 hops.
There are 16 hosts as shown in Figure 2 in the bottom
layer.
The BCube network shown in Figure 3 has 16
relaying hosts in the middle layer. The network
diameter in the 24-node BCube network is 4 hops.
Figure 2: FatTree topology used for simulation.
All the figures in this section are tested following
lognormal distribution. Because of the ON-OFF
pattern of traffic the average arrival rate is smaller
than the arrival rate of a continuouslognormal process
having the same mean and standard deviation. The
tests are conducted over burst distribution of our
proposed traffic model discussed in section 5.
Figure 3: BCube topology used for simulation.
Figure 4 motivates the use of MPTCP in data
center networks for improving throughput. Figure 4
is tested using the lognormal distribution with mean
µ = 1.8 and standard deviation σ = 1, corresponding
to an arrival rate of 7.12/tu in BCube topology.
Figure 4 shows the throughput comparison between
TCP (K = 1) and MPTCP (K = 2,3,4), where K is
the number of paths (i.e., number of subflows) used
by each MPTCP connection. It can be observed that,
MPTCP gives much higher throughput as compared
to single path TCP. It can also be observed that
MPTCP performs better with increasing number of
paths. Similar results were achieved for FatTree
topology.
Figure 5 shows the ability of QAMO algorithm to
achieve QoS differentiation when tested for bursts of
various sizes and priority levels as proposed in our
traffic model. The dropping probability comparison
for six priority levels is shown with increasing load in
a FatTree topology. For lognormal traffic, the mean
values used in this test are from µ = 1 to µ = 3 and
standard deviation σ = 1. It can be observed that the
algorithm achieves substantial QoS differentiation for
all priority levels. For example, P6 being the highest
OPTICS2014-InternationalConferenceonOpticalCommunicationSystems
60

0
5
10
15
20
25
30
35
TCP(K=1) MPTCP(K=2) MPTCP(K=3) MPTCP(K=4)
Throughput (TBits/sec)
Throughput comparison - BCube Topology
Figure 4: Arrival Rate /µs = 7.12, W = 64.
priority level, experiences the least dropping at all
values of input load. Similar results were achieved
for BCube topology.
0
0.02
0.04
0.06
0.08
0.1
0.12
3.19 4.77 7.14 10.58 15.79 23.61
Dropping Probability
Arrival rate /µs
Dropping Probability - FatTree Topology
P6
P5
P4
P3
P2
P1
Figure 5: Variable arrival rate, W = 64.
Figure 6 shows the average throughput
comparison of TCP, MPTCP (K = 4) and QAMO.
The lognormal mean values used in this test are
from µ = 0.5 to µ = 1.75 and standard deviation
σ = 1. It can be observed that QAMO and MPTCP
(K = 4) both performs much better than standard
TCP. The throughput of QAMO is slightly less than
MPTCP (K = 4) at small values of input load while
the difference in throughput becomes less at higher
loads. The reason for QAMO’s degraded throughput
is its preferential treatment for higher priority bursts,
which are mostly very small in size.
Figure 7 provides deeper analysis of throughput
breakdown in terms of burst priorities at one of the
loads from Figure 6, specifically at arrival rate =
2.49 bursts/tu. The lognormal mean in Figure 7
is µ = 0.75 and standard deviation σ = 1. It can
be observed that in TCP and MPTCP the greatest
share of throughput is achieved by low priority
background traffic, giving less importance to the
time sensitive foreground flows in the absence of
QoS provisioning. The throughput of QAMO is
well distributed between high priority (foreground)
and low priority (background) traffic. Hence, the
slight degradation of QAMO throughput compared
0
5
10
15
20
25
30
35
1.94
2.49
3.18
4.09
5.26
6.74
Throughput( TBits/sec)
Arrival rate /µs
Throughput Comparison
-
FatTree Topology
TCP
MPTCP
QAMO
Figure 6: Variable arrival rate, W = 64.
to MPTCP is acceptable for achieving better share
of network resources for more critical traffic in data
centers.
0
2
4
6
8
10
12
14
TCP
MPTCP
QAMO
P6
P5
P4
P3
P2
P1
Throughput (TBits/sec)
Throughput distribution per priority level - FatTree
Topology
Figure 7: Arrival Rate /µs = 2.49, W = 64.
7 CONCLUSIONS
In this paper we have shown the benefits of the
newly emerging transport protocol MPTCP over OBS
networks in data centers. We have seen that MPTCP
improves the throughput and reliability in data
center networks by parallel transmission on multiple
paths. We have presented and evaluated QoS-
aware MPTCP over OBS (QAMO) scheme to provide
service differentiation in data center traffic. QAMO
algorithm provides tangible QoS differentiation to
bursts of various classes without impacting the
throughput of the system. For future work, we plan to
improve basic QAMO scheme to an adaptive and self
configurable algorithm that can change its dynamics
based on current network feedback. These extensions
will make it applicable in software defined networks
(SDN) for future datacenters.
QAMO:QoSAwareMultipath-TCPOverOpticalBurstSwitchinginDataCenters
61

REFERENCES
Akar, N., Karasan, E., Vlachos, K. G., Varvarigos, E. A.,
Careglio, D., Klinkowski, M., and Sol´e-Pareta, J.
(2010). A survey of quality of service differentiation
mechanisms for optical burst switching networks.
Optical Switching and Networking, 7(1):1–11.
Alizadeh, M., Greenberg, A., Maltz, D. A., Padhye, J.,
Patel, P., Prabhakar, B., Sengupta, S., and Sridharan,
M. (2011). Data center tcp (dctcp). ACM SIGCOMM
computer communication review, 41(4):63–74.
Benson, T., Akella, A., and Maltz, D. A. (2010a). Network
traffic characteristics of data centers in the wild. In
Proceedings of the 10th ACM SIGCOMM conference
on Internet measurement, pages 267–280. ACM.
Benson, T., Anand, A., Akella, A., and Zhang, M. (2010b).
Understanding data center traffic characteristics.
ACM SIGCOMM Computer Communication Review,
40(1):92–99.
Chen, Y., Hamdi, M., and Tsang, D. H. (2001). Proportional
qos over obs networks. In Global Telecommunications
Conference, 2001. GLOBECOM’01. IEEE, volume 3,
pages 1510–1514. IEEE.
Chen, Y., Jain, S., Adhikari, V. K., Zhang, Z.-L., and Xu,
K. (2011). A first look at inter-data center traffic
characteristics via yahoo! datasets. In INFOCOM,
2011 Proceedings IEEE, pages 1620–1628. IEEE.
Diop, C., Dugu´e, G., Chassot, C., and Exposito, E. (2011).
Qos-aware multipath-tcp extensions for mobile and
multimedia applications. In Proceedings of the
9th International Conference on Advances in Mobile
Computing and Multimedia, pages 139–146. ACM.
Diop, C., Dugu´e, G., Chassot, C., Exposito, E., and
Gomez, J. (2012). Qos-aware and autonomic-oriented
multi-path tcp extensions for mobile and multimedia
applications. International Journal of Pervasive
Computing and Communications, 8(4):306–328.
Exposito, E., Gineste, M., Dairaine, L., and Chassot,
C. (2009). Building self-optimized communication
systems based on applicative cross-layer information.
Computer Standards & Interfaces, 31(2):354–361.
Gao, X. and Bassiouni, M. A. (2009). Improving
fairness with novel adaptive routing in optical burst-
switched networks. Journal of Lightwave Technology,
27(20):4480–4492.
Ghosh, A., Ha, S., Crabbe, E., and Rexford, J.
(2013). Scalable multi-class traffic management in
data center backbone networks. Selected Areas in
Communications, IEEE Journal on, 31(12):2673–
2684.
Hong, C.-Y., Caesar, M., and Godfrey, P. (2012).
Finishing flows quickly with preemptive scheduling.
ACM SIGCOMM Computer Communication Review,
42(4):127–138.
Lazzez, A., Boudriga, N., and Obaidat, M. S. (2008).
Improving tcp qos over obs networks: a scheme based
on optical segment retransmission. In Performance
Evaluation of Computer and Telecommunication
Systems, 2008. SPECTS 2008. International
Symposium on, pages 233–240. IEEE.
Liu, S., Xu, H., and Cai, Z. (2013). Low latency
datacenter networking: A short survey. arXiv preprint
arXiv:1312.3455.
Peng, L., Youn, C.-H., Tang, W., and Qiao, C.
(2012). A novel approach to optical switching for
intradatacenter networking. Lightwave Technology,
Journal of, 30(2):252–266.
Pustisek, M., Humar, I., and Bester, J. (2008). Empirical
analysis and modeling of peer-to-peer traffic flows. In
Electrotechnical Conference, 2008. MELECON 2008.
The 14th IEEE Mediterranean, pages 169–175. IEEE.
Raiciu, C., Barre, S., Pluntke, C., Greenhalgh, A., Wischik,
D., and Handley, M. (2011). Improving datacenter
performance and robustness with multipath tcp. In
ACM SIGCOMM Computer Communication Review,
volume 41, pages 266–277. ACM.
Raiciu, C., Pluntke, C., Barre, S., Greenhalgh, A., Wischik,
D., and Handley, M. (2010). Data center networking
with multipath tcp. In Proceedings of the 9th ACM
SIGCOMM Workshop on Hot Topics in Networks,
page 10. ACM.
Ranjan, S., Rolia, J., Fu, H., and Knightly, E. (2002). Qos-
driven server migration for internet data centers. In
Quality of Service, 2002. Tenth IEEE International
Workshop on, pages 3–12. IEEE.
Rygielski, P. and Kounev, S. (2013). Network
virtualization for qos-aware resource management
in cloud data centers: A survey. Praxis
der Informationsverarbeitung und Kommunikation,
36(1):55–64.
Saha, S., Deogun, J. S., and Xu, L. (2012). Hyscaleii: A
high performance hybrid optical network architecture
for data centers. In Sarnoff Symposium (SARNOFF),
2012 35th IEEE, pages 1–5. IEEE.
Shihada, B., Ho, P.-H., and Zhang, Q. (2009). A novel
congestion detection scheme in tcp over obs networks.
Journal of Lightwave Technology, 27(4):386–395.
Song, Y., Wang, H., Li, Y., Feng, B., and Sun, Y.
(2009). Multi-tiered on-demand resource scheduling
for vm-based data center. In Proceedings of the
2009 9th IEEE/ACM International Symposium on
Cluster Computing and the Grid, pages 148–155.
IEEE Computer Society.
Sowailem, M. Y., Plant, D. V., and Liboiron-Ladouceur, O.
(2011). Implementation of optical burst switching in
data centers. IEEE PHO, WG4.
Tariq, S., Bassiouni, M., and Li, G. (2013). Improving
fairness of obs routing protocols in multimode
fiber networks. In Proceedings of the 2013
International Conference on Computing, Networking
and Communications (ICNC), pages 1146–1150.
IEEE Computer Society.
Tariq, S. and Bassiouni, M. A. (2013). Hop-count fairness-
aware protocols for improved bandwidth utilization
in wdm burst-switched networks. Photonic Network
Communications, 25(1):35–46.
Wei, J. Y. and McFarland, R. I. (2000). Just-in-time
signaling for wdm optical burst switching networks.
Journal of lightwave technology, 18(12):2019.
OPTICS2014-InternationalConferenceonOpticalCommunicationSystems
62

Wilson, C., Ballani, H., Karagiannis, T., and Rowtron, A.
(2011). Better never than late: Meeting deadlines in
datacenter networks. In ACM SIGCOMM Computer
Communication Review, volume 41, pages 50–61.
ACM.
Yoo, M., Qiao, C., and Dixit, S. (2000). Qos performance
of optical burst switching in ip-over-wdm networks.
Selected Areas in Communications, IEEE Journal on,
18(10):2062–2071.
Zats, D., Das, T., Mohan, P., Borthakur, D., and Katz, R.
(2012). Detail: reducing the flow completion time tail
in datacenter networks. ACM SIGCOMM Computer
Communication Review, 42(4):139–150.
Zhang, Q., Vokkarane, V. M., Chen, B., and Jue, J. P.
(2003). Early drop scheme for providing absolute
qos differentiation in optical burst-switched networks.
In High Performance Switching and Routing, 2003,
HPSR. Workshop on, pages 153–157. IEEE.
Zhang, Q., Vokkarane, V. M., Wang, Y., and Jue,
J. P. (2005). Analysis of tcp over optical
burst-switched networks with burst retransmission.
In Global Telecommunications Conference, 2005.
GLOBECOM’05. IEEE, volume 4, pages 6–pp. IEEE.
QAMO:QoSAwareMultipath-TCPOverOpticalBurstSwitchinginDataCenters
63