Professional Documents
Culture Documents
275
ONDM 2013 Brest, France
276
ONDM 2013 Brest, France
prior to an interval of an offset time. This control packet is at edge nodes those do not have the global knowledge of
processed at each intermediate node to reserve wavelength network states.
resources and configure optical cross-connects (OXCs) for In the proposed adaptive-burst assembling procedure, bursts
the interval of the burst arrival such that the burst can be are assembled at edge nodes while the burst size is estimated
transmitted all-optically, and thus, the bottlenecks of electrical at the controller. The controller estimates the burst size based
processing are eliminated. The performance of OBS is highly on the available network resources and the latency requirement
influenced by a burst assembling procedure, a burst routing Le of a traffic class C. Let Z denotes the burst configuration
and scheduling procedure, and a control packet signaling time that is defined as Z = (2 x T� x D�) + ToDe +
protocol. In the prior research, the proportional-QoS based Toxe, where T� denotes OpenFlow control message prop
burst assembling algorithm [6], the source-based routing and agation latency per unit distance, D� denotes the distance
burst scheduling algorithm [7], and the JET-based signaling between a controller and an edge node, ToDC is control
protocol [8] have gained a lot of attentions. However, these packet processing time at ODC, and Toxe is an upper bound
algorithms and protocol may not directly be applicable to the of optical cross-connect configuration time. Thus, the burst
centralized software-defined optical network architecture. configuration time is independent of the route length. The
In SDON, a centralized controller is responsible for making procedure constraints a burst size between BMin and BMax,
control plane decisions. The control decisions represented where BMin and BMax denote the minimum and maximum
in terms of actions, rules and policies are cached in the allowed burst size respectively, since very small burst size
flow tables [4] or cross-connect tables [1] of programmable increases the control plane overhead while very large burst
switches. A controller maintains the network state information, size may not be supported at edge nodes due to the buffer
such as network connectivity and wavelength occupancy in capacity limitations. Let, Be denotes the estimated burst size.
formation centrally. In the wavelength occupancy information, Algorithm 111.1: ADAPTIVE BURST ESTIMATION(De, M)
time is slotted, and the state of each wavelength in a time
, if M < BMin or De < BMin
{
slot is represented by a binary variable; l' denotes that the
then { Be +- BMin
wavelength is available in a time slot, and 0 denotes that the
else if !vI > BMax and De > BMax
' '
QoS-aware unified control protocol consists of adaptive-burst denotes the estimated burst size based on the network state
assembling, latency-aware burst routing and scheduling, and information, where Y and PAvg denote channel capacity per
OpenFlow based signaling those are described as follows. unit time slot and an average packet size respectively. M
is selected based on the fact that an availability of a void
A. Adaptive-Burst Assembling Algorithm island with an average size is highly likely in the near future
Various applications have heterogeneous latency require compared to that with the maximum island size. The controller
ments. For example, voice-over-IP, online gaming, video con also estimates a burst size based on the application latency
ference, and IPTV applications require smaller packet latency requirements. If the packets of class C are buffered since
compared to email, online backups, and software download Le Z units of time or in other words the buffered packets
-
applications. An edge node classifies these traffic into different of class C are just the configuration time away by meeting
Forwarding Equivalent Classes (FECs) based on their latency their latency requirements, then the burst size is estimated to
requirements. In the past, many QoS provisioning policies be the size of class C buffer denoted as Dc.
have been investigated to form and treat each class of traffic As shown in Algorithm 111.1, if both the estimated burst
separately [6], [9]. However, these schemes are executed only sizes, M and Dc are greater than BMax, then the procedure
277
ONDM 2013 Brest, France
equipped with wavelength convertors and fiber delay lines, Here, we propose extensions to OpenFlow protocol to
and also ignore deflection routing in the events of burst realize QoS-aware unified control protocol for OBS switching.
contentions. Thus, if the procedure fails to find all optical In [1], the format of a cross-connect table is introduced,
end-to-end connection for the duration of a burst transmission, which defines circuit flows based on time division multiple
then the burst is considered to be blocked. Let us consider for access (TDMA), SONET/SDH, Virtual Concatenation switch
class C 1,2,3, l traffic, and their latency requirements are
=
ing, wavelength switching, and fiber switching. In OBS, such
L1 < L2 < L3 < < Ll, then the latency-aware burst routing circuits are established for predefined time intervals. In order
and scheduling algorithm can be described as follows. to explicitly specify OBS connections, circuit establishment
• Step 1: Construct the bit-map of each wavelength along
and termination time instances need to be included in the
the precalculated K-alternate shortest routes connecting cross-connect tables. Thus, a burst configuration message from
the pair of end nodes by performing the bitwise logical a controller to each core node along the route consists of
end operations on the bit-maps of the same wavelength cross-connection entries with the scheduled establishment and
in all links along the route. termination time instances. A burst establishment message
• Step 2: The procedure finds a set of void islands those
from an edge node to a controller incorporates parameters
have at least a width of Bp' in the bit-map of each such as burst length, source edge node, destination edge node,
wavelength along the K-alternate shortest routes, and the QoS requirement, forwarding equivalent class (FEC), buffer
starting time slot of the found void island is recorded in size of FEC at an edge node, and arrival time of the oldest
a set Q, where Bbst denotes a burst size. buffered packet of FEC at the edge node. On the other hand, a
• Step 3: Among the found void islands Q, a void island
burst release message from a controller to a source edge node
is selected and returned for a traffic class C as follows. incorporates an offset time after which the burst needs to be
minqEQ,q:'OL, ( q ) if C= 1 released and an estimated burst size Be for the future burst
transmissions of class C.
minqEQ,L,:'Oq:'OL2 ( q ) if C=2 D. QoS-Aware Unified Control Protocol
maxqEQ,q:'OL, ( q ) if C=2 In this section, we describe the QoS-aware unified protocol
for OBS switching (demonstrated in Fig 2) and summarized
minqEQ,L2:'Oq:'OL3 ( q ) if C=3 as follows.
maxqEQ,q:'OL2 ( q ) if C=3 • Step 1: A source OBS edge node buffers packets with
while avoiding over-provisioning of network resources. The message to a controller in order to request an edge-to
procedure finds the bit-maps of wavelengths along the K edge optical connection.
alternate routes. In the found bit-maps, void islands of at • Step 3: A controller finds the burst routing and scheduling
least B.;;:" width are searched, and the starting time slots at using the latency-aware burst routing and scheduling
278
ONDM 2013 Brest, France
procedure. The controller also estimates a burst size Figure 3 evaluates burst blocking in the QoS-aware OBS
based on the network state information, QoS (latency) protocol and GMPLS-based distributed protocol as the packet
requirements of applications, and the state of a queue arrival rate increases for each class of traffic. The QoS
at the edge node using the adaptive burst assembling aware OBS protocol improves the burst blocking by at least
procedure. 50% for each class of traffic compared to the GMPLS
• Step 4: The controller sends the estimated burst size based protocol. Such a significant performance improvement
along with a burst release message to the edge node, is due the cumulative effects of the optimized adaptive burst
and cross-connection entries for a scheduled time interval assembling procedure, the latency-aware burst routing and
with a burst configuration message to each core node scheduling procedure, and OpenFiow-based signaling pro
along the selected route. tocol. The adaptive burst assembling algorithm takes into
• Step 5: Upon reception of the burst release message from account additional information such as the network state
the controller, an edge router updates the estimated burst and application latency requirements while forming a burst
size parameter Be for class C traffic. Upon reception compared to the proportional-QoS based burst assembling pro
of the burst configuration message, the received cross cedure that is rigid as shown in Fig 4. Figure 4 illustrates the
connect entry for a scheduled duration is cached in the average burst size in the proportional-QoS based and adaptive
cross-connect table of each node along the selected route, burst assembling algorithms. The adaptive burst assembling
and the core nodes are configured at the scheduled time algorithm intelligently adapts the burst size based on the
instance. network load. At low load, the proposed procedure can find
• Step 6: The burst is released at the schedule time from sufficient network resources for a large burst size with high
the source edge node, and transported to the destination probability. On the other hand, as the network load increases,
edge node transparently. this likelihood decreases, and thus, the procedure decreased the
burst size with the load. On the contrary, the proportional-QoS
IV. SIMULATION RESULTS burst assembling algorithm always assembles a burst of the
The QoS-aware unified OBS protocol is evaluated through a same size irrespective of the network load and the application
custom-built event driven simulator. We consider 6-node mesh, latency requirements. In the latency-aware burst routing and
lO-node mesh, and 14-node NSF networks. Each fiber link is scheduling algorithm, routing and scheduling operations are
assumed to have 10 wavelengths, and each wavelength offers optimized based on the global network state information. On
10 Gb/s line rate. Packet arrival distribution is Poisson with the other hand, in the source-based routing and scheduling
mean arrival rate R. The size of each packet is assumed to be algorithm, these operations are independently optimized at
10 KB. The network supports three classes of traffic; Class A, each edge node based on the local network state information.
Class B, and Class C, those arrive with probability 0.3, 0.3, Thus, lack of up-to-date network state information in the
and 0.4 respectively among the cumulative arrival rate. The source-based routing and scheduling procedure contributes
latency requirements of Class A, Class B, and Class C traffic higher blocking in the GMPLS-based distributed protocol.
are 100 msec, 150 msec, and 200 msec respectively. The upper Furthermore, the latency-aware burst routing and scheduling
limit of OXC configuration time is assumed to be 10 msec. algorithm finds a solution that satisfies the latency requirement
The light propagation latency in the fiber is 0.005 msec/Km. of each class of traffic while maximizing the probability of
The latency-aware burst routing and scheduling algorithm is provisioning low latency traffic.
evaluated by considering K 3 alternate shortest routes, and
= The packet latency for various classes of traffic is demon
the adaptive burst assembling algorithm bounds the minimum strated in Fig. 5 as the network load increases. The QoS-aware
burst size of 500 packets and the maximum burst size of 2000 OBS protocol suffers from higher packet latency compared to
packets. Each traffic class is of type non-preemptive. the GMPLS-based protocol since the bursts that are blocked in
The QoS-aware unified OBS protocol is evaluated with the GMPLS protocol are accepted in the QoS-aware OBS protocol
conventional GMPLS-based distributed protocol that consists with higher latency. As shown in the Figure, the proposed
of the proportional-QoS based burst assembling algorithm, the protocol meets the latency requirement of each class of traffic.
source-based burst routing and scheduling algorithm, and the Furthermore, the OpenFlow-based signaling in the proposed
JET-based signaling protocol. In the proportional-QoS based protocol reduces the burst configuration time compared to the
burst assembling algorithm, a burst is released for a class of JET-based signaling in the GMPLS based distributed protocol
traffic that has the maximum ratio of the current buffer size as shown in Fig. 6. Instead of sequential OXC configurations
and the latency requirement. If the buffered queue size is larger in the distributed protocol, parallel OXC configurations in
than 2000 Packets, then a burst of maximum size is generated. OpenFlow based centralized protocol dramatically improves
In the source-based routing and burst assembling algorithm, the burst configuration time. Furthermore, the gain in the
the source node selects a route out of K-alternate shortest configuration time increases with the network size. Thus, the
routes on which a void island is available at the earliest starting proposed protocol reduces the service provisioning time and
time slot based on the local network state information. Any improves QoS offered to end users.
network event triggers OSPF protocol to update network state Figure 7 demonstrates the network throughput of the QoS
information in the GMPLS-based distributed protocol. aware OBS protocol and GMPLS-based distributed protocol as
279
ONDM 201 3 Brest, France
0.40
- ---i_
III -- - - -_ III
180
~
!W _••_ 0 · _ •. _ . • _ . . _ . • - •• _ 0· _ .... 160.0M ~ IIt--i_ _
III III III
0.35 ~
,/ -e- Proportional·QoS A1go (ClassA) 160
r~
0.30 140.0M
~
j!1 "' I~ ~ -~ Proportional-QoS Alga (ClassC)
.. no AdaptiveA1go (ClassA)
140
:g
~
Q.
0.25
0.20
'"
.5
~
0;
120.0M
, ~ AdaptiveAlga (Class B)
-+- Ada iva AI 0 Class C
~
E 120
c
...-...-....._._._.-... _ ._ ._ ._ ._ ._ ._ ._ .~
- _ GMPLS·basedc asteless A)
~ 0.15 i? ~ 100 _ GMPLS-based OaS (e less B)
~ ~ 1 00.0M ~ - Ii> GMPLS·basedOaS(e lessC)
~ 0.10
0.05 .. . _<It"
..... .k . _ . _ . _ . - _ GMPLS.basedCBS (ClassA)
. . GMPLS-basedaBS (ClassB)
s 80.0M
..J 80
60
-* OoS-Aware OBS(ClassA)
-+- OoS-Aware OBS(ClassB)
--<if- OoS-AwareCBS ClassC
. - GMPLS·basedCBS (ClassC)
J\o- CoS-AwareCBS (ClassA)
*--*-----*----- ------*
0.00
--+--OoS-AwareaBS (ClassB) 40 ~.:::...--..-....... ~ ..-~-~ . _ . _ . _ . _ . ':..:.=_':..:-1
--<if- CoS-AwareCBS ClassC
60 .0M +-~-r-~__.-~__.~-_,_~___,~
4x1!f 5x1cf o 2x1cf 3x1cf 4x10' 5x10' 1x10' 2x1cf 3x10' 4x10' 5x1cf
Packet Arrival Rate (R) Packet Arrival Rate (R) Packet Arrival Rate (R)
Fig. 3. Blocking vs. Load Fig. 4. Burst Size vs. Load Fig. 5. Latency vs. Load
280