REDUCING LATENCY IN AFRICAN NRENS USING PERFORMANCE-BASED LISP/SDN TRAFFIC ENGINEERING

David C. Wyld et al. (Eds) : NETCOM, NCS, WiMoNe, CSEIT, SPM - 2015
pp. 83–98, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.51608
REDUCING LATENCY IN AFRICAN
NRENS USING PERFORMANCE-BASED
LISP/SDN TRAFFIC ENGINEERING
Josiah Chavula, Melissa Densmore, and Hussein Suleman
Department of Computer Science, University of Cape Town, South Africa
{jchavula,mdensmore,Hussein}@cs.uct.ac.za
ABSTRACT
Active topology measurements on the African Internet have showed that over 75% of the intra-
Africa traffic destined for Africa's National Research and Education Net- works (NRENs) uses
intercontinental links, resulting in high latencies and data transmission costs. The goal of this
work is to investigate how latency-based path selection using Locator/Identifier Separation
Protocol (LISP) and Software Defined Networking (SDN) in NRENs can be used to reduce
inter-NREN latencies. We present aspects of an experimental prototype implementation for real-
time topology probes to discover lower-latency remote gateways and dynamic configuration of
end-to-end Internet paths. Simulation results indicate that ranking remote ingress gateways, and
dynamic configuration of end-to-end paths between gateways can lower the average latency for
inter-NREN traffic exchange.
KEYWORDS
Software Defined Networking, Traffic Engineering, Latency, Internet exchange points, African
National Research and Education Networks (NRENs)
1. INTRODUCTION
Research collaboration and e-resource sharing in sub-Saharan Africa continues to be hampered by
the limited interconnectivity that is not only expensive, but fails to meet the quality of service
(QoS) required for collaborative applications among the National Research and Education Net-
works (NRENs). Despite the ongoing interconnection efforts by African NRENs, over 75% of the
inter-NREN traffic is being exchanged through circuitous routes traversing inter-continental links
and Internet exchange points in Europe and North America, thereby experiencing high latencies
[1].
Efforts have been made to improve the Internet traffic exchange among the NRENs in Africa. A
key player in this effort is the UbuntuNet Alliance, an association of the NRENs in Eastern and
Southern Africa. Between 2011 and 2014, the UbuntuNet Alliance has been implementing the
Africa-Connect Project to create a regional research and education Internet network
interconnecting the NRENs in the region. The project has aimed to interconnect Southern and
Eastern African NRENs into a regional network through the use of terrestrial network facilities.
The project involves establishment of Points of Presence (PoPs) in major cities in the region -

84 Computer Science & Information Technology (CS & IT)
notably in Mtunzini, Maputo, Dar es Salaam, Nairobi, Kampala and Kigali, and interconnecting
them with broadband cross-border links to create a regional research network. So far, the intra-
Africa interconnection serves six NRENs: TENET (South Africa), MoRENet (Mozambique),
TERNET (Tanzania), KENET (Kenya), RENU (Uganda) and RwEdNet (Rwanda). Furthermore,
transcontinental links between Nairobi and the Ubuntunet Alliance PoP in Amsterdam, as well as
from Cape Town to London, have been established, thereby linking UbuntuNet Alliance with
GEANT, the European research network (Figure 1).
Fig. 1. The UbuntuNet Alliance regional network
The establishment of multiple PoPs, multiple intra-Africa Internet links, as well as multiple
transcontinental links, provides new opportunities for implementing multipath routing and traffic
engineering mechanisms with the aim of improving performance of traffic exchange among
Africa's NRENs. NRENs could implement mechanisms that would allow them to announce to
each other, and make use of multiple Internet attachment points to exchange traffic. One protocol
that allows edge networks to exchange traffic over multiple Internet gateways is the
Locator/Identifier Separation Protocol (LISP) [2, 3]. LISP allows edge networks to announce

Computer Science & Information Technology (CS & IT) 85
multiple Internet gateways, known as Route Locators (RLOCs), and to influence the selection of
incoming paths. Through a mapping system, LISP allows networks to announce preferences for
multiple RLOCs. The availability of multiple locators for the same destination increases path
diversity[4] and enable multipath routing, as source networks can select among multiple gateways
to reach a destination network.
This paper demonstrates the potential for performance improvement in the Pan-African NRENs
by employing, at the network edge, traffic engineering techniques that are based on end-to-end
multipath ranking. Using a Software Defined Network (SDN) experimental topology, and a LISP
mapping system, the paper examines the potential for dynamically ranking egress and ingress
links between multihomed NRENs based on end-to-end path metrics. The objective is to
minimize latency for two-way delay sensitive traffic flows (e.g. real-time classroom streaming
and video conferences between universities), and minimize intercontinental bandwidth utilization.
2. BACKGROUND AND RELATED WORK
Given the opportunities for traffic engineering provided by the multiple intra-continental and
transcontinental links provided by the Africa Connect network, one way of improving the
performance and optimization of traffic exchange across African NRENs is to enable dynamic
selection of optimal traffic exchange routes based on application QoS needs. For example, path
selection for delay sensitive applications can be made based on prevailing end-to-end latencies
through either the intercontinental transit links or through the intra-Africa links.
2.1 LISP-based Multipath Traffic Engineering
The Locator/Identifier Separation Protocol (LISP) [3] simplifies multihoming. LISP divides the
Internet's address space into two - locally routable Endpoint IDentifiers (EIDs) and global Route
LOCators (RLOCs). By separating the host address space from the locator address space, LISP
introduces a level of indirection that allows networks to specify preferences for multiple ingress
gateways (locators). The availability of multiple locators for the same destination increases path
diversity since the source networks are able to forward traffic for a particular destination through
multiple remote locators (gateways).
Thus, additional end-to-end performance gains can be achieved with the ability to dynamically
select the ingress link at the destination network. In Figure 2, for example, edge networks A and
B are multi-homed to networks x,q and y,z respectively. Depending on how the routing is done in
the Internet core, the choice of the egress link by network A, i.e. (A,x or A,q), has potential to
influence selection of ingress link towards B i.e. 3,y or 4,z. Since each end-to-end path has its
own unique path metrics in terms of bandwidth, delay, and loss, selection of particular egress and
ingress links at A and B impacts the overall quality of the end-to-end path.

Fig. 2. Multihomed Networks A and B, multihomed through providers (x,q) and (y,z) respectively
One prominent work that uses LISP for traffic engineering is the ISP-Driven Informed Path
Selection(IDIPS) [5]. IDIPs is a request/response service where centralized server nodes perform
network measurements towards popular destinations, and clients request path rankings for a set of
sources and destinations. The IDIPS server ranks the available paths based on a client's ranking
preference and measured path metrics. In IDIPS implementation, path ranking is further
influenced by the destination's preferences in the locator mapping. The selected paths therefore
reflect not only the source network's ranking criteria, but also the destination's preferences for
incoming traffic.
2.2 SDN-based Path Enforcement
One challenge with inter-domain multi-path routing and end-to-end traffic engineering is with
regard to enforcement of paths across different domains. Software Defined Networking (SDN)
provides new opportunities for flexible management of Internet routing and packet forwarding
[6]. An SDN-based IXP [7] allows IXP participants to have access to an SDN controller and to
write policies that override the default policies of the IXP's BGP route server. SDN has three
important characteristics that are useful for interdomain traffic engineering [7]. Firstly, in contrast
to traditional switches that forward traffic based only on the destination MAC address, SDN
enables packet forwarding based on multiple header fields. Secondly, an SDN controller
consolidates control messages from multiple remote networks, such that source and destination
networks can remotely configure forwarding paths through a controller. Thirdly, the controller's
direct control of the data plane enables dynamic/programmatic configuration of the forwarding
tables. With these SDN opportunities, it is possible to allow edge networks some control over
selection of inter-domain forwarding paths at Internet exchange points, thereby having more
control on the end-to-end paths.
3. A MODEL FOR PERFORMANCE-BASED PATH SELECTION
The traffic engineering framework depicted in Figure 3 is based on the ability of traffic source
gateways to select the destination's ingress link based on metrics of the edge-to-edge path. This
requires that the source gateway should have a mechanism for learning the destination's multiple
ingress links.

Fig. 3. LISP/SDN Multi-Path Traffic Engineering Model
In summary, the model works in the following manner: For each new traffic flow from a local
source node to some remote network, the source network queries the LISP mapping system
(through an RLOC lookup API) to obtain the destination network's RLOCs. After obtaining the
remote RLOCs, a Locator Ranking module performs active measurements towards remote
RLOCs. The Locator Ranking module uses the network metrics obtained from the active
measurements to rank the local and remote gateways, after which it updates the local mapping
cache. For each new traffic flow, the source network selects the egress and ingress gateways
based on the rankings in the mapping cache. Once the egress and ingress links are selected, a
Circuit Pusher module is invoked to configure, through an Open Flow SDN control, an end-to-
end switching path between the source and destination RLOCs via the selected egress/ingress
links. If no mapping exists in the mapping cache for a destination network, the RLOC Mapping
module is invoked to perform the lookup, after which the RLOC Ranking module performs the
ranking.
3.1 Path Performance Measurement
The key network metrics in this model are end-to-end latency, jitter and packet loss. Latency,
measured as round trip time (RTT) for traffic to move from source to destination, and for the
acknowledgement packet to be received by the sender, is an important characteristic that affects
the performance and responsiveness of Internet applications. Jitter, on the other hand, is the
variation in latency over time between a traffic source and destination node. Packet loss is a
measure of the percentage of packets lost along the data path for each traffic flow. To obtain
traffic characteristics of the network in terms of the key path metrics, active measurement
techniques are used. In particular, a ping-based tool is used for sending probe packets, from each
edge-network gateway, to the destination networks' RLOCs. Upon retrieving the destination
network's RLOCs from the mapping server, the source gateway sends a ping probe, through each
of its egress links, to each of the destination's RLOCs. By analysing the solicited responses,
topological characteristics such as round-trip-times, jitter and packet loss are obtained. The values
obtained from the responses are used to rank the different paths.

3.2 LISP-based Path Ranking
With LISP, multihomed edge networks are able to achieve some degree of path diversity, as
multiple alternate gateways become visible between source and destination networks. Achieving
optimal end-to-end performance in such environments requires that the source and destination
networks should be able to evaluate the alternate paths, and to dynamically select both the source
network's egress link and the destination network's ingress link. In particular, the source network
needs a way of discovering and evaluating end-to-end links through alternate egress and ingress
links.
The routing cost for an edge-to-edge path can be modelled as a vector comprising the measured
performance metrics and the network RLOC preferences [8]. Let P(Axy); P(Byx) be the
performance cost vectors for the two edge networks A and B, with respect to forwarding traffic
through their access links x and y respectively. The performance cost P from each edge network
comprises a set of end-to-end path metrics K, (eg. latency, packet loss, jitter) weighted by
variable _, such that:
To calculate the total cost, T(Axy); T(Byx), the source preference cost (ɸ(x)) is combined with the
performance cost P, using a variable scaling factor λ:
In LISP implementation, selection of the local egress RLOC as well as the remote ingress RLOC
is determined by priority and weight values recorded in RLOC records that are retrieved from the
mapping system and stored in a local cache. If multiple locators for the same destination exist, the
priority values, ranging from 0 to 255, are used to select the locator that is most preferred. In this
work, the calculated path costs are translated into RLOC priority values using a log function that
scales the costs into values between 0 and 255. The egress-ingress RLOC pair that has the lowest
resultant cost is the one that is used for the outgoing traffic flow between the locator pair. While
the mapping remains valid, all subsequent matching flows between the locator pair uses the
cached locator ranking. The cached locator mapping remains valid for a configurable TTL period
of 60 seconds, after which the mapping and locator ranking process is repeated.

3.3 End-to-end Path Configuration
Each edge network gateway has a Circuit Pusher, an SDN module for setting up end-to-end
circuits between locators. After the source gateway has selected both the local and remote
locators, it invokes the SDN module to configure, through an SDN network controller, a
unidirectional end-end circuit between locators. This is achieved by installing flow entries on all
switches that are part of the shortest path between two selected locators. The installed path is
unidirectional because each source gateway independently performs RLOC lookups and path
ranking, and sets up a circuit toward the remote RLOC.
4. TESTBED IMPLEMENTATION
The overarching purpose of this study is to evaluate the extent to which LISP and SDN can
support dynamic selection of end-to-end paths between multi-homed edge networks. The
objective was therefore to assess network performance in a topology that uses a dynamic
performance-based path selection, versus static default path routing. Figure 4 depicts the testbed
implementation.
Fig. 4. SDN/LISP Traffic Engineering for dual homed NRENs
4.1 Topology
The testbed was built using a LISP-based SDN simulated network consisting of an SDN
(Openflow) controller, an Open flow switch, and a LISP mapping server. The topology was built
in a virtual environment, using the Mininet network emulator [9]. Mininet allows the creation of
virtual hosts, switches, controllers and links. Furthermore, Mininet nodes run a standard Linux
kernel and net-work stack, and can therefore run real network applications. The switches and
network controllers are based on the standard SDN protocol Open Flow. An Internet-like
topology was built in Mininet and was used together with a LISP mapping system and LISP
transit routers (xTRs), which are based on an open source implementation of LISP called Open
LISP[10]. The interconnection among the edge networks is through ISPs that interconnect at a
common Internet exchange point. More specifically, the network was designed with the following
features:

• 48 routers connected to the IXP switch; the routers represent participant networks at the
IXP. The JINX has, as of December 2014, 48 participant networks.
• Each IXP participant has 10 edge networks connected to it, and each edge network is
dual-homed and serves a total of 5 hosts, representing campus networks. In total, there
were 1,200 end hosts as potential sources and destinations for traffic flowing through the
IXP fabric.
• To simulate multi-homing, each edge network is connected to two provider networks
present at the IXP. The primary link represents the intra-Africa link and is configured
with lower latency, while the secondary link represents an transcontinental link and is
configured with higher latency.
• The access links between the edge networks and the provider networks are configured
with band-width evenly distributed between 2 Mbps and 4 Mbps. The links between the
IXP participants (provider networks) and the IXP are equally provisioned with bandwidth
of 100 Mbps.
• The end-to-end latencies are modelled on latencies measured for traffic exchanged
between African NRENs. Round trip latencies in this experiment are distributed between
60 ms and 700 ms.
4.2 Network Traffic
A realistic evaluation of a network model requires emulating the network with traffic that
characteristically resembles the traffic pattern of the emulated networks. This is important, as a
major scalability issue with centralized network architectures, such as the Open flow controllers
and LISP mapping systems, hinges on their ability to cope with the traffic flow characteristics in
the network [11].
Researchers have characterized Internet traffic based on flow metrics such as byte volume, packet
volume, flow duration, and flow inter-arrival time. The percentage of UDP traffic has increased
with the advent of many UDP based P2P applications and streaming multimedia, which transport
large volumes of data[12]. In 2009, a CAIDA survey showed that the ratio of UDP to TCP traffic
was almost 0.21 in terms of packet numbers, 0.11 in terms of byte count, and 3.09 in terms of
flows.
The length of data flows impacts the relative latency introduced at the controller and,
furthermore, the number of active flows has implications for the size of the forwarding tables
maintained at each Open Flow forwarding device. For example, a characterization of university
campus network traffic [13] established that 21.4% of the traffic was carried by flows longer than
10 minutes, 12.6% by flows longer than 20 minutes, and nearly 2% was carried by flows longer
than 100 minutes. Short flows are bursty and have flow speeds ranging from 1 Bps to over 10
kBps, while longer flows are slower, around 50 Bps for 40 min flows, [13].
The test traffic for the experiment is therefore based on the following Internet traffic
characteristics:
1. Protocol flow: UDP to TCP ratio: 3:1
2. Flow Duration:
- 0 - 2 sec : 45% of all the traffic
- 2 sec - 5 mins : 55% of all the traffic

3. Flow inter-arrival time: 4 ms - 40 ms (or 25 to 250 new flows per sec)
4. Flow rate:
- Short flows (0 - 60 sec) : 1 Bps - 10 kBps
- Medium flows (1 min - 5 mins) : 100 Bps - 50 Bps
4.3 Test Traffic Generation
Some of the most widely used traffic generators include Iperf, PackETH, D-ITG, and Ostinato
[14,15]. PackETH [16] is a stateless packet generator designed for Ethernet networks, and
supports a number of protocols including UDP, TCP and ICMP. Iperf[17] is mostly used for
evaluating topology parameters such as bandwidth, delay, window size and packet loss, for both
TCP and UDP traffic. Iperf provides an estimation of received and transmitted data rates.
Ostinato[14] is a user level traffic generator tool that supports UDP and TCP protocols at multiple
rates. DITG (Distributed Internet Traffic Generator) [18] can generate Internet traffic with a user
defined packet inter-departure times. This work made use of D-ITG for traffic generation and for
obtaining performance metrics.
5. RESULTS
A key objective of the RLOC ranking was to discover and direct traffic flows through lower
latency paths towards multi-homed remote networks. TCP traffic is particularly impacted by
network round-trip-times, and this paper has considered performance of TCP traffic when RLOC
ranking and dynamic path configuration is employed. Furthermore, the evaluation also considers
how jitter is affected due to path ranking and circuit configuration.
5.1 Round Trip Times
The key results from a series of experiments suggest that in cases where the paths towards the
different RLOCs of an edge network have significantly different RTTs, latency based ranking and
selection of RLOCs does help to lower the overall latency in the network. Figure 5 shows the
RTT dispersion and mean for TCP traffic in a LISP/SDN topology, with each flow lasting
between 1sec and 300 sec. The vertical lines represent the dispersion of flow RTTs over the time
interval, with each traffic flow RTT averaged over 2 sec intervals. The instantaneous average
RTT for all the flows is indicated by the blue and red lines. The results show that RLOC ranking
results in a 20 % lower overall latency compared to the default gateway forwarding.
The performance gains from RLOC ranking appear to diminish significantly with increased
network load. As the network gets more congested, the observable gain from RLOC ranking is
significantly reduced. This can be explained from the fact that as the network links get more
congested, the otherwise shorter links begin to exhibit equally higher RTTs. With higher RTTs,
many of the probe packets time-out, prompting the egress RLOC to forward packets towards the
destination's default RLOC. Even when the probe packets generate responses, the RTT values of
the otherwise shorter paths tend to be just as high, thereby being ranked lower and resulting in
selection of the other paths for traffic forwarding. Figure 6 shows the dispersion and mean of the
RTT at the point when the network is congested. Figure 7 and Figure 8 further illustrates the
RLOC ranking effects in normal network operation and when there is congestion.

Fig. 5. Round-trip times for network operating with LISP gateway ranking vs non-path ranking LISP
operation
5.2 Jitter
Apart from latency, jitter is another key metric that affects performance of interactive Internet
applications. Some causes of Internet jitter include congestion in the core network as well as in
the access links. Results from the simulation network suggest that RLOC ranking and dynamic
path configuration does increase the overall jitter in the network. Given a network topology and
traffic profile, RLOC ranking in a LISP network results in higher overall jitter than when no
RLOC ranking is employed. Figure 9 shows the jitter for both RLOC ranking and the normal
LISP operation. On average, RLOC probing is seen to increase jitter by 20 %. The jitter can be
attributed to the delay experienced by some packets during the time the local gateway performs
path measurement and ranking.
However, as the network reaches congestion point, both the RLOC ranking and non-ranking
scenarios experience similarly higher jitter. This is illustrated in Figure 10, where jitter for both
ranking and non-ranking LISP operations have their average jitter increase significantly. As
congestion occurs, the probe engine fails to discover any lower latency RLOCs and resorts to
using default paths.

Fig. 6. Round trip times for path ranked and non-path ranked LISP operation at network congestion point
Fig. 7. RTT without congestion

Fig. 8. RTT with congestion
Fig. 9. Jitter with no congestion

Fig. 10. Jitter with congestion
6. DISCUSSION
6.1 RLOC Ranking During Congestion
The RTT results show that the RLOC ranking mechanism fails to produce positive results under
network congestion. In general, centralized systems are venerable to performance bottlenecks
under system overload. For instance, inter-arrival times of traffic flows have implications on the
performance of centralized network controllers [11]. In an Open Flow network architecture, for
example, a scalability challenge stems from the fact that the first packet of each flow is forwarded
to a central controller, which is responsible for determining and configuring the forwarding path
for the flow. Similarly, for LISP, the egress gateway performs a lookup from a mapping server to
determine each new flow's destination network's RLOC. In either case, the flow inter-arrival time
has a scalability impact on the network, as higher rates for new flows result in bottlenecks at the
SDN and LISP controllers, thereby introducing latency and jitter. Although the scalability and
performance bottleneck would affect both the ranking and normal LISP operations, the RLOC
ranking mechanism would experience more severe impact as it is dependent on receiving replies
from probe packets, which take longer when there is congestion. One way of dealing with RLOC
ranking during congestion is to reduce the amount of probing required by using historical per-
formance information to select the RLOCs. Also, the amount of probing needs to be reduced by
performing ranking only for critical flows (eg. delay intolerant applications).
6.2 Effects of RLOC Ranking on Jitter
The observed jitter in the experiments reveals that the process of RLOC ranking and path
configuration does increase the overall jitter in the network. One way of minimising the jitter is to
reduce the path setup time. In the presented model, a new end-to-end path is not setup until path

probing/ranking and SDN path configuration is complete. Reducing jitter would require pre-
ranking the RLOCs, so that new flows don't wait too long before circuits are set up. Pre-ranking
of paths could be based on RLOC performance history as well as known traffic characteristics in
the network.
6.3 Model Limitations
One challenge with the model presented in this paper is the assumption that NRENs are
multihomed. This is true to a large extent as, in general, campus networks that are part of NRENs
have more than one Internet attachment points; a specific network attachment point for traffic
destined within the NRENs' network and another for traffic destined to non-NREN networks.
Where an NREN has only one Internet attachment point, it can still appear multi-homed by
performing prefix de-aggregation and announcing separate prefixes with different RLOCs. This
would enable them to still benefit from multipath traffic engineering.
Another challenge is to do with independent selection of the remote RLOCs by the source
network, which could result in violation of the destination network's preferences and policies.
This could negatively impact on the destination's policies. For edge networks that have some
form of cooperation, such as the case with NRENs within the UbuntuNet Alliance, a mutually
beneficial approach would be to employ some level of coordination in selection of gateways.
While each NREN would aim to optimize traffic cost and QoS performance (latency),
collectively, the NRENs can optimize the performance of some common preferred applications.
Balancing the individual NREN's optimization objectives with those of the peering domains
requires coordination and routing cooperation among the peers. For UbuntuNet Alliance, benefits
from this level of cooperation could include better performance of the network applications,
reduction in usage of intercontinental links, as well as reduction in the cost of inter-NREN traffic
exchange.
7. CONCLUSION AND FUTURE WORK
Results from this study show that through dynamic ranking of the local and remote LISP locators,
a source network can perform traffic engineering towards a destination without requiring any
form of cooperation in the network. It is evident that by leveraging LISP capabilities through
integration with SDN, there is potential for improving traffic exchange performance. This how-
ever implies granting NRENs more exibility and control of routing and traffic engineering across
Internet exchange points, such as to be able to dynamically select routing paths among multiple
ingress and egress links. This could have important applications for NRENs that experience high
bandwidth costs [19]. The ability to perform multipath routing has potential to over significant
performance enhancements and cost savings. For example, delay-sensitive applications such as
VoIP, can be mapped onto lower latency intra-continental links, while traffic for bandwidth
intensive _le-sharing applications can be routed through higher capacity inter-continental links.
Furthermore, by using Openow's ability to customize the packet forwarding rules, and by
appropriately matching packets into ows using header tags, it is possible to set up application
specific traffic engineering mechanisms. With this capability, a group of NRENs can jointly form
traffic engineering strategies specifically for certain applications of common interest, e.g. inter-
university video streaming, or access to e-library sites within the domain.

This paper has motivated for and described a performance based traffic engineering mechanism
that involves path measurement, gateway ranking, and SDN-based edge-to-edge path
configuration. While Internet infrastructure being implemented by the UbuntuNet Alliance
through the Africa Connect project has great potential to improve traffic exchange among African
NRENs, the inability of traditional protocols to fully take advantage of the available path
diversity remains a challenge. For this reason, African NRENs should, apart from implementing
the physical interconnectivity, also consider appropriate traffic engineering mechanisms to allow
individual NRENs to discover and use optimal inter-NREN paths. However, a globally optimal
solution requires coordination and collaboration among several domains that form part of the end-
to-end path. Future work will investigate mechanisms that would enable African NRENs to
perform collaborative performance based and application specific traffic engineering.
REFERENCES
[1] J. Chavula, N. Feamster, A. Bagula, and H. Suleman, Quantifying the effects of circuitous routes on
the latency of intra-africa internet traffic: A study of research and education networks," vol. 147, pp.
64{73, 2015.
[2] K. Li, S. Wang, and X. Wang, Edge router selection and traffic engineering in lisp-capable
networks,"Communications and Networks, Journal of, vol. 13, pp. 612{620, Dec 2011.
[3] D. Saucez, L. Iannone, O. Bonaventure, and D. Farinacci, Designing a deployable internet: The
locator/identifer separation protocol," Internet Computing, IEEE, vol. 16, pp. 14{21, Nov 2012.
[4] S. Secci, K. Liu, G. Rao, and B. Jabbari, Resilient traffic engineering in a transit-edge separated
internet routing," in Communications (ICC), 2011 IEEE International Conference on, pp. 1{6, June
2011.
[5] D. Saucez, B. Donnet, and O. Bonaventure, Idips: Isp-driven informed path selection," 2008.
[6] C. E. Rothenberg, M. R. Nascimento, M. R. Salvador, C. N. A. Correa, S. Cunha de Lucena, and R.
Raszuk, Revisiting routing control platforms with the eyes and muscles of software-defined
networking," pp. 13{18, ACM, 2012.
[7] A. Gupta, M. Shahbaz, L. Vanbever, H. Kim, R. Clark, N. Feamster, J. Rexford, and S. Shenker,
Sdx: A software defined internet exchange," 2013.
[8] S. Secci, K. Liu, and B. Jabbari, Efficient inter-domain traffic engineering with transit-edge hierar-
chical routing," Computer Networks, vol. 57, no. 4, pp. 976{989, 2013.
[9] B. Heller, N. Handigol, V. Jeyakumar, B. Lantz, and N. McKeown, Reproducible network
experiments using container based emulation," in Proc. ACM CoNEXT, Dec. 2012.
[10] L. Iannone, D. Saucez, and O. Bonaventure, Openlisp: an open source implementation of the loca-
tor/id separation protocol," IEEE INFOCOM, Demo paper, 2009.
[11] T. Benson, A. Akella, and D. A. Maltz, Network traffic characteristics of data centers in the wild," in
Proceedings of the 10th ACM SIGCOMM conference on Internet measurement, pp. 267{280, ACM,
2010.
[12] M. Zhang, M. Dusi, W. John, and C. Chen, Analysis of udp traffic usage on internet backbone links,"
in Applications and the Internet, 2009. SAINT'09. Ninth Annual International Symposium on, pp.
280{281, IEEE, 2009.
[13] L. Quan and J. Heidemann, On the characteristics and reasons of long-lived internet
ows," in Proceedings of the 10th ACM SIGCOMM conference on Internet measurement, pp.
444{450, ACM, 2010.
[14] A. Botta, A. Dainotti, and A. Pescap, A tool for the generation of realistic network workload for
emerging networking scenarios," Computer Networks, vol. 56, no. 15, pp. 3531 { 3547, 2012.
[15] S. S. Kolahi, S. Narayan, D. D. T. Nguyen, and Y. Sunarto, Performance Monitoring of Various
Network Traffic Generators," in International Conference on Computer Modeling and Simulation,
2011.
[16] M. Jemec, packeth{ethernet packet generator," 2012.

[17] A. Tirumala, F. Qin, J. Dugan, J. Ferguson, and K. Gibbs, Iperf: The tcp/udp bandwidth measure-
ment tool," htt p://dast. nlanr. net/Projects, 2005.
[18] S. Avallone, S. Guadagno, D. Emma, A. Pescape, and G. Ventre, D-itg distributed internet traffic
generator," in Quantitative Evaluation of Systems, 2004. QEST 2004. Proceedings. First International
Conference on the, pp. 316{317, IEEE, 2004.
[19] B. Barry, C. Barton, V. Chukwuma, L. Cottrell, U. Kalim, M. Petitdidier, and B. Rabiu, egy-africa:
better internet connectivity to reduce the digital divide," in IST-Africa, 2010, pp. 1{15, IEEE, 2010.
AUTHORS
Josiah Chavula is a PhD candidate at the Hasso-Plattner-Institute in Computer
Science, at the University of Cape Town. His research focuses on Internet
performance and traffic engineering in Africa's National Research and Education
Networks (NRENs). He is interested in approaches for achieving flexible
collaborative Internet traffic engineering using Software Defined Networking
(SDN). He obtained a Master of Science degree (Networking and Internet Systems)
from Lancaster University (UK), and a BSc (Computer Science) from University of
Malawi.
Melissa Densmore is a Senior Lecturer in the Department of Computer Science at
University of Cape Town. Melissa completed her PhD at University of California,
Berkeley in Information Management and Systems, a 3 year ethnographic study of
the use of Internet and mobile technologies by health practitioners and NGO staff
in a health financing program in Uganda, has an MSc in Data Communications,
Networks and Distributed Systems from University College London, and holds a
BA in Computer Science from Cornell University.
Hussein Suleman is an Associate Professor in Computer Science at the University
of Cape Town. His research is situated within the Centre for ICT for Development
(ICT4D) and the Digital Libraries Laboratory. Hussein's main research interests are
in digital libraries, ICT4D, information retrieval, cultural heritage preservation,
Internet technology, high performance computing and computer science education.
He completed his undergraduate degrees and MSc at the then University of
Durban-Westville (now University of Kwazulu-Natal) and finished a PhD at
Virginia Tech in 2002, in the area of component-based digital libraries.

REDUCING LATENCY IN AFRICAN NRENS USING PERFORMANCE-BASED LISP/SDN TRAFFIC ENGINEERING

More Related Content

What's hot (19)

Similar to REDUCING LATENCY IN AFRICAN NRENS USING PERFORMANCE-BASED LISP/SDN TRAFFIC ENGINEERING (20)

More from cscpconf (20)

Recently uploaded (20)

REDUCING LATENCY IN AFRICAN NRENS USING PERFORMANCE-BASED LISP/SDN TRAFFIC ENGINEERING