Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2023 6th International Conference on Information Systems and Computer Networks (ISCON)

GLA University, Mathura, India. Mar 3-4, 2023

Implementation and Performance Evaluation of


Load Balanced Routing in SDN based Fat Tree
Data Center
2023 6th International Conference on Information Systems and Computer Networks (ISCON) | 979-8-3503-4696-1/23/$31.00 ©2023 IEEE | DOI: 10.1109/ISCON57294.2023.10112200

Dhanya Raj P. Anitha V. S.


Department of Computer Science and Engineering Department of Computer Science and Engineering
Government Engineering College, Wayanad Government Engineering College, Wayanad
APJ Abdul Kalam Technological University APJ Abdul Kalam Technological University
Kerala, India Kerala, India
dhanyaraj@gecwyd.ac.in anithavs@gecwyd.ac.in

Abstract—As the number of data intensive and distributed corresponding to precomputed paths to be stored in the
applications increase, data centers are highly utilized for their forwarding element at the time of flow creation and avoids
storage, processing and accessing. These applications cause per packet computation [2]. So selection of a path from
throughput sensitive incast traffic in the data center. multiple available paths can be done proactively.
Traditional shortest path routing algorithms cause traffic
hotspots in the shortest path links and lead to congestion. POX [4], Beacon [6], Floodlight [5], and Ryu [7] are the
Therefore, load balanced routing strategies need to be major controller platforms for implementing Software
employed to spread the traffic across multiple paths. This work Defined Networking. None of these controllers provide
employs the idea of multi path routing to achieve load inbuilt module for multi path support. They provide
balancing in SDN based Fat tree Data center. It implements Spanning Tree Protocol(STP) to deal with ARP broadcast
Equal Cost Multi Path routing algorithm on Fat tree Data streams occurring in the topologies having loops. But STP
Center Network in Ryu SDN controller and evaluates its load result in the selection of a single path and load balancing
balancing performance against traditional Shortest Path along multiple available paths cannot be performed. This
Routing algorithm under different data center benchmark paper implements a load balanced routing strategy for Fat
traffic scenarios. We have conducted an exhaustive study about tree based data center that works on software defined
the load balancing effect of ECMP on Ryu based SDN Fat tree network architecture. Here, Mod-N based Equal Cost Multi
data center and found that ECMP increases Average network
Path routing algorithm is implemented on top of the Ryu
throughput by 68% and Average Link utilization by 60%. Also
SDN controller. Performance of this algorithm in terms of
packet drop is decreased by 44% with ECMP.
average network throughput, average link utilization, and
Keywords—Data Center, Multi path routing, Software packet drop are compared with the implementation of
Defined Networking, Load Balancing, Fat Tree traditional Shortest Path Routing algorithm under different
benchmark traffic patterns to observe its load balancing
effect in the data center. This work utilizes parallel TCP
I. INTRODUCTION flows to conduct an exhaustive study of the load balancing
Data Centers Networks(DCN) host thousands of effect of ECMP Mod-N strategy in Ryu SDN controller
networked servers to provide storage and computing to based Fat tree data canter.
business organizations. They are used to store, process and
access data intensive and distributed applications like Map The rest of the paper is organized as follows: Section II
Reduce applications, social networking platforms, large scale explains the background topics and Section III gives related
web search etc. These applications create high volume of works. The proposed ECMP routing framework for SDN
throughput intensive east west traffic in the DCN. To provide based Fat tree DCN is given in section IV. Section V
network efficiency and to satisfy the service requirement of describes the implementation details and section VI gives
diverse flows in the data center, congestion possibility should results and a discussion on it. Paper concludes in section VII.
be avoided by sharing the traffic load across available
alternative links. Traditional shortest path routing in a heavy II. BACKGROUND
traffic scenario results in the creation of congestion hot spots Data Centers are highly utilized for their storage,
due to single best path selection. This demands for load processing and accessing by different applications.
balanced routing techniques in the data center. Traditional TCP mechanisms are not suitable for data center
Inefficiency of traditional network technologies in traffic and fails to prevent the occurrence of congestion [15].
handling the large network of thousands of servers in the Also commonly used shortest path routing strategy results in
data center network introduced Software Defined the selection of the same least cost path in all the
Networking into it. SDN uses centralized software communications between a source-destination pair. Thus
controllers to connect with switches in the network and shift Data centers need to avoid the creation of traffic hot spots
the routing logic implementation from the switch to the through appropriate load balance based network layer or
controller. Also, in SDN paradigm, switching/forwarding transport layer mechanisms.
element can store flow state in flow table as flow entries. Fat tree is a symmetric, non-blocking, fault tolerant, and
Thus each forwarding element can store flow state without switch-centric data center topology. Fat tree connect the
any additional overhead. This allows next hops

979-8-3503-4696-1/23/$31.00 ©2023 IEEE 1


Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on December 11,2023 at 03:39:45 UTC from IEEE Xplore. Restrictions apply.
network elements with three layers. A k-ary fat tree consist the literature in this context. Static mechanisms select a path
of 5k2/4 k-port switches and k3/4 servers. A pod in the fat from the alternative choices by some fixed rules. They do not
tree is a group of (k/2)2 servers, k/2 lower layer Edge consider network load or network element status while
switches, and k/2 middle layer Aggregation switches. Each making the decision whereas dynamic mechanism collect
Edge switch connects with k/2 servers and k/2 Aggregation network statistics, measure network load and choose path
switches. Each Aggregation switch connect with k/2 core according to the load. Simplicity of selection policy and low
switches to complete the connectivity. There exists (k)2 /4 overhead of implementation are the characteristics of static
equal cost paths between any two hosts on different pods. mechanisms.
This work [10] implements ECMP Mod-N Hash on a Fat
tree topology on a POX controller. It runs three TCP and
UDP background traffic between two selected hosts in the
network and measures network performance against shortest
path routing algorithm. Nikolaosc et al in the work [14]
analyze multi path routing algorithms on a Spine-leaf data
center and argues for packet-level multi pathing. In this work
[9] Eric Jo et al model load balanced multipath routing
problem as a linear program and implements ECMP and
their own dynamic load balanced multi path routing. This
Fig. 1. Fat Tree Topology
work [16] implements ECMP in a tree topology over Ryu
Software Defined Networking(SDN) is a new network SDN controller and performs UDP traffic test between fixed
paradigm that makes network more flexible and controllable hosts. Here in this work [17] ECMP and a weighted-ECMP
through their global network visibility and programmable are implemented and compared. X. Shi et al in [18] give
nature. SDN architecture comprises of three layers namely, different treatment for elephant flows in path selection and
Data plane, Control plane, and Application plane [1]. tests their strategy on a custom topology with six hosts and
Switches in the control plane communicate with the switches over a Pox controller.
controller using OpenFlow protocol. The interface between Our work is limited to static load balanced routing
control and data plane is known as Southbound interface. method and it implements Equal Cost Multi Path routing on
OpenFlow is the standard south bound interface. The SDN based fat tree data center using the next hop selection
interface between control and application plane is known as method, Mod-N with hash over packet header fields. Here,
Northbound interface. There is no standard North bound we measure the throughput of the network, packet drop and
interface. Controllers provide their own API to write average link utilization of the ECMP routing algorithm along
application programs. with that of Shortest Path Routing algorithm under different
Multi Path routing is applied where multiple equal cost data center benchmark traffic patterns to analyze their load
paths are available between a source-destination pair. As the balancing effect.
selection of a path is from a set of equal cost paths, this type
of routing is known as Equal Cost Multi Path routing IV. PROPOSED ECMP ROUTING FRAMEWORK ON
(ECMP). ECMP implementation is defined in RFC 2991. SDN FAT TREE DATA CENTER
Both flow based and packet based implementations of ECMP Figure 2 shows the architecture of the proposed ECMP
are possible. In the case of packet based ECMP, next hop is routing framework. Arrival of a flow with no matching flow
selected without regard to other packets in the flow. rule installed in the switch causes Packet-In message to the
Generally flow based implementations of ECMP routing are SDN controller. Controller invokes Packet-In handler in
preferred over packet based implementations as latter cause order to find the port of the sending switch through which
packet reordering at the receiving side [3]. Also in SDN
architecture, OpenFlow switches support flow matching and
hence per flow state can be maintained in it. Fat tree based
data center has many equal cost paths between a given
source destination pair. This makes it suitable for applying
ECMP based routing for load balancing.
Multi path load balancing problem on fat tree data
centers can be represented using a directed graph, where
nodes correspond to hosts and switches. Edges correspond to
host to switch links and switch to switch links. Then for an
incoming flow, routing solutions that provide assignment of
a path from the set of equal cost paths, which keeps the
average link utilization and throughput at optimum values
are required.

III. RELATED WORK


The term load balancing in network by default refers to
server load balancing. But here we discuss about balancing
traffic load among alternative available equal cost paths.
Fig. 2. Load Balanced Routing Framework in Ryu based Fat tree DCN
Both static and dynamic load balancing mechanisms exists in

2
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on December 11,2023 at 03:39:45 UTC from IEEE Xplore. Restrictions apply.
this packet is to be forwarded, so that it will reach the for the interested events and can handle it through its own
destination. Source to destination path computation is done at instruction.
this point to find multiple equal cost shortest paths between
them. We emulate a 4-ary Fat tree that has 16 hosts and 20
switches to make Data center topology. Emulation
Selection of one path from these multiple equal cost parameters are given in the Table I. Each link is given 1GB
paths are done using Hash based Mod-N operation. bandwidth. The Data center topology is connected with Ryu
Algorithm 1 describes it. Here G indicates the graph to implement SDN based architecture. An application is
representation of the network, H is the set of hosts, S is the written to compute equal cost shortest paths among a
set of switches and L denotes the set of links. particular source destination pair. Path selection module
selects a path based on the ECMP Hash Mod-N algorithm
discussed in the design section. Newly selected path
information is installed among the switches along the path.

TABLE I. EXPERIMENT SET UP

Experiment Set up Values


Number of hosts 16
Number of switches 20
Link Bandwidth 1 GB

In Mod-N hash based selection, flow is identified at the Switch to controller protocol Open Flow V1.3
micro level using source and destination ip addresses, source
and destination port numbers and protocol fields. A unique
hash value is generated for each such flow. A mod-N Two types of traffic can be seen in the data centers, east-
operation on this hash value, where N is the number of equal west traffic and north-south traffic. East-west traffic is the
cost paths available between the source and the destination, traffic between server racks in a data center and north-south
is performed to get the choice number for the selection of traffic is the communication coming and going from internet
path. Path corresponding to this choice number is selected to the data center. Studies[12] have shown that east-west
from the list of all paths. traffic contributes about seventy seven percentage of total
data center traffic. Due to the difficulty in getting real data
Major steps in the load balanced routing framework center traffic traces, we have generated data center
design are listed below: benchmark communication patterns as per[13] namely
stride(i), staggered(Pe,Pp), and random. Stride pattern with
1) Collect network topology information into a graph G
argument value 1, 2, 4 and 8 generates east-west traffic
2) Start network status monitoring to collect performance between the server hosts. When 1 is used as the argument,
evaluation parameters each server sends to the next neighbour in the fat tree
topology and the traffic pattern is h1 to h2, h2 to h3...., h15
3) Initiate parallel network traffic between source to h16 and h16 to h1. With arguments 2, 4 and 8
destination pairs, that can test the effectiveness of load communication distance between the hosts increases. Stride
balanced routing strategy pattern generates both intra pod and inter pod
4) For an incoming flow with no matches in the flow communications.
table of a switch, find the list of alternative paths between the With staggered pattern a server sends traffic to a server
source-destination of that flow within the same edge switch with a probability of Pe and to
5) Select a path based on Modulo-N Hash based ECMP the server in the same pod with a probability Pp and to the
load balancing route selection strategy, as explained in the remaining servers in the network with 1 − Pe − Pp. In
Algorithm 1 Random pattern, for each source host, the destination host is
selected randomly. Iperf tool [8] is used to generate TCP
6) Install flow entry on all switches along the selected traffic flows between hosts in the network. Parallel flows are
path applied to create multiple flow between a source destination
ECMP Hash Mod-N algorithm assigns each flow to a pair. All the traffic patterns are implemented and applied
particular path between the source destination pair. So it do over the hosts in the fat tree data center multiple times to take
not require packet reordering at the receiver side. average measurements.

V. IMPLEMENTATION VI. RESULTS AND DISCUSSION


We implement SDN based fat tree Data center using Ryu We measured the average throughput of the network,
SDN controller and Mininet emulation tool. Mininet[11] average link utilization and packet drop under the above
creates realistic virtual network on a single machine using mentioned traffic patterns for traditional Shortest Path
Linux network name spaces. The virtual network consist of Routing and ECMP hash based Mod-N. Received bytes from
virtual hosts, switches, controllers and connection links. each host in the network is collected during fixed time
Ryu[7] is a Software defined networking controller, written intervals. This data is used to compute average network
in Python. It provides APIs to write network management throughput. The values of average network throughput on
applications. A Ryu application can write decorator functions running the ECMP Mod-N path selection algorithm on
different traffic patterns are plotted in the Figure 5. ECMP

3
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on December 11,2023 at 03:39:45 UTC from IEEE Xplore. Restrictions apply.
Mod-N give high throughput compared to single shortest
path routing in all the patterns.

Fig. 4. Throughput comparison of the ECMP algorithm on different traffic


patterns with Time

Fig. 3. Throughput comparison of the SPR algorithm on different traffic


patterns with Time TABLE II. PERFORMANCE ANALYSIS
Performance Improvement
Figures 3 and 4 shows the throughput performance Parameter
SPR ECMP Mod-N Difference %
comparison of the different patterns over time with SPR and Throughput(Gbps) 5.51 9.23 3.72 68
ECMP Mod-N algorithms. On closely analyzing the Link Utilization(%) 34.27 54.87 20.61 60
performance, we can see that even though both the
Packet Drop(x1000) 153.47 86.71 66.76 44
algorithms performs equally on stride(1) pattern that has very
less number of inter pod traffic, on stride(4) and stride(8)
patterns in which inter pod communication is high and core average of link utilization of all links in the network is
switch is involved in the path from source to destination computed under different traffic patterns for the two
route, ECMP-Mod N clearly gives high throughput. Also on algorithms. Figure 6 clearly shows that average link
Staggered(0.2,0.3) pattern that has fifty percentage inter pod utilization is high for ECMP Mod-N compared to SPR in all
communication, ECMP Mod-N algorithm over perform the traffic patterns. Overall increase in link utilization is
single Shortest path routing. Average increase in throughput 60%. This clearly shows the spreading of flows along
of all patterns, in ECMP Mod-N is 68%, as shown in Table multiple available paths by ECMP Mod-N strategy.
II.
Packet drop in the network is measured and plotted in the
Plotting average link utilization caused by different Figure 7. ECMP Mod-N has reduced packet drop in all the
routing algorithms under same network load provides a clear traffic patterns. Overall reduction in packet drop with ECMP
indication of their load balancing effect. Here, link utilization over SPR is 44%. High network throughput, link utilization
is calculated from the switches with the number of bytes and reduced packet drop clearly shows that ECMP transmits
received information on each switch port. Threads are the traffic successfully to the destination and spreads it
created to collect port statistics from all switches in the fat evenly.
tree data center topology. Port status request messages are
sent from the controller and a reply handler is written to save
relevant information and to calculate link utilization. The

Fig. 5. Average Network Throughput comparison of SPR and ECMP Mod-N algorithms on different traffic patterns in SDN Fat tree Data Center

4
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on December 11,2023 at 03:39:45 UTC from IEEE Xplore. Restrictions apply.
Fig. 6. Average Link Utilization comparison of SPR and ECMP Mod-N algorithms on different traffic patterns in SDN Fat tree Data Center

Fig. 7. Packet Drop comparison of SPR and ECMP Mod-N algorithms on different traffic patterns in SDN Fat tree Data Center

[3] C. Hopps, ”Analysis of an equal-cost multi-path algorithm. RFC


VII. CONCLUSION 2992, Internet Engineering Task Force”, 2000.
[4] Pox. https://noxrepo.github.io/pox-doc
This paper presents a load balanced routing framework
[5] Floodlight.
for a Ryu based SDN Fat tree data center. ECMP Mod-N https://floodlight.atlassian.net/wiki/spaces/floodlightcontroller
static path selection strategy is implemented in Ryu SDN
[6] D. Erickson, ”The beacon openflow controller.”, Proceedings of the
controller to select a path from available multiple paths in the second ACM SIGCOMM workshop on Hot topics in software defined
Fat tree data center topology. The performance of this networking, 2013.
algorithm along with traditional Shortest Path Routing [7] Ryu. https://ryu-sdn.org/
algorithm, in terms of average network throughput, average [8] Iperf. https://iperf.fr
link utilization and packet drop are measured for different [9] E. Jo, D. Pan, J. Liu, L Butler, ”A simulation and emulation study of
data center benchmark traffic patterns, generated with SDN-based multipath routing for fat-tree data center networks.”
parallel flows. ECMP Mod-N increased the throughput by Proceedings of the Winter Simulation Conference 2014. IEEE, 2014.
68%, link utilization by 60% and reduced the packet drop by [10] F. Rhamdani, N.A. Suwastika, and M.A. Nugroho, ”Equal-cost
44%. The performance of ECMP Mod-N shows that multipath routing in data center network based on software defined
distributing traffic across multiple paths can increase average network.” 2018 6th International Conference on Information and
Communication Technology (ICoICT). IEEE, 2018.
network utilization and throughput and decrease packet drop.
[11] Mininet. http://mininet.org/
This inspires to conduct further studies that can utilize the
[12] Index, Cisco Global Cloud. ”Forecast and methodology, 2015-2020
multi path routing for further improvement in Fat tree Data white paper.” Retrieved 1st June (2016): 15.
center load balancing.
[13] M. Al-Fares, S. Radhakrishnan, B. Raghavan, N. Huang, and A.
The implementation of ECMP Mod-N on Ryu based Fat Vahdat, ”Hedera: dynamic flow scheduling for data center networks.”
Nsdi. Vol. 10. No. 8. 2010.
tree data center with the benchmark communication pattern
[14] N. Chrysos, et al. ”High performance multipath routing for
can be used as a test bed for future investigations towards datacenters.” 2014 IEEE 15th International Conference on High
load balancing routing. Further research is needed to find Performance Switching and Routing (HPSR). IEEE, 2014.
enhanced techniques to spread the traffic more uniformly [15] ] T. Hafeez, N. Ahmed, B. Ahmed, and A.W. Malik, ”Detection and
across the available multiple paths and thereby improving the mitigation of congestion in SDN enabled data center networks: A
load balancing performance of the routing strategies in the survey.” IEEE Access 6 (2017): 1730-1740.
data center. [16] H.E. Wahanani, M. Idhom, and E.P. Mandyartha, ”Equal cost
multipath ryu controller analysis in software-defined networking.”
2020 6th Information Technology International Seminar (ITIS). IEEE,
REFERENCES 2020.
[1] Goransson, P. Software Defined Networks (Second Edition), A [17] E. Nepolo, and G.A.L. Zodi, ”A Predictive ECMP Routing Protocol
Comprehensive Approach (2017). for Fat-Tree Enabled Data Centre Networks.” 2021 15th International
[2] D. Thaler and C. Hopps., ”Rfc2991: Multipath issues in unicast and Conference on Ubiquitous Information Management and
multicast next-hop selection.”, 2000 Communication (IMCOM). IEEE, 2021.

5
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on December 11,2023 at 03:39:45 UTC from IEEE Xplore. Restrictions apply.
[18] X. Shi, L. Yangyang, X. Haiyong, Y. Tengfei, Z. Linchao, L. Panyu,
Z. Heng, and L. Zhiyao, ”An openflow-based load balancing strategy
in SDN.” Comput. Mater. Contin 62, no. 1 (2020): 385-398.

6
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on December 11,2023 at 03:39:45 UTC from IEEE Xplore. Restrictions apply.

You might also like