You are on page 1of 8

(IJCSIS) International Journal of Computer Science and Information Security,

Vol. 8, No. 3, June 2010

A STUDY OF VARIOUS LOAD BALANCING TECHNIQUES IN


INTERNET
M.Azath1, Dr.R.S.D.Wahida banu2,
1
Research Scholar, Anna University, Coimbatore.
1
mailmeazath@gmail.com
2
Research Supervisor, Anna University, Coimbatore.
2
drwahidabanu@gmail.com

Abstract traffic has to be carried in the network in


such a way that performance objectives are
One of the most important applications of fulfilled. In computer networking, load
traffic engineering is load balancing. balancing is a technique to spread work
Successful implementation of load balancing between two or more computers, network
depends on the underlying routing protocol links, CPUs, hard drives, or other resources,
that provides connectivity through the in order to get optimal resource utilization,
Internet by determining the routes used by throughput, or response time.
traffic flows. But the load-balancing
problem is not yet solved completely; new One of the most common applications of
applications and architectures are required to load balancing is to provide a single Internet
meet the existing or incoming fastest service from multiple servers, sometimes
Internet world. And, for greatest impact, known as a server farm. Commonly load
these new capabilities must be delivered in balanced systems include popular web sites,
toolkits that are robust, easy-to-use, and large Internet Relay Chat networks, high
applicable to a wide range of applications. bandwidth File Transfer Protocol sites,
For balancing traffic in internet, packets NNTP servers and DNS servers.
should be reorder, reordering also having a
problem for flows in internet. In Internet, The idea of load balancing is to move traffic
unresponsive flows easily occupy the from congested links to other parts of the
limited buffers, there by reducing the network in a well-controlled way. Traffic
Quality of Service (QoS). In this paper, engineering seeks to effectively balance
various techniques that are adopted for load traffic load throughout existing networks,
balancing in Internet are analyzed. thus achieving QoS demands and
minimizing typical costs of adding hardware
Keywords: and software implementations, common to
Traffic engineering, Load Balancing, network engineering. When dealing with
Internet Services, unresponsive flows, QOS, real-world cases of load balancing, both
Buffer, Traffic splitting and Router. network and traffic engineering is general
purpose tools used throughout all steps of an
1. Introduction implementation [12].

Traffic engineering refers to the 2. Motivation


performance optimization of operational
networks. On one hand, traffic offered Load balancers are an integral part of today's
between origin and destination nodes loads Web infrastructure. They're also complex
the network and on the other hand, this and under-documented pieces of hardware.

149 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

Today's Web sites are complex beasts. site. This is why a firewall is often a suspect,
Every component must work together to too, but to a lesser degree since it is
create a site that is greater than the sum of generally a simpler device than load
its parts. balancers [1].

Figure 2: Load Balancer implementation

3. Needs of load balancing

Load balancing is common in ISP networks.


Figure 1: Traffic flow for a load balancer
If the traffic demands are known, the load
balancing can be formulated as an
The Internet is connected to the routers,
optimization problem. However, knowledge
which pass traffic through a firewall to the
of traffic demands is often lacking. Recent
load balancers, which distribute the traffic to
trends in load balancing points toward
the Web servers, which pass information to
dynamic protocols [13]. These protocols
the application server bone, and the
map the traffic of router onto multiple paths
application server bone is connected to the
and adapt the share of each path in real-time
database server bone. We get the picture. If
to avoid hot-spots and cope with failures.
one component or piece of the process fails,
Dynamic load balancing needs schemes that
it can take down the entire site.
split traffic across multiple paths at a fine
granularity.
Load balancers are also in the direct path of
all traffic to a particular Web site. By
Since the power of any server is finite, a
looking at Figure 2 below, we can see that if
web application must be able to run on
the load balancer stops working, the entire
multiple servers to accept an ever increasing
site stops working. This critical position in
number of users. This is called scaling.
the infrastructure can make it appear as
Scalability is not really a problem for
though the load balancer is the problem,
intranet applications since the number of
even in cases where it is not (such as a
users has little chances to increase.
firewall issue, a back-end database problem,
someone tripping over a cable, etc.). Unlike
However, on internet portals, the load
a broken or malfunctioning Web server, a
continuously increases with the availability
miss configured or malfunctioning load
of broadband Internet accesses. The site's
balancer will result in a dead-to-the-world
maintainer has to find ways to spread the

150 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

load on several servers, either via internal are commonly based on the Border Gateway
mechanisms included in the application Protocol, used for data-exchange between
server, via external components, or via large Internet operators.
architectural redesign. 3. DNS-based load balancing is a popular
way of distributing traffic amongst a set of
Further complication increases due to Internet addresses by returning a list of
various factors such as: active addresses to the requesting client.
These addresses can point to a set of servers
• Sizes of objects might not be the or even a set of geographically separate
same. sites.
• Object IDs might not be chosen at 4. Sites can connect to the net through
random. several links, a practice known as multi-
• Heterogeneity in the capabilities of homing. This enables both incoming and
nodes. outgoing load balancing, in addition to the
increased redundancy.
4. Types of load balancing and its 5. Dispatcher-based load balancing is used
approaches within a site to balance load between a set of
real servers. Generally, the dispatcher
Load balancing algorithms can be classified assumes a virtual address for a service and
into three main classes: static algorithms, receives requests which it then redirects to
dynamic algorithms, and adaptive an appropriate server based on given
algorithms. Static algorithms decide how to criteria.
distribute the workload according a prior 6. The real servers can again operate some
knowledge of the problem and the system form of balancing mechanism to decide
characteristics. Dynamic algorithms use whether to handle the request or redirect it to
state information to make decisions during a more suitable server or site.
program execution. Finally, Adaptive 7. Content servers could access back-end
algorithms are a special case of dynamic servers – typically running databases and
algorithms. They dynamically change its low-level services – in a load-balanced
parameters in order to adapt its behavior to fashion.
the load balancing requirements [10]. 8. Back-end servers could also incorporate
balancing amongst themselves to avoid
As though there are three main algorithms over-utilization.
for load balancing, there are different load
balancing approaches available. They are: When talking about varying levels of load
balancing, it is fair to identify the level to be
1. Client-side load balancing: Client-side proportional to the distance from the content
load balancing is not a normal practice, but served – long distance equals high level, and
it is indeed possible. For some times, vice versa. In an informal manner, we can
Netscape incorporated a simple balancing designate steps 1through 4 in the figure as
algorithm in their Navigator browser, high-levels, and steps 5 through 8 as low-
making it choose a random Netscape web- levels of load balancing [12].
server when visiting www.netscape.com.
[2]. 5. Adopted techniques
2. Core Internet routing uses protocols and
agreements that allow for automated load There are various papers which explain
balancing and fail-over mechanisms. These about the load balancing in Internet. Each of

151 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

them is routed in different directions and comparable performance to 16-bit CRC. As


with various considerations. though hashing methods provide best
performance, new hash based algorithms are
Router mechanisms designed to achieve fair needed that have less computational
bandwidth allocations, like Fair Queueing, complexity.
have many desirable properties for
congestion control in the Internet. In [6],
architecture is proposed that significantly In [7], they proposed a per-class queue
reduces this implementation complexity yet management and adaptive packet drop
still achieves approximately fair bandwidth mechanism in the routers for Internet
allocations. This architecture has two key congestion control. An active queue
aspects: management is modeled as an optimization
1. To avoid maintaining per flow state at problem and the proposed mechanism
each router, we use a distributed algorithm provides congestion control and fairness for
in which only edge routers maintain per different types of traffic flows. An optimal
flow state, while core (non-edge) routers do packet drop rate is obtained to maintain a
not maintain per flow state but instead relatively small queue occupancy, which
utilize the per-flow information carried via a provides a less queue delay delivery of
label in each packet’s header. This label packets. Moreover, the queue occupancy
contains an estimate of the flow’s rate; it is and the packet drop rates obtained are both
initialized by the edge router based on per- upper bounded, which is meaningful for
flow information, and then updated at each providing the class-based guaranteed delay
router along the path based only on services for real-time multimedia
aggregate information at that router. applications. They model the general AQM
2. To avoid per flow buffering and as an optimization problem, and try to obtain
scheduling, as required by Fair Queueing, a minimal packet drop rate that results in
we use FIFO queueing with probabilistic low queue occupancy. Compared with RED
dropping on input. The probability of that controls the average queuing delay in
dropping a packet as it arrives to the queue the router, the per-class queue management
is a function of the rate estimate carried in and optimal packet drop mechanism can
the label and of the fair share rate at that obtain the minimal queuing delay and hence
router, which is estimated based on the end-to-end delay. The major drawback
measurements of the aggregate traffic. The in this paper is that the packet drop rate is
limitation here is in the destination, packet minimal and the resulting queue occupancy
reorganizing should be there. is also kept minimal.

In [18], they evaluated 5 direct hashing In [19], a novel packet scheduler is proposed
methods and one table-based hashing and is called Stratified Round Robin, which
method. While hashing schemes for load has low complexity, and is amenable to a
balancing have been proposed in the past, simple hardware implementation. In
this is the first comprehensive study of particular, it provides a single packet delay
performance using real traffic traces. They bound that is independent of the number of
find that hashing using 16-bit CRC over flows. This property is unique to Stratified
TCP five tuples gives excellent load Round Robin among all other schedulers of
balancing performance. Load-adaptive table- comparable complexity.
based hashing uses the exclusive-OR of the
source and destination IP addresses achieve

152 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

An important component of the many QoS two categories: the aggressive and the
architectures proposed is the packet normal, and applies different scheduling
scheduling algorithm used by routers in the policies to the two classes of flows. They
network. The packet scheduler determines have stated that their work is unique in
the order in which packets of various exploiting flow-level Internet traffic
independent flows are forwarded on a shared characteristics.
output link. One of the simplest algorithms
is First Come First Served (FCFS), in which In [13], a new mechanism called Flow-let
the order of arrival of packets also Aware Routing Engine (FLARE) is
determines the order in which they are proposed in which a new traffic splitting
forwarded over the output link. FCFS algorithm that operates on bursts of packets,
clearly cannot enforce QoS guarantees, as it carefully chosen to avoid reordering. Using
allows rogue flows to capture an arbitrary a combination of analysis and trace-driven
fraction of the output bandwidth. simulations, it is shown that FLARE attains
accuracy and responsiveness comparable to
Stratified Round Robin operates as a two- packet switching without reordering packets.
step scheduler: FLARE is simple and can be implemented
with a few KB of router state. Highly
1. The first step uses the flow class accurate traffic splitting can be implemented
mechanism to assign slots to each with little to no impact on TCP packet
flow fi in proportion to its reordering and with negligible state
approximate weight as defined by overhead. Owlets can be used to make load
the flow class Fk to which it belongs. balancing more responsive, and thus help
2. The second step uses the weight- enable a new generation of real-time
proportional credit mechanism to adaptive traffic engineering.
ensure that each flow fi receives
service in proportion to its actual Static load balancing is presented in [3]. The
weight wi. static problem is possible to be formulated
and solved only if we have precise
The advantage of this approach is that it information on all the traffic demands.
considerably simplifies the scheduling However, such information is not available
decision to be made. But we are having the or traffic condition change unexpectedly. In
difficulty of considerable packet dynamic load balancing, dynamically
reorganizing in the destination. changing network status information are
utilized [5],[8],[14],[4],[15].
In this paper [11], based on measurements of
Internet traffic, they examined the sources of Traffic Engineering (TE) of dynamic
load imbalance in hash-based scheduling methodologies is classified into two basic
schemes. They proved that under certain types: time-dependent and state-dependent.
Zipf-like flow-size distributions, hashing In time-dependent TE, traffic control
alone is not able to balance workload. They algorithms are used to optimize network
introduced a new metric to quantify the resource utilization in response to long time
effects of adaptive load balancing on overall scale traffic variations. In state-dependent
forwarding performance. To achieve both TE, traffic control algorithms adapt to
load balancing and efficient system resource relatively fast network state changes. State-
utilization, they proposed a scheduling dependent load balancing is a key technique
scheme that classifies Internet flows into for improving the performance and

153 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

scalability of the Internet. The fundamental splitter algorithm. As though it is adaptive


problem in dynamic load balancing in algorithm for load balancing, it does not
distributed nodes involves moving load satisfy the needs when we are coming across
between nodes. Each node can transfer load the unresponsive flows.
to at most one neighbor also, any amount of
load can be moved along a communication 6. Conclusion
link between two nodes in one step. The
dynamic load balancing is formed with Static load balancing is suffering from lack
incomplete information. More precisely, it is of all the available information while we
assumed that the traffic demands are perform for load balancing. Similarly
unknown but the link loads are periodically dynamic load balancing needs schemes that
measured using a measurement system as in split traffic across multiple paths at a fine
[9], [16] [17]. granularity. In adaptive load balancing the
performance of the network can be
Floyd et al. [20] study congestion collapse improved by parameter adjustments of the
from undelivered packets. This situation routes, traffic splitting, and scheduling. Thus
arises when bandwidth is continuously in this current fast Internet scenario, a new
consumed by packets at the upstream that adaptive load balancing algorithm is needed.
are dropped at the downstream. Several The adaptive load balancing is also not
ways to detect unresponsive flows are sufficient for some situations like
presented. It is suggested that routers can unresponsive flows of traffics. Hence there
monitor flows to detect whether flow is is a need that for load balancing algorithm
responsive to congestion or not. If a flow is with the additional feature of detecting
not responsive to congestion, it can be unresponsive flows of the incoming traffic
penalized by discarding packets to a higher in Internet.
rate at the router. According to the authors
there are some limitations of these tests to References
identify non-“TCP-friendly flow”. It does
not help to save bandwidth at the upstream if [1]
the flow sees the congestion at the http://www.oreillynet.com/pub/a/oreilly/net
downstream because this solution does not working/news/slb_0301.html, Feb 2010.
propagate the congestion information from
downstream to upstream [21]. [2] Dan Mosedale, William Foss, and Rob
McCool, “Lessons learned administering
In [22], architecture which contains an netscape’s internet site.” IEEE Internet
adaptive packet scheduler with a bursty Computing, 1(2):28– 35, 1997.
traffic splitting algorithm is proposed for
better load balancing. The scheduler has a [3] Grenville Armitage, 2000. MPLS the
classifier which classifies the flows into magic behind the myths. IEEE Commun.
aggressive and normal flows. Aggressive Mag., 38: 124-131.
flows are treated as high priority flows.
Based on the buffer occupancy threshold, a [4] Dinan, E., D. Awduche and B. Jabbari,
trigger handler checks for load un-balance of “Analytical framework for dynamic traffic
the network and automatically triggers the partitioning in MPLS networks”, IEEE
load adapter. The load adapter reroutes the International Conference on
high-priority aggressive flows into the least Communications, (ICC’00), 18-22 June
loaded best path, using the bursty traffic 2000, New Orleans, Volume-3, pp: 1604-

154 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

1608. Parallel Forwarding” Networking, Ieee/Acm


Transactions On, Aug. 2005.
[5] Elwalid, A., C. Jin, S. Low and
I.Widjaja, MATE: MPLS adaptive traffic [12] Sven Ingebrigt Ulland, “High level load
engineering. IEEE Infocom., Twentieth balancing for web services”,University of
Annual Joint Conference of the IEEE Oslo, 20th May 2006.
Computer and Communications Societies,
Volume-3, April 22-26 2001, Anchorage, [13] Srikanth Kandula, Dina Katabi,
Alaska, USA. ISBN 0-7803-7016-3, 1300- Shantanu Sinha, Arthur Berger ”Dynamic
1309. Load Balancing Without Packet Reordering”
Acm Sigcomm Computer Communication
[6] Core-Stateless Fair Queuing: A Scalable Review ,Volume 37 , Issue 2 (April 2007).
architecture to approximate fair bandwidth
allocations in high speed networks, Ion [14] Murugesan, G. and A.M. Natarajan,
Stoica, Scott Shenker, Hui Zhang, 2003. “Adaptive granularity algorithm for
effective distributed load balancing and
[7] Mei-Ling Shyu, Shu-Ching Chen, implementation in multiprotocol label
Hongli Luo, “Per-class queue management switching networks”. IEEE, International
and adaptive packet drop mechanism for Conference on Advanced Computing and
multimedia networking”, Proceedings of the Communication (ADCOM’07) 18-21
2003 International Conference on December 2007, pp: 626-631.
Multimedia and Expo - Volume 3 (ICME
'03) - Volume 03. [15] Dengyin Zhang, Zhiyun Tang and
Ruchuan Wang, “Automatic traffic balance
[8] Song, J., S. Kim, M. Lee, H. Lee and T. algorithm based on traffic engineering”. J.
Suda, “Adaptive load distribution over Network Syst. Manage., 14: 317-325.
multipath in MPLS networks”. IEEE
International Conference on [16] Ashwin Sridharan, Roch Guerin and
Communications (ICC’03), Anchorage, Christophe Diot, “Achieving near-optimal
Alaska, Date: 11-15 May 2003 pp: 233-237. traffic engineering solution for current
OSPF/IS-IS networks”. In the IEEE/ACM
[9] Butenweg, S., “Two distributed reactive Trans. Network., 13: 234-247.
MPLS traffic engineering mechanism for
throughput opt’imization in Best effort [17] G. Murugesan, A.M. Natarajan and C.
MPLS networks, In the Eighth IEEE Venkatesh, “Enhanced Variable Splitting
International Symposium on Computers and Ratio Algorithm for Effective Load
Communications (ISCC’03). 379-384 vol.1, Balancing in MPLS Networks” Journal of
30th June-3rd July 2003. Computer Science 4 (3): 232-238, 2008
ISSN 1549-3636 2008 Science Publications.
[10] Mohammed Aldasht, Julio Ortega,
Carlos G. Puntonet; Antonio F. Diaz, “A [18] Zhiruo Cao, Zheng Wang, Ellen
Genetic Exploration of Dynamic Load Zegura, “Hashing-based traffic splitting
Balancing Algorithms”, IEEE 2004. algorithms for Internet load balancing”,
Georgia Institute of Technology 1999.
[11] Shi, W. Macgregor, M.H.
Gburzynski, P. “Load Balancing For [19] S. Ramabhadran and J. Pasquale,
"Stratified Round Robin: A Low

155 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 3, June 2010

Complexity Packet Scheduler with


Bandwidth Fairness and Bounded Delay,"
Proc. Acm Communications Architectures
and Protocols Conf. (Sigcomm), Karlsruhe,
Germany, Pp. 239-249, Aug. 2003.

[20] S. Floyd and K. Fall. Promoting the use


of end-to-end congestion control in the
Internet. IEEE/ACM Transactions on
Networking, Aug. 1999.

[21] Ahsan Habib, Bharat Bhargava,


“Network Tomography-based Unresponsive
Flow Detection and Control”, Department of
Computer science, Purdue University, IN
47907-1398.

[22] M.Azath, Dr.R.S.D.Wahida banu,


“Load balancing in Internet Using Adaptive
Packet Scheduling and Bursty Traffic
Splitting”, International Journal of
Computer Science and Network Security,
Vol.8, No.10, Oct 2008.

156 http://sites.google.com/site/ijcsis/
ISSN 1947-5500