A study of load distribution algorithms in distributed scheduling

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 02 Special Issue: 02 | Dec-2013, Available @ http://www.ijret.org 37
A STUDY OF LOAD DISTRIBUTION ALGORITHMS IN DISTRIBUTED
SCHEDULING
Shounak Chakraborty1
, Ajoy Kumar Khan2
1
Department of Information Technology, Assam University, India
2
Department of Information Technology, Assam University, India
Abstract
As we know that in distributed systems several autonomous computers are interconnected to provide a single coherent view of a
powerful system and these autonomous computers work independently in a team-like fashion such that the domain of tasks is shared
between all of them. Now to take full advantage of this distributed scenario, we need good resource allocation schemes. Load
distribution algorithm’s work is to deliverately distributed and re-distributes the tasks (loads) among all the participating nodes so
that the overall performance of the entire system is maximized. in this paper we study the details of the load distribution algorithms
and their suitableness in various load scenerios.
Keywords: loads, stability, Load balancing, load sharing, location affinity
----------------------------------------------------------------------***---------------------------------------------------------------------
1. INTRODUCTION
A distributed scheduler is a resource management component of
a distributed operating system that focuses on judiciously and
transparently redistributing the load of the system among the
individual units to enhance overall performance. Users submit
tasks at their host computers for processing. The need for load
distribution arises in such environments because, due to the
random arrival of tasks and their random CPU service time
requirements, there is a good possibility that several computers
are idle or lightly loaded and some others are heavily loaded,
which would degrade the performance. In real life applications
there is always a possibility that one server or system is idle
while a task is being waited upon at another server. [1]
Let us consider a system of N independent servers and let ρ be
the utilization of each server. Then P=1-ρ is the probability of a
server being idle. The probability P is given by the expression
[1] as following:
P= (Eq. 1.1)
Where is the probability that a given set of i server are idle
and is the probability that a given set of (N-i) servers is
not idle.
Even in a homogeneous distributed system, system
performance can potentially be improved by transferring the
load from the heavily loaded systems (sender) to lightly loaded
systems (receiver). The performance is most often quantified by
average response time of tasks (time interval between
submission and completion of task) in case of distributed
systems. Moreover, resource queue length and CPU queue
length are good indicators of load since they co-relate well with
the task response time. [2]
2. COMPONENTS OF A LOAD DISTRIBUTION
ALGORITHM
2.1 Transfer Policy
Transfer policy indicates when a node (system) is in a suitable
state to participate in a task transfer. The most popular proposed
concept for transfer policy is based on a optimum threshold.
[1], [3], [4], [5]. Thresholds are nothing but units of load. When
a load or task originates in a particular node and the number of
load goes beyond the threshold T, the node becomes a sender
(i.e. the node is overloaded and has additional task(s) that
should be transferred to another node ). Similarly, when the
loads at a particular node falls bellow T it becomes a receiver.
2.2 Selection Policy
A selection policy determines which task in the node (selected
by the transfer policy), should be transferred. If the selection
policy fails to find a suitable task in the node, then the transfer
procedure is stopped until the transfer policy indicates that the
site is again a sender. Here there are two approaches viz.: pre-
emptive and non-pre-emptive. Non-pre-emptive the approach is
simple, we select the newly originated task that has caused the
node to be a sender, for migration. But often this is not the best
approach as the overhead incurred in the transfer of task should
be compensated for by the reduction in the response time
realised by the task. Also there are some other factors, firstly
the overhead incurred by the transfer should be minimal (a task
of small size carries less overhead) and secondly, the number of
location dependent system calls made by the selected task

__________________________________________________________________________________________
should be minimal. This phenomenon of location dependency is
called location affinity and must be executed at the node where
the task originated because they use resources such as windows,
or mouse that only exist at the node.
2.3 Location Policy
A location policy decides to which node a task selected for
transfer should be transferred. A widely used method by the
researchers is polling, [1],[3],[4],[5],[6], which involves
querying another node whether it can accept a task to be
transferred to it. The variations are:
 Random: It is a simple dynamic technique where a
task is simply transferred to a random node. Obviously
there is a limit to the number of transfers.
 Threshold: Here polling is done to the node before
task transfer to check its status.
 Shortest: This technique selects the most lightly
loaded node after polling the nodes at random.
2.4 Information Policy
This is responsible for deciding when information about the
other nodes in the system should be collected, where it should
be collected from and what information should be collected.
The variants of this scheme are demand driven, periodic and
state change driven.
2.5 Stability
We first informally describe two views of stability: the queuing
theoretic perspective and the algorithmic perspective.
2.5.1 Queuing Theoretic Perspective
According to the queuing theoretic perspective, when the long-
term arrival rate of work to a system is greater than the rate at
which the system can perform work, the CPU queues grow
without bound. Such a system is termed unstable.[7]
For example we consider a load distribution algorithm
performing excessive message exchanges to collect state
information. The sum of the load due to external work arriving
and the load due to the overhead imposed by the algorithm can
become higher than the service capacity of the system, causing
instability.
We use the effectiveness of an algorithm as the evaluating
criterion. A load distributing algorithm is said to be effective
under a given set of conditions if it improves the performance
relative to that of a system not using load distributing.
2.5.2 The Algorithmic Perspective
If an algorithm can perform fruitless action with finite
probability, the algorithm is said to be unstable. [6] We
consider processor thrashing, the situation where the transfer of
a task to a receiver may increase the receiver’s queue length to
a point of overload, necessitating the transfer of that task to yet
another node. If this process repeats indefinitely then according
to [6] it is unstable.
3. TYPES OF LOAD DISTRIBUTION ALGORITHM
Load distribution algorithms can be categorized according to
the taxonomy introduced by Casavant and Kuhl in [9].
3.1 Classification According to Approach
Load distribution algorithms can be classified as static, dynamic
or adaptive. Static schemes are those when the algorithm uses
some priori information of the system based on which the load
is distributed from one server to another. The disadvantage of
this approach is that it cannot exploit the short term fluctuations
in the system state to improve performance. This is because
static algorithms do not collect the state information of the
system. These algorithms are essentially graph theory driven or
based on some mathematical programming aimed to find a
optimal schedule, which has a minimum cost function. But
unfortunately Gursky has shown that the proble of finding an
optimal schedule for four or more processing elements is NP-
hard.
Dynamic scheduling collect system state information and make
scheduling decisions on these state information. An extra
overhead of collecting and storing system state information is
needed but they yield better performance than static ones.
Dynamic load distribution for homogenous systems was studied
by Livny and Melman in [1], and the scenario of task waiting in
one server and other server being idle was regarded as “wait
and idle” (WI) condition. Significantly for a distributed system
of 20 nodes and having a system load between 0.33 and 0.89,
the probability of WI state is greater than 0.9. Thus, at typical
system loads there is always a potential for improvement in
performance, even when nodes and process arrival rates are
homogeneous. Load sharing facility for large, heterogeneous
system is studied in Utopia [8].
Adaptive load balancing algorithms are a special class of
dynamic load distribution algorithms, in that they adapt their
activities by dynamically changing the parameters of the
algorithm to suit the changing system state.
3.2 Pre-emptive and Non pre-emptive Type
A pre-emptive transfer involves transfer of task which are
partially executed. These transfers are expensive because the
state of the tasks also needs to be transferred to the new
location.
Non pre-emptive transfers involves transfer of tasks which has
not been started. For a system that experiences wide
fluctuations in load and has a high cost for the migration of
partly executed tasks, non pre-emptive transfers are
appropriate.[7]

__________________________________________________________________________________________
3.3 Load Sharing and Load Balancing
Although both type of algorithms strive to reduce the likelihood
of unshared state i.e. wait and idle state, load balancing goes a
step further by attempting to equalize loads at all computers.
Because a load balancing algorithm involves more task
transfers than load sharing algorithms, the higher overhead
incurred by load balancing types may outweigh its potential
improvement.
3.4 Initiation Based
In general the algorithms are also categorized on which node
initiates the load distribution activity. The variations are sender
initiated, receiver initiated or symmetrically initiated (by both
sender and receiver).
A sender initiated algorithm was studied by Eager et. al. in [4]
and a receiver initiated algorithm was studied in [3], where as a
symmetrically initiated algorithm was adopted in [10].
Moreover an adaptive stable symmetrically initiated algorithm
was put forward in [5] and a stable sender initiated algorithm
was discussed in [7].
All the load distribution algorithms are based on one of more of
the types discussed above.
4. COMPARISON AND CONCLUSION
The observations on various kinds of algorithms are as follows:
 Sender initiated algorithms work well in low system
load, but in case of high system load when most of the
nodes are senders they send query to each other
resulting in wastage of CPU cycles and incurring
more delay due to which the system becomes unstable.
 This un-stability happens with receiver initiated
algorithms when the system load is low and most
nodes are receiver.
 For symmetrically initiated algorithms, they cannot
use the previous gathered information and so in
stateless.
 Adaptive algorithms use the previous information to
query a new node and also adjust their threshold
themselves according to the information.
Fig 1: Average response time versus system loads [7]
Based on the performance trends of load sharing algorithms, the
recommendations [11], for selection of a load distribution
algorithm are:
 If the system under consideration never attains high
load, sender-initiated algorithms will give an improved
average response time over no load sharing at all.
 Stable scheduling algorithms are recommended for
systems that can reach high loads. These algorithms
perform better than non-adaptive algorithms for the
following reasons:
a) For overloaded processors are even more
hurled with the extra task of message
handling, in case of sender initiated
algorithms
b) And for the receiver initiated algorithm,
which works well on high loads, but the pre-
emptive transfers are expensive.
 For a system that experiences a wide range of load
fluctuations, the stable symmetrically initiated
scheduling is recommended because it provides
improved performance and stability over entire
spectrum of system loads.
 For a system that experiences wide fluctuations in
loads but has a high cost for the pre-emptive transfers,
a stable sender initiated is recommended.
 For a system that experiences heterogeneous work
arrival, adaptive stable algorithms are preferred.

__________________________________________________________________________________________
REFERENCES
[1]. M. Livny and M. Melman, “Load Balancing in
Homogeneous Broadcast Distributed Systems,” Proc. ACM
Computer Network Performance Symp., pp. 47-55, 1982.
[2]. Zhou S., “An Experimental Assesment of Resource Queue
Lengths as Load Indices” Proc. of USENIX, Washington, pp
73-82, 1987.
[3]. D.L. Eager, E.D. Lazowska. and J. Zahorjan, “A
Comparison of Receiver-Initiated and Sender-Initiated
Adaptive Load Sharing,” Performance Evaluation, Vol. 6, No.
1, pp. 53-68, 1986.
[4]. D.L. Eager, E.D. Lazowska, and J. Zahorian, “Adaptive
Load Sharing in Homogeneous Distributed Systems,” IEEE
Trans. Software Eng., Vol. 12. No. 5, pp. 662-675, 1986
[5]. N.G. Shivaratri and P. Krueger. “Two Adaptive Location
Policies for Global Scheduling,” Proc. 10th Int’l Conf.
Distributed Computing Systems. IEEE COMPUTERS, pp. 502-
509, 1990
[6]. Bryant R.M. and R.A. Finkel. “A Stable Distributed
Scheduling Algorithm.” Proc. 2nd
Int’l Conference On
Distributed Computing, pp 314-323, 1981
[7]. N.G. Shivaratri, P.Krueger and M.Singhal. "Load
Distributing for Locally Distributed Systems". Proc. of IEEE
Computers, 1992.
[8]. Zhou S., Z.Zheng, J.Wang and P.Delisle, “Utopia: A Load
Sharing Facility for Large Heterogeneous Distributed Computer
Systems”. University of Toronto Press, 1992.
[9]. Casavant T.L. and J.G. Kuhl. “A taxonomy of Scheduling
in General-Purpose Distributed Computing Systems”, IEEE
Transactions on Software Engg., pp. 141-154, 1988
[10]. P. Krueger and M. Livny, “The Diverse Objectives of
Distributed Scheduling Policies.” Proc. 7th Int’l Con&
Distributed Computing Systems, IEEE CS, pp. 242-249, 1987.
[11]. N.G. Shivaratri and M. Singhal. “Advanced Concepts In
Operating Systems”. Tata McGraw Hill Edu. Pvt. Ltd., ISBN:
13:978-0- 07-047268-6, 30th
edition, 2012.

A study of load distribution algorithms in distributed scheduling

More Related Content

What's hot (17)

Viewers also liked (20)

Similar to A study of load distribution algorithms in distributed scheduling (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

A study of load distribution algorithms in distributed scheduling