2009 International Conference on Electronic Commerce and Business Intelligence

Research of Grid Scheduling Algorithm Based on P2P_Grid Model

Xiu-mei Wen, Wei Zhao, Fan-xing Meng

AbstractTask scheduling is an integrated component of Grid sever that based on Super-Peer model, each participant
computing among the key technology of the Grid. P2P_Grid organization can configure one or more nodes and operate as
model, a new resource manage model of Grid is brought up in Super-Peer. The nodes in each organization will change
this paper. In this model, we make some Super-Peers to join information of monitor and resource find, but for the nodes
Grid, it divides the Grid into corresponding sub-Grid. Each in the different organization will change information by P2P
Super-Peer composes one local Grid system and answers for manner. At the same time P2P_Grid model has many
dealing with events of a local Grid that it belongs to; these characteristics such as independent of concentrate control,
Super-Peers in different local Grid are equal and they change distribute, extension and can adapt resource state change etc.
messages by P2P mode. A modified Grid scheduling algorithm
We will give a new grid task scheduling algorithm in
is drawn out based on P2P_Grid model environment and the
idea of traditional Min_min scheduling algorithm. At last, the
order to improve quality of the system load balancing and
new algorithm is compared with traditional Min_min utilization of the system resource in this paper. The new
algorithm by simulate experimentation, it shows that the new algorithm is better logical and efficiency than conventional
algorithm has a better quality of system load balancing and the algorithms by simulation testing. There are five sections in
utilization of system resource. this paper, the material content are as follows: the first
section is Introduction; the second section is Review and
Keywords-grid, task scheduling, resource scheduling, P2P, contrast of existent grid task scheduling algorithms; the third
super-peer section is New algorithm idea; the forth section is the
Simulation testing model and the testing result; the last
I. INTRODUCTION section is Conclusion and Future work.
Grid is a new network infrastructure that based on II. REVIEW PREVIOUS HEURISTIC SCHEDULING ALGORITHMS
Internet, it is an important science and technology progress In this section, we will review and contrast a set of
after Internet [1]. The Grid technology has become one of heuristic scheduling algorithms in heterogeneous computing
the research hotspot in the computer fields. The idea of Grid system. First, we mention meta-task concept: A meta-task is
comes from Power Grid, its target is that hope people to get defined as a collecting of independent tasks with no data
high-powered computing capacity and share various dependences. A meta-task is mapped onto machine
heterogeneous resources as they using power-net in the daily statically; each machine executes a single task at a time. In
life. But Grid has some characteristics such as distribute general, we take into account a set of meta-tasks in the Grid
widely, numerous and complicated types, huge amount and environment. There are a large number of heuristic
dynamic etc. Because different Grid technology may be used algorithms to be designed to schedule task to machines on
in the different fields, it leads that resource distributing heterogeneous computing system. In this section, we will list
density is not symmetrical in the whole Gird. Both P2P and several basic heuristics scheduling algorithms as follows:
Grid are new-style distribute computing models, their aims OLB: Opportunistic Load Balancing (OLB) assigns each
are similar, so it makes possible that setting up P2P_Grid task, in arbitrary order, to the next available machine[6, 6,7];
model.P2P technology has advantage of load balancing and UDA: User-Directed Assignment (UDA) assigns each
abundant information resources, thus we can use these task, in arbitrary order, to the machine with the best expected
advantages to make up for the fault of scatteration of Grid execution time for the task [6, 6].
resources. Fast Greedy: Fast Greedy assigns each task, in arbitrary
P2P technology is applied to the Grid and we adopt P2P order, to the machine with the minimum completion time for
protocol to deal with Grid Computing, namely join some that task [6].
Super-Peers into Grid and come into being a new Grid Min-min: In Min-min, the minimum completion time for
resource organization model-P2P_Grid model. It makes each each task is computed respect to all machines. The task with
center Super-Peer to form one local Grid System and answer the overall minimum completion time is selected and
for dealing with local task. Various local Grid systems are assigned to the corresponding machine. The newly mapped
connected by P2P technology. In this model, for a group of task is removed, and the process repeats until all tasks are
clients, every Super-Peer is used as a server, but it is equal mapped [6, 6, 9];
between every two Super-Peer. The topology structure Max-min: The Max-min heuristic is very similar to the
carries out the balance between centrality search and Min-min algorithm. The set of minimum completion times is
anonymous efficiency, it also keeps the robustness of load calculated for every task. The task with overall maximum
balancing and distribute search. In the Grid information

978-0-7695-3661-3/09 $25.00 2009 IEEE 41

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 26, 2009 at 23:06 from IEEE Xplore. Restrictions apply.
completion time from the set is selected and assigned to the receives the task, and it submits lookup resource request to
corresponding machine [6, 6, 9]. Resource Manager. Resource Manager searches those valid
Greedy: The Greedy heuristic is literally a combination logical resources that registered by a certain searching
of the Min-min and Max-min heuristics by using the better algorithm, and registers these valid resources attribute
solution [6, 6]. information into the valid Resource Manager (resource)
GA: The Genetic algorithm (GA) is used for searching according to a certain strategy and level order. At the same
large solution space [5]. time, it must select the best resource or resource set assign to
SA: Simulated Annealing (SA) is an iterative technique the submitted task.
that considers only one possible solution for each meta-task Resource Receiver must control the thing of load in the
at a time. SA uses a procedure that probabilistically allows system, it will transfer the task to other idle Super-Peer as
solution to be accepted to attempt to obtain a better search of soon as it discovers that loading of its local system is not
the solution space based on a system temperature [3]. balancing, so the new algorithm can overcome the fault of
Tabu: Tabu search is a solution space search that keeps loading unbalance.
track of the regions of the solution space which have already
been searched so as not to repeat a search near these areas [4,
8].
A*: A* is a tree search beginning at a root node that is
usually a null solution. As the tree grows, intermediate nodes
represent partial solutions and leaf nodes represent final
solutions. Each node has a cost function, and the node with
the minimum cost function is replaced by its children. Any
time a node is added, the tree is pruned by deleting the node
with the largest cost function. This process continues until a
complete mapping (a leaf node) is reached [2].
The testing results show that OLB, UDA, Max_min, SA,
GSA, and Tabu do not produce good schedules in general.
Min_min, GA, and A* are able to deliver good performance.
The difference between the completion times of the
schedules (makespans) generated by these three algorithms is
within 10%. GA is consistently better than Min_min by a
few percents, since it is seeding the population with a
Min_min chromosome. A*, on the other hand, produces Figure 1. Super-Peer local resource management model
better or worse schedules than Min_min and GA in different
situations. Among the three algorithms, Min_min is the
fastest algorithm, GA is much slower, and A* is very slow. IV. SCHEDULING MODEL
For 512 tasks and 16 machines, the running time of Min_min Resource scheduling plays an important role among the
is about 1 second, GA 30 seconds, and A* 1200 seconds. whole task scheduling, a good resource scheduling algorithm
The conventional task scheduling of Grid is divided two can increase load balancing and reduce the implement time
processes actually, resource scheduling and task scheduling. of task scheduling .From the discussing of P2P_Grid inner
As a result, it must be lead to ignore some parameters of management model, we can know that in this model, Super-
existing actually. But the task scheduling algorithm that Peer schedules resources among a local range, so it will
given in this paper, we take into account not only the avoid search resources from whole Grid range and make
attributes of tasks themselves but also the delay that resource resource scheduling time to reduce, as a result, the
scheduling is brought. So new algorithm will embed resource scheduling time of all tasks will be reduced. Secondly,
scheduling into task scheduling, and produce the new grid because P2P technology has advantage of load balancing, it
task scheduling algorithm. will increase resources utilization of Grid system.
III. P2P_GRID MODEL AND ITS INNER MANAGEMENT MODEL A. New algorithm idea
In general, Super-Peer is constituted the region of In our study, we embed the QoS request/supply into the
resources concentrating to reduce communications delay grid scheduling algorithm to make a better match among
when resources are scheduled. Fig.1 shows a Super-Peer different level of QoS request/supply. In general, the
local resource management model. practical request/supply of QoS is related to resource that
In Fig.1, Grid Resource submits resource attribute tasks apply, for example, QoS for a network may mean the
information to request to register resources by Information desirable bandwidth for the application; QoS for CPU may
Receiver. Information Receiver classes receiving resources mean the requested speed, like FLOPS, or the utilization of
and submits them to PC Manager, HPC(High Performance the underlying CPU. In our study, a one dimension QoS is
Computer)Manager or LAN Manager according to different considered. Now, QoS request/supply is considered few in
sort. User submits task to Task Receiver, Task Receiver the most of existed scheduling algorithms .Consequently,
responds the task resource request when Task Receiver while a task with no QoS request can be executed on both

42

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 26, 2009 at 23:06 from IEEE Xplore. Restrictions apply.
high QoS and low QoS resources, a task that requests a high having been scheduled and return the information to
QoS service can only be executed on a resource providing Resource Controller. Resource Controller visits resource
high quality of service. Thus, it is possible for low QoS tasks based on IP and locates resource at once, then it must return
to occupy high QoS resources while high QoS tasks wait as to the new state of resource. Resource Manager can select a
low QoS resources remain idle. As a result, it leads to reduce group of resource assign this task based on a certain
the system resource utilization. compounding and optimizing principle after it analyzes the
We divide the scheduling takes into two set, high QoS received information.
request(Task_H) and low QoS request(Task_L). Because (3) If there are other tasks to be waiting for resources and
new algorithm combines resource scheduling and task all this kinds of resources have been occupied in the task
scheduling to deal with the grid task scheduling together, we queue memorizer that user submitted, but this task must be
must consider a significant parameter, that is resource completed in a certain time, thus Resource Manager can
scheduling time, let ST be the resource scheduling time. First make this task to join waiting queue, or submits it to the
we assume several parameters as follows : Let ET be other Super-Peer, Super-Peer deals with it according to (1) or
expectation execution time of task, let START be start (2) that mentioned above after it receives this task, finally,
execution time of task, so expectation completion time CT is return the result to user.
defined as CT=START+ET, and START=A+ST; Let A be When user submits task to Super-Peer, they may be meet
arrived time of task. We expect that ST should be enough such things that Super-Peer is not satisfied with QoS
short, thus the whole tasks scheduling time would be request/supply enough of user because local system loading
reduced. To get this aim, we introduce P2P technology into is too much. The Super-Peer needs to transfer some tasks to
the new algorithm, each Super-Peer makes up of a local Grid other Super-Peer at once, in this case, we need to search
system. When users submit tasks, they consider not only whether exist Super-Peer of satisfaction demand and transfer
their task QOS request but also the Super-Peer that the tasks to it.
submitted task belongs to. Users always submit the tasks to
the Super-Peer which they belong to. Super-Peer would deal B. A new grid task scheduling algorithm
with task scheduling after task is submitted. It can be divided Every waiting for scheduling task has two attribute
three scenarios when Super-Peer schedules its own task, values: the Super-Peer that task belongs to; the type of
listed as follows: QoS that task belongs to;
(1) If there are not other tasks that being executed or In the new algorithm, we still use the idea of Min_min to
waiting for scheduling in the task queue memorizer of the select waiting for scheduled tasks for the same kind of tasks
Task Manager, when Task Manager receives the first this that belong to the same Super-Peer. As a result, it ensures
kind of task that user submits, it can search satisfaction that expectation execution time of task ET is the shortest.
demand resources in Resource Manager. Then, Resource For the sake of simpleness, we divided all tasks into two
Manager search whether existing valid resource queue in the sets (Task_H and Task_L) according to tasks bandwidth
respond resource group according to receiving resource. If request.
resource queue is empty, search sub-tree of directory tree of Algorithm procedure:
resource group or leaf node, return collected information to (1) divide all submitted tasks into two sections according
Task Manager and Resource Controller. Resource Receiver to tasks QoS request/supply and they belong to Super-Peer,
divides the resources to different group based on the range of place different set Task_H and Task_L;
IP address and compounds them again after it receives (2) for each task Ti in Task_H set, it is submitted the
information. These compounded resources must be saved at Super-Peer that it belongs to, Super-Peer deals with this task
the same time. Resource Controller can search the best new according to the 3 situations that mentioned above;
state of these resources in the database and return it to Task FOR I=1 TO m// m is a number of tasks in Task_H set
Manager after Resource Controller receives the information {
about received resources. Task Manager makes resource IF(task_queue= =null AND resource_queue= =null)
scheduling last disposing based on receiving resource // scheduling this kind of task firstly
information twice. At last, Task Manager selects a group of {
the best quality resources to assign this task based on Super-Peer searches resources that the task request;
compounding and optimizing principle. Return the new state of resources;
(2) If there is other task or this kind of task to being Scheduling this task;
executed in the task queue memorizer that user submitted, }
when task queue is empty in the Task Manager and resource IF(task_queue!=null AND resource_queue=!=null)
queue of resource group is not empty in the Resource //there is other task or this kind of task to being executed
Manager, because Resource Manager has saved a set of {
better QoS resources that has been completed this task when Super-Peer makes this task to join in the waiting queue;
it receives this kind of task firstly, namely, there is resource Assign the resources to this task and waiting for
queue that has been scheduled and is waiting for being scheduling;
assigned, thus, it is not necessary to submit searching Return the new state of resources;
resource request to Resource Manager. We can select those }
not be assigned resource group straight in resource queue of IF(task_queue!=null AND resource_queue= =null)

43

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 26, 2009 at 23:06 from IEEE Xplore. Restrictions apply.
 Super-Peer makes this task to join in the waiting queue or VI. CONCLUSION AND FUTURE WORK
transfer it to other Super-Peer according to the urgent We first introduce several conventional task scheduling
situation of task; algorithms in heterogeneous computing system and compare
} them, analyze the fault of conventional algorithms deeply.
(3) for each task Tj in Task_L set, the Super-Peer still We produce a new algorithm based on it. In our new
works according to the above procedure. algorithm, we embed P2P technology into Grid, compose
V. EXPERIMENT RESULTS local Grid system and take advantage of Super-Peer to
complete local task scheduling. Comparing with basic
We use SimGrid simulator [8] to make simulation testing heuristic Min_min algorithm, there have more improving in
for the new grid task scheduling algorithm by us. We put the quality of loading balancing and system resource
Super-Peer into SimGrid simulator and compose a local Grid utilization.
system. For task QoS request, we consider bandwidth This study is a first attempt to embed QoS and P2P into
request only. QoS requests have a big effect on the the Grid task scheduling. There exist more study space in
performance of task scheduling. Based on actual this field, for example, embedding multi-dimensional QoS
applications, three different scenarios are used in our into task scheduling is still a topic of research; for Super-
simulation testing, we set high QoS request ratio to 20%, Peer, we hope find the best fittest node when the task is
50%, 80% in all tasks. For each scenario and each heuristic transferred other Super-Peer. So, it must produce searching
we create 150 tasks 150 times independently and get the strategy to find the best fittest node of Super-Peer.
average makespan of the 150 times. Fig.2 and Fig.3 show the
comparison between new algorithm and Min_min in the REFERENCES
average makespan and system resource utilization. [1] Du Zhi-Hui, Chen Yu, Liu Peng. Grid Computing[M] BeiJing
As shown in Fig.2, we can get such conclusion as Tsinghua University Publish,20021 .
follows: where the tasks that require high QoS and the tasks [2] K. Chow and B. Liu. On mapping signal processing algorithms to a
that require low QoS are evenly distributed, the performance heterogeneous multiprocessor system. In ICASSP 91, pages 1585
gain reaches as high as 15.74%. While where the tasks that 1588, May 1991.
require high QoS are in lower density (20%) or in higher [3] M. Coli and P. Palazzari. Real time pipelined system design through
density(80%), the performance gain are lower than evenly simulated annealing. Journal of Systems Architecture, 42(6-7):465
475, Dec. 1996.
density (50%). As shown in Fig.3, we find that for the three
[4] I. D. Falco, R. D. Balio, E.Tarantino, and R. Vaccaro. Improving
scenarios, the system resource utilization as much high as search by incorporating evolution principles in parallel tabu search. In
Min_min. So, new grid task scheduling algorithm shows IEEE Conference on Evolutionary Computation, pages 823828,
better quality than Min_min in two aspects. 1994.
[5] L. Wang, H. J. Siegel, V. P. Roychowdhury, and A. A. Maciejewski.
Task matching and scheduling in heterogeneous computing
environments using a genetic-algorithm-based approach. Journal of
Parallel and Distributed Computing, 47(1):115, Nov. 1997.
[6] R. Armstrong, D. Hensgen, and T. Kidd. The relative performance of
various mapping algorithms is independent of sizable variances in
run-time predictions. In 7th IEEE Heterogeneous Computing
Workshop (HCW 98), pages 7987, Mar. 1998.
[7] Sun Xian-He, Wu Ming, GHS: A performance prediction and task
scheduling system for Grid computing. In Proc. of 2003 IEEE
Figure 2. Makespan for three scenarios for two algorithms International Parallel and Distributed Processing Symposium
[8] Henri Casanova, Graziano Obertelli, Francine Berman and Rich
Wolski. The AppLeS parameter sweep template: User-level
middleware for the Grid. In Proc. the Super Computing Conference
(SC'2000), Dallas, Texas, Nov. 2000.
[9] Buyya R, Murshed M, Abramson D. A deadline and budget
constrained cost-time optimization algorithm for scheduling task
farming applications on global Grids, In 2002 International
Conference on Parallel and Distributed Processing Techniques and
Applications (PDPTA'02), Las Vegas,Nevada, USA, June 2002.

Figure 3. system resource utilization for three scenarios for two

44

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on August 26, 2009 at 23:06 from IEEE Xplore. Restrictions apply.