# Introduction

raditionally multiprocessors were used as parallel computing platform to execute large-scale grand challenging applications. But for over the past decade, there have been unprecedented technological advances in the commodity personal computers (PCs) and network performance, mainly as a result of faster hardware and more sophisticated software. Another predominant trend witnessed during this era was the falling prices of these technologies. Both of these trends intuitively stimulated the creation of new cost-effective and high-performance networked-computing based parallel and distributed paradigm centering on the use of cluster of low-cost PCs (and/or workstations) interconnected with low-latency, high-bandwidth networks (like ATM, switched Fast or Gigabit Ethernet etc.). Clusters of PCs are becoming a commonplace high-performance computing platform in universities which enjoy the in-house availability of cluster constituents such as PCs and internetworking devices as commodity-off-the-shelf (COTS) components.

On par with the development of clusters as a parallel processing platform to execute large-scale applications (also known as jobs), scheduling on clusters has been an interesting research area to work with in recent years. Job schedulers are generally designed to resolve the contention among multiple competing jobs to acquire the available computational resources (such as processors, memory, storage etc.).

Parallel job scheduling problem is widely studied in traditional multiprocessor systems [3][4][5][6] and to a relatively less extent in cluster computing systems [1][2]. A common assumption in most of the existing parallel job scheduling research in both these systems, has been that all processors in the system have equal processing capacity (i.e., homogeneous) and dedicated. In contrast, in this paper we focus on proposing a scheduling algorithm to allocate processors to jobs in non-dedicated and heterogeneous cluster computing environment.

The rest of the paper is organized as follows: Section 2 discusses background knowledge on cluster computing systems and scheduling. Section 3 states the problem statement. Section 4 gives an overview of previous literature work related to the problem and describes the details of the proposed solution. Section 5 describes simulation model which discusses the workload and system model used. Section 6 evaluates the performance of new policies and compares them with existing solutions and Section 7 concludes the paper.


# II.


# Background Knowledge

In recent times, paradigm of parallel processing in various organizations has been shifted from traditional expensive multiprocessors to commodity-based highperformance clusters due to their high-performance and cost-effectiveness. Clusters of PCs or workstations based on the duration and availability of amount of processing capacity can be generally classified into two classes; 1) High Performance Computing (HPC) systems, and 2) High Throughput Computing (HTC) systems. HPC systems are suitable for interactive parallel jobs as they deliver enormous amount of processing capacity over short periods of time. HTC systems provide large amounts of processing capacity over long periods of time and hence suitable for batch parallel jobs. Within these two classes, cluster computing systems are further classified into three categories [1][2] as follows: 1. Volunteer-owned Cluster Computing (VCC): Individual computer in the system may be homogeneous or heterogeneous and is assumed to be privately owned. The machines may be used for executing externally submitted jobs only if the owner is not using it. Therefore machines in the VCC systems are non-dedicated as these are not simultaneously available to external and local users. The VCC cluster type systems are commonly used as High throughput computing (HTC) platforms. Space-sharing policies are commonly used to schedule parallel jobs in distributed-memory parallel systems such as multiprocessors as well as cluster computing systems. In space-sharing policy, parallel system of multiple processors is divided into disjoint set of processors (known as partitions) so that each partition can be assigned to a single job. In this way, number of jobs can be executed side-by-side by simultaneously providing processor partitions. The number of processors in each partition to be assigned to a job is known as partition size. The primary reason for preferring space-sharing over time-sharing for cluster systems is to avoid the cost of context switching due to frequent preemptions in time-sharing systems.

Space-sharing policies can be broadly divided into fixed, variable, adaptive and dynamic policies [3][4] based on the decision that whether the partition size once assigned to the jobs can be changed during execution time or not. In fixed policies, partition sizes are fixed by the administrator before the system actually starts operating and any modification to these partition sizes require a system reboot. Variable policies require partition sizes to be specified by the user at the time of job arrival. In adaptive policies, partition sizes are determined by the scheduler at the time of job scheduling on the basis of current system load and any available job characteristics. However partition size once assigned to a job cannot be changed during job execution. In dynamic policies, partition size of a job can be changed during its execution.

High performance applications for cluster computing systems are mostly presented as parallel jobs. A parallel job is said to consist of a set of tasks running concurrently to achieve a certain common objective. Each task runs to completion on its assigned processor. The number of tasks (and hence processors required) a certain job has is referred to as the job size.

Characteristics of on-line job streams that act as input workload to the job schedulers influence the performance of the schedulers. Parallel jobs can be classified into four types [3][4]; (i) Rigid, (ii) Moldable (iii) Evolving, and (iv) Malleable, depending upon the number of processor to be allocated at submission time or during execution. A rigid job demands a fixed number of processors at the time of submission and executes on these processors exclusively until completion. Moldable jobs can be made to execute on different number of processors based on the current system load. For example if system load is high, then few processors can be assigned to the moldable job and if system load then large number of processors can be allotted to the job. However this flexibility is only available at job start time and partition size cannot be reconfigured during execution. The processor requirements of both Evolving and Malleable jobs can be changed during execution. For evolving jobs, requirement changes are initiated by the application itself during the various phases of its execution. If the system cannot satisfy the job's demand, the job has to wait for exact processor allocation. For malleable jobs, the decision to change the number of processors is made by an external job scheduler.


# III.


# Problem statement

Most of existing parallel job scheduling policies especially adaptive space-sharing polices have been focused on homogeneous parallel systems such as distributed-memory multiprocessors and cluster computing systems (for example PCC or VCC systems) in which all the processors are dedicated and of equal capacity. It should be noted that scheduling polices for distributed-memory multiprocessors can be directly used in PCC systems without any modification due to similarity in architectures of both the systems. However clusters of PCs such as CCC systems tend to be heterogeneous due to the fact that over the passage of time, new fast machines are regularly added to cluster or some of the obsolete cluster hardware is replaced to improve cluster computing performance. Moreover in The problem seems significant as the partition sizes obtained in non-dedicated heterogeneous parallel systems (e.g. CCC systems) will be different from those obtained in dedicated homogeneous systems. When we partition a dedicated homogeneous cluster (such as PCC and VCC), partition size is obtained by dividing total number of physical processors by the total number of jobs in the system. But in case of non-dedicated heterogeneous systems, partition size is calculated by dividing the total available computing power of all processors by the number of jobs currently available in the system. However total available computing power will be different at different times due to variations in the computing power of individual processors in the presence of varying background workload. Hence corresponding calculated partition size changes continuously. Moreover existing adaptive policies focus on using certain job characteristics (which may not be readily and cheaply available) to calculate partition size. Therefore an efficient adaptive scheduling policy is required which can take care of heterogeneity of processor speeds as well as run-time load variations due to background workloads executing at individual processors and above all requires no job characteristics to calculate partition size.

IV.


# Related Works and proposed algorithms

The focus of the current job scheduling research in distributed-memory parallel systems is towards adaptive algorithms to schedule moldable jobs [8][9][10][11][12][13][14][15] as they have shown to achieve better mean response time than the scheduling algorithms for rigid jobs. This is due to the fact that adaptive algorithms decide the partition sizes by adapting to current system load at job scheduling time whereas rigid jobs only require a fixed number of processors resulting into increased processor fragmentation and mean response times. Dynamic policies are shown to more suitable for shared-memory parallel systems in which the associated overheads of dynamic-partitioning are outweighed by the benefits.

Adaptive scheduling algorithms for assigning partition sizes to moldable jobs have been extensively studied in homogeneous parallel systems [5][6][7][8][9][10][11][12] and to less extent in heterogeneous parallel systems [2] [13]. Existing adaptive algorithms in both homogeneous and heterogeneous cluster systems share one common assumption that processors are dedicated to execute only cluster applications (no other applications can be executed locally). Available adaptive policies also differ from each other by the amount of job characteristics used in making processor allocation decisions.

In [5][6], Rosti et al. introduced several adaptive partitioning policies (known as Fixed Processors per Job (FPPJ)), Equal Partitioning with a Maximum (EPM), Insurance Policy and Adaptive Policies (known as AP1, AP2, AP3, AP4 and AP5)) for distributed-memory multiprocessors over a wide range of workload types and with different possible arrival rates. These policies try to allocate equal-sized partitions to the waiting applications since no a priori job characteristics were assumed to be available. However these policies differ from each other in how the target partition-size is computed.

Out of these adaptive policies, AP2 (known as work-conserving policy) seems to be an interesting policy that reserves one additional partition for the future job arrivals. The partition size in the AP2 policy is calculated as shown in (1).

(1)

In [7], Dandamudi and Yu show that AP2 considers only queued jobs to calculate partition size. This will lead to a situation that contravenes the principal of allocating equal-sized partitions to all jobs. Dandamudi and Yu, suggested a modified version of AP2 known as Modified adaptive policy (MAP) which considers waiting as well as running jobs to calculate partition size as shown in (2).

(2)

Target partition size to be finally allocated to the waiting job is calculated using equation (3). It is the minimum of the partition size calculated using equation ( 2) and maximum parallelism of the job.

(3)

The parameter f (whose value lies between 0 and 1) is used to control the contribution of the "running" jobs to the partition size. It has been shown that the MAP policy provides significant improvement in performance over policies like AP2, ASP and ASP-max etc. that do not consider the contribution of running jobs while calculating partition size. The amount of improvement obtained is a function of parameter f, system load, and workload.

The adaptive policy proposed in [8][10] is more restrictive, in that users must specify a range of the number of processors for each job. Availability of service demand knowledge of an individual job is assumed in the paper. Schedulers will select a number which gives In [2], a variation of MAP, known as Heterogeneous Adaptive Policy (HAP) was suggested by Dandamudi and Zhou to work with heterogeneous parallel systems. The work introduced the concept of Basic Processor Unit (BPU) to differentiate the heterogeneous processors from each other. Partition sizes are allocated to the jobs on the basis of their computation power in terms of number of BPUs rather than using a physical processor level as in homogeneous systems. The research paper showed the supremacy of HAP over MAP and AP2 policies. Partition size in HAP is calculated as in equation ( 5) and target partition size is calculated using equation (3).

(5)

In [13], Shim suggested various adaptive policies for shared heterogeneous network of workstations (NOW) considering the priority of sequential local jobs as well as the parallel jobs. No indepth details about the working of the algorithms are provided in the paper and no comparisons are made with the existing policies. The shortcoming of this paper is that it considers only waiting jobs to calculate the partition size which usually lead to worse results.

In [14], Doan et al. suggested priority-based adaptive policy for homogeneous PC-based cluster systems for both rigid and moldable jobs. The user can assign priority to both types of jobs. The jobs with higher priority are given preference in execution. Since rigid jobs require the fixed number of processors (e.g. partition size), so partition-function for only moldable jobs is derived from equation (2) as given in [7].

In [15], Abawajy proposed another adaptive policy known as SOUL for heterogeneous multi-cluster systems which calculates partition size on the basis of mean service rate of heterogeneous processors, local load at processors and maximum parallelism information of waiting jobs. It has been shown that SOUL policy tends to produce shorter mean job response times as compared to both AEP and MAP at various workloads. But no comparison between HAP and SOUL policy is available.

From the literature survey, following lessons have been learnt which will help us to design a robust adaptive policy for non-dedicated heterogeneous parallel systems. 1) Adaptive policies which consider both current waiting and running jobs in the parallel system perform better than those policies which consider only current waiting jobs. 2) In heterogeneous systems, BPU mechanism is used frequently to differentiate the computing power of different physical processors. 3) When no job knowledge is available, equal-sized (or equivalent) partitioning mechanism is preferred over weighted square-root fair-share strategy which requires the service demand knowledge of jobs. 4) Significant cost in terms of various overheads is involved in obtaining the a priori knowledge of various job characteristics such as maximum parallelism, average parallelism, service demand knowledge etc.  In [12], Srinivasan et al. argue that an equalsized partition strategy tends to benefit jobs with small computation size (light jobs). On the other hand, allocating processors to jobs proportional to the job computation size tends to benefit heavy jobs significantly. A compromise policy is that each job will have a partition size proportional to the square root of its computation size (Weight) as in ( 4). This equation is used to calculate partition size in an enhanced backfilling scheme proposed in [11]. Using these observations and lessons, we have suggested few modifications to HAP policy which have shown good results over various policies in dedicated heterogeneous systems. The new policy is named as Robust Heterogeneous Adaptive Policy (RHAP) to schedule jobs in non-dedicated heterogeneous cluster environment and requires no job characteristics (as opposed to HAP) to calculate final target partition size for the current waiting jobs.

Since cluster processors can be shared between local and parallel jobs, therefore at any point of time, current available computing power for execution of parallel workload at each processor in the presence of local workload is given as in equation ( 6).

Ideal partition size in RHAP is then calculated on the basis of current available computing as shown in (7). (7) It should be noted that job scheduler is invoked only at arrival and departure time of jobs. Information about local load and computing power of each processor is also collected by the job scheduler at these times. Here "Max" is system-wide Maximum Allocation parameter which imposes an upper limit on the number of BPUs to be allocated to jobs. It is equal to some fixed percentage of the total BPUs available in the system. The number of BPUs finally allocated is calculated as follows in (8). (8) Since no knowledge about maximum parallelism of the jobs is "a priori" available to the scheduler, so there is no distinction between short and long jobs. In the absence of "Max" parameter, shorter jobs can be assigned large number of processors resulting into internal processor fragmentation and increased waiting times for other jobs. Max puts a cap on the smaller jobs that tend to retain larger partition sizes. Moreover it also avoids the situations when a long job is holding up large number of processors.

V.


# Simulation model

We have implemented a discrete event performance of proposed adaptive scheduling algorithms under various workload conditions. Simulation modeling is preferred over the actual experimentation as it gave us the greater flexibility of covering a wide range of application characteristics and controlled parameters like arrival rates, system utilization etc. and allowed us to abstract away trivial details of the environment under study, which otherwise would complicate the performance evaluation procedure.

The developed simulator takes the on-line job stream as input parallel workload, executes parallel workload with the specified adaptive policy and generates the output in the form of mean response time. Response time of a job is defined as the sum of its execution time and waiting time. Waiting time of job is the difference between job arrival time and job scheduling time. Execution time is the actual time spent to execute the job. It should be noted that at the time of job arrival, no job characteristics (such as number of processors required, job service demand etc.) are available to the scheduler.


# a) System Model

We have used an open system model of community-owned cluster of 64 independent commodity single-processor personal computers and each computer is used in a shared mode i.e. it is able to service local sequential tasks as well as the tasks of parallel job submitted by the central job scheduler. The computers differ from each other in terms of heterogeneity in processor speeds i.e. computing power they possess. Computer and processor terms are used interchangeably in context of this paper. We assume that computers in the cluster are connected using 100Mbps Ethernet switch. Relative computing power of different physical processors is represented in terms of Basis Processing Unit (BPU) [2] which can either be derived with the help of SPECfp2000 ratings based on the processor speeds or by executing independent benchmarking programs on heterogeneous processors. We have used two types of processors in the computers of cluster system; First 32 computers contain Type I processors; Next 32 computers contain Type II processors that are twice faster than Type I processors. Hence each processor in Type I has 1 BPU and Type II processor has 2 BPUs.


# b) Parallel Workload Model

Parallel workload model containing online stream of parallel jobs for scheduling contains two components; 1) job arrival process and 2) job service demand. The job arrival process is characterized by job arrival rate (?) and coefficient of variation of inter-arrival times (CVa). High arrival rate represents that inter-arrival time between successive jobs is small. We have modeled the job arrival process using exponential distribution with CVa equal to one.

Mean service demand (D) parameter is the uncorrelated cumulative mean service demand which represents the total time required to execute the job in a dedicated environment, independent of how many processors are used. Service demand of jobs is generated using 2-stage hyper-exponential distribution with coefficient of variation of service demand (CVs) greater than one. Since moldable jobs can be made to run on the varying number of processors, therefore time (tj) taken by the parallel job varies based on the number of processors (pj) assigned to it when the job starts executing. It should be noted that dj= (tj)*(pj) as we have ignored the communication and synchronization overheads, when overall mean service demand of a parallel job (dj) is distributed equally among tasks (which are always equal to "pj" processors assigned to the job) of the job. 


# Performance Evaluation and Results

In this section we will evaluate the performance of proposed algorithms in terms of mean response time and also compare the simulation results with the existing approaches. The default parameters and values used in simulation experiments are for various simulation Table 1 : Default parameters and values used Average load or utilization of the cluster system due to parallel jobs is derived using equation (9).


# ?????????????? ??????????????????????


# =

?????? ?????????????? ???????? * ???????? ?????????????? ???????????? ???????????? ???? ????????????????????

a) Relative performance of the scheduling policies

In this section we compare the performance of the proposed adaptive scheduling policy i.e. RHAP with the HAP and MAP policy. Since no maximum parallelism information is a priori available to the scheduler, therefore target partition size for HAP policy computed in (3) will be equal to the partition size computed in (2). The default value of 'f' in the partitioning-function for RHAP, HAP and MAP policies is set to 0.5 which is suggested as a reasonable value in existing similar research works [2][7].

RHAP policy tends to produce shorter MRT values at system loads of interest (i.e. at medium to high loads) as shown in figure 1. This is due to two reasons; 1) RHAP policy produce smaller partition sizes as compared to both HAP and MAP as it considers the background workload into account. 2) Max parameter also restricts allocation of the larger partition sizes to jobs. On the other hand, both HAP and MAP try to allocate larger partition sizes since they are not aware of any background workload. But in reality the total available computing power of all processors is much less than that of assumed by MAP and HAP. Therefore jobs have to wait for a long time to receive calculated partition sizes. HAP and MAP policies also tend to produce bigger partition sizes at low to medium system Fig. 1 : Performance of scheduling policies Utilization since they impose no upper limit on the number of processors to be allocated to jobs. This will apparently result into allocation of large partition sizes to even smaller jobs. In this section, we study the sensitivity of the three policies to variances in inter-arrival and service times. When the arrival CV is varied, the service CV is held at 4. Similarly arrival CV is fixed at 1 when the performance sensitivity to service time CV is studied. The system utilization for parallel load is fixed at 80%. The impact of service time variance on three policies is more than the impact of arrival time variance. This is due to the fact that all of these policies use FCFS as a job selection policy which is known to be sensitive of variance in service demand, to allocate processors to jobs. FCFS allocation of processors to jobs results in a situation where small jobs could be blocked by an earlier arrived large job. This problem gets more serious as the variance in service demand increases.


# VII.


# Conclusion

Space-sharing algorithms are preferred in distributed-memory cluster systems to avoid the overhead due to frequent preemptions involved in timesharing systems. Adaptive space-sharing algorithms are used in distributed-memory parallel systems such as cluster computing systems and dynamic space-sharing algorithms are more suited to shared-memory multiprocessors. Most of popular adaptive algorithms are only designed for dedicated homogeneous parallel systems such as multiprocessors, PCC or even VCC type of clusters. Existing few adaptive algorithms for heterogeneous parallel systems require the knowledge of job characteristics to schedule jobs. This paper suggests a robust adaptive policy for non-dedicated heterogeneous cluster systems to schedule parallel jobs without requiring any knowledge of job characteristics. Comparative results have shown the dominance of the proposed policy over the existing similar policies at medium to high system loads of interest. 
![the utilization of computing machines, the processors in CCC systems are often shared by local users to execute local jobs (hereafter known as background workload).](image-2.png "")
![performance. Schedulers in[8][10] use a submit-time greedy strategy to schedule moldable jobs.(4) ](image-3.png "")
6![In a cluster system with P processors, BPUk represents the computing power of kth processor and Local_load k denotes the load at individual processor due to the execution of local jobs.](image-4.png "( 6 )")
![], Srinivasan et al. have some improvement to[8][10]: (i) using schedule timescheduler which defers the choice of partition size until the actual job schedule time instead of job submission time and, (ii) using aggressive backfilling instead of conservative backfilling.](image-5.png "")
![c) Background Workload ModelWe assume abstract model for representing load due to background jobs at each processor by hiding the internal details of arrival and execution times of sequential local jobs. Each cluster processor is assumed to service a stream of background jobs that arrive at individual computers independently. Local_load at each processor indicates the load due to the execution of sequential local jobs. As the local load Net environment to evaluate the increases, computing power available to service parallel workload decreases. We model the local load using discrete uniform distribution ranging from 0% to 30% i.e. U[0%, 30%] and this information is only available to job scheduler at job arrival and departure times.VI.](image-6.png "")
2![Fig. 2 : Sensitivity to arrival time CV b) Sensitivity Analysis](image-7.png "Fig. 2 :")
3![Fig. 3 : Sensitivity to service time CV](image-8.png "PARAMETERSFig. 3 :")
			© 2012 Global Journals Inc. (US)
		
		
			

## i. Sensitivity to Arrival Time Variations

The performance sensitivity of the three policies to inter-arrival CV is shown in figure 2. The mean response time increases with increasing inter-arrival CV for the three policies. The RHAP policy maintains its performance superiority over HAP and MAP policy at 80% system utilization.

The increase in arrival time variance means the clustered arrival of jobs into the system. This also led to longer gaps in the job arrivals. The impact of variance in arrival time is more on HAP and MAP policies as shown fragmentation induced by the background workload and the way the partition-size is computed for the jobs. Since the partition sizes are computed on the basis of total number of BPUs (in case of HAP) and total number of processors (in case of MAP), the actual number of available BPUs (in case of HAP) and available processors (in case of MAP) can be lower than the partition-size computed. This is due to the fact that there is possibility of background tasks running on some of processors at the time and both HAP and MAP doe not consider background workload when computing partition size. But RHAP policy tend to produce smaller partition sizes due to consideration of background workload as well as upper limit imposed by Max parameter, therefore the impact of arrival time variance is reduced as compared to other two policies.


## ii. Sensitivity to Service Demand Variations

The figure 3 shows that MRT of the three policies increases with the increase in the variance in the service demand. With the increase in service demand variance, there will few large service demand jobs and large number of small service demand jobs. As the service time CV increases, the service demand of the larger jobs will increase even though their number goes down as a fraction of the total jobs.
			
			

				


* 
	
		
			JHAbawajy
		
		Parallel Job Scheduling Policies on Cluster Computing Systems
				Ottawa, Canada
		
			November, 2003
		
		
			Ottawa-Carleton Institute for Computer Science, Carleton University
		
	
	Ph.D. Thesis




* 
	
		Performance of Adaptive Space-Sharing Policies in Dedicated Heterogeneous Cluster Systems
		
			SPDandamudi
		
		
			ZZhou
		
	
	
		Future Generation Computer Systems
		
			20
			5
			
			2004
		
	




* 
	
		Theory and practice in parallel job scheduling
		
			DGFeitelson
		
		
			LRudolph
		
		
			USchwiegelshohn
		
		
			KCSevcik
		
		
			PWong
		
	
	
		Job Scheduling Strategies for Parallel Processing
		Lecture Notes in Computer Science
		Berlin
		
			Springer-Verlag
			1997
			1291
			
		
	




* 
	
		
			DGFeitelson
		
		
			LRudolph
		
	
	
		Parallel Job Scheduling -A Status Report
		Lecture Notes in Computer Science
		
			Springer
			2005
			3277
		
	




* 
	
		Robust Partitioning Policies for Multiprocessor Systems
		
			ERosti
		
		
			ESmirni
		
		
			LWDowdy
		
		
			GSerazzi
		
		
			BMCarlson
		
	
	
		Performance Evaluation
		
			19
			
			1994
		
	




* 
	
		Processor saving scheduling policies for in figure 2. These two policies suffer from processor Year multiprocessor systems
		
			ERosti
		
		
			ESmirni
		
		
			LWDowdy
		
		
			GSerrazi
		
		
			KCSevcik
		
	
	
		IEEE Transactions on Computers
		
			47
			2
			1998
		
	




* 
	
		Performance of Adaptive Space Sharing Processor Allocation Policies for Distributed-Memory Multicomputers
		
			SPDandamudi
		
		
			HYu
		
	
	
		Journal of Parallel and Distributed Computing
		
			58
			
			1999
		
	




* 
	
		Adaptive Selection of Partition Size for Supercomputer Requests
		
			WCirne
		
		
			FBerman
		
	
	
		Lecture Notes in Computer Science
		
			1911
			
			2000
			Springer
		
	




* 
	
		Using Moldability to Improve the Performance of Supercomputer Jobs
		
			WCirne
		
		
			FBerman
		
	
	
		Journal of Parallel and Distributed Computing
		
			62
			
			2002
		
	




* 
	
		A Comprehensive Model of the Supercomputer Workload
		
			WCirne
		
		
			FBerman
		
	
	
		Proc. of IEEE 4th Annual Workshop on Job Scheduling Strategies for Parallel Processing
				of IEEE 4th Annual Workshop on Job Scheduling Strategies for Parallel essing
		
			2005
		
	




* 
	
		Effective Selection of Partition Sizes for Moldable Scheduling of Parallel Jobs
		
			SSrinivasan
		
		
			VSubramani
		
		
			RKettimuthu
		
		
			PHolenarsipur
		
		
			PSadayappan
		
	
	
		Lecture Notes In Computer Science
		
			2552
			
			2002
			Springer
		
	




* 
	
		A Robust Scheduling Strategy for Moldable Scheduling of Parallel Jobs
		
			SSrinivasan
		
		
			SKrishnamoorthy
		
		
			PSadayappan
		
	
	
		Proc. of 2003 IEEE International Conference On Cluster Computing
				of 2003 IEEE International Conference On Cluster Computing
		
			2003
		
	




* 
	
		Performance evaluation of scheduling schemes for NOW with heterogeneous computing power
		
			Young-ChulShim
		
	
	
		Future Generation Computer Systems
		
			20
			2
			
			2004
		
	




* 
	
		An Adaptive Space-Sharing Scheduling Algorithm for PC-Based Clusters, Modeling, Simulation and Optimization of Complex Processes
		
			VHDoan
		
		
			2008
			
		
	




* 
	
		An Efficient Adaptive Policy for High-Performance Computing
		
			JHAbawajy
		
	
	
		Future Generation Computer Systems
		
			25
			
			2009