REVISING RESOURCE MANAGEMENT AND SCHEDULING

SYSTEMS

Mehdi Sheikhalishahi and Lucio Grandinetti

Department of Electronics, Computer and System Sciences, University of Calabria, Rende (CS), Italy

Keywords:

Green Computing, Resource Management, Performance, Computing Paradigms, Scheduling, Technologies,

Resource Contention.

Abstract:

With the explosive growth of Internet-enabled cloud computing and HPC centers of all types, IT’s energy

consumption and sustainability impact are expected to continue climbing well into the future. Green IT recog-

nizes this problem and efforts are under way in both industry and academia to address it. In this paper, we take

into account green and performance aspects of resource management. Components of resource management

system are explored in detail to seek new developments by exploiting contemporary emerging technologies,

computing paradigms, energy efﬁcient operations, etc. to deﬁne, design and develop new metrics, techniques,

mechanisms, models, policies, and algorithms. In addition, modeling relationships within and between various

layers are considered to present some novel approaches.

1 INTRODUCTION

Climate Change and Global Warming are the two

most important challenging problems for the Earth.

These problems pertain to a general increase in world

temperatures caused by increased amounts of car-

bon dioxide around the Earth. Researchers in vari-

ous ﬁelds of science and technology in recent years

started to carry out research in order to address these

problems by developing environmentally friendly so-

lutions. Green IT and in particular Green Computing

are two new terms introduced in ICT community to

address the aformentioned problems in ICT.

At the ﬁrst sight of exploring green technolo-

gies and capabilities, we identify power management

operations such as DVFS and IDLE-states capabili-

ties at the processor level and virtualization technol-

ogy at the middleware layer of computing systems

as immediate solutions to address energy efﬁciency

goals i.e. energy consumption minimization and heat-

dissipation (limitations in wasted energy-removal).

On the other hand, resource management as the

main middleware management software system plays

a central role in addressing computing system prob-

lems and issues. Energy efﬁcient operations, new

technologies for green computing, and emerging

computing paradigms should be exploited at the re-

source management to coordinate multiple elements

within a system for manifold objectives. We be-

lieve green and performance objectives converge to

the same point. In the evolution of resource manage-

ment systems, we should take into account this note

as we demonstrate this insight in this paper. In ad-

dition, resource management design and architecture

should evolve according to the advances in contempo-

rary technologies, computing paradigms, and energy

efﬁcient operations to provide new techniques, algo-

rithms, etc. For instance, a comparison between cloud

and other paradigms provides some guidelines and

insights in the design and development of resource

management components. Economic model and ac-

curacy of allocations’ parameters (requests’) are the

main two different characteristics of HPC and Cloud

paradigms which highly impact scheduling.

In Cloud computing, pay-as-you-go on a utility com-

puting basis is the economic model so users pay based

on how much time they used cloud resources, while

in HPC paradigm there is no general or speciﬁc eco-

nomic model. Similarly, the requested allocation time

and capacities of resources as allocation parameters

are not accurate in HPC whereas they are precise and

accurate in Cloud computing model. In fact, this is

the result of economic model.

In HPC environment, a scheduler makes decisions

and reservations according to the requested runtime

of jobs which is an estimate of the real runtime; this

means jobs might ﬁnish earlier or later than the spec-

iﬁed requested runtime, however in Cloud the re-

121

Sheikhalishahi M. and Grandinetti L..

REVISING RESOURCE MANAGEMENT AND SCHEDULING SYSTEMS.

DOI: 10.5220/0003955401210126

In Proceedings of the 2nd International Conference on Cloud Computing and Services Science (CLOSER-2012), pages 121-126

ISBN: 978-989-8565-05-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

quested runtime is the actual runtime.

Therefore, in cloud model users provide which re-

sources they need, the capacity of those resources and

precisely for how much time. We might take into ac-

count these differences in these two paradigms to de-

sign and develop resource management components.

There are many approaches, mechanisms and al-

gorithms in the literature on these problems and is-

sues; however, most of them are special purpose.

A complete approach should be a multi-level and

general-purpose (holistic) approach that is architected

over all layers of computing paradigms and systems.

For instance, in cloud paradigm from the highest

level of resource management stack i.e. Cloud pric-

ing strategies and Admission control policies to the

lower levels i.e. policies (to direct a scheduler in mak-

ing various decisions e.g. host selection for a spe-

ciﬁc job), and ﬁnally core scheduling algorithms are

some of the research work which could be carried out

in holistic scheduling approach. Such an approach

should model the relationship between these layers,

for example core scheduling information about jobs

and resources might be considered in designing Cloud

pricing strategies, Admission control policies and so

on.

In brief, our paper lays some groundwork for fu-

ture researchers in the ﬁeld of resource management

and scheduling to easily deﬁne, design and develop

new objectives and it makes the following contribu-

tions in the ﬁeld:

• Evolution of resource management and schedul-

ing as new technologies, paradigms, etc. emerge

• Finding and establishing relationships within and

between various layers of resource management

• Being general rather than special purpose solution

for all computing paradigms i.e. Cluster, Grid and

Cloud

The next parts of this paper are structured as follows.

In Section 2, we describe computing system problems

from the resource management point of view. Sec-

tion 3 to Section 7 discuss and to some extent ex-

plore resource management components. Then, Sec-

tion 8 points out to some key notes. Finally, Section 9

presents our conclusions and future work.

2 RESOURCE MANAGEMENT

In computing systems from resource management and

scheduling point of view in general, there are the main

problems and issues such as Low Utilization, Over-

load, Poor Performance, Resource Contention. In ad-

dition, if we consider Energy Efﬁcient computing the

High Energy Consumption is another issue. More-

over, sustainability and reliability are other issues to

be addressed by resource management to some extent.

Figure 1 depicts the whole picture of our research

including resource management components, con-

temporary technologies, computing paradigms, Green

IT and their relationships. In this paper, in order

to address these problems components of resource

management system are explored in detail to seek

new developments and evolutions and in parallel this

process exploits contemporary emerging technolo-

gies, computing paradigms, energy efﬁcient opera-

tions, etc. to deﬁne, design and develop new met-

rics, techniques, mechanisms, models, policies, and

algorithms. Furthermore, ﬁnding and establishing re-

lationships within and between various components is

a key consideration in this development process. For

example, a metric could be modeled as an approxi-

mate function of some other (well deﬁned) metrics.

A model could take advantages of several techniques,

etc. A consolidation policy might exploit resource

contention and utilization metrics in order to address

Resource Contention and Utilization issues i.e. to

achieve distribution and packing at the same time, re-

spectively. Thus, in this case consolidation policy is

modeled as a function of utilization and resource con-

tention metrics.

We can enumerate many other relationships in

resource management. However, in some relation-

ships there are tradeoffs, we should model these trade-

offs as well. For example, if we improve utilization,

we might face some performance issues such as Re-

source Contention and Overload. On the other hand,

if we improve resource contention, utilization might

degrade.

The solution to address the aforementioned prob-

lems and issues is all about to answer complex re-

source management questions which start by When,

Which and Where with the help of well established

relationships. For instance, Which types of appli-

cations might be consolidated together in a server?,

When some workloads should be migrated to other

servers?, Where a workload should be placed?, and

lots of other general and speciﬁc questions.

2.1 Metrics

The ﬁrst step in evolving resource management sys-

tem is to deﬁne, model and develop new metrics to

address some problems in better ways than already

developed well known metrics. For example, in green

computing Energy Consumption is a good metric to

address energy related issues.

Utilization, Wait time, Slowdown, QoS, SLA

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

122

Figure 1: Resource management evolution.

are some traditional performance oriented metrics in

HPC, Grid and Cloud paradigms. Resource Con-

tention is another quasi-performance oriented metric

(Sheikhalishahi et al., 2011b) and Revenue is an eco-

nomic metric introduced by Cloud paradigm. In ad-

dition, more attention is paid to Energy Consumption

metric since the appearance of the GreenIT term.

There are many metrics to be considered in re-

source management. We seek to establish some re-

lationships between some of these metrics in order

to have a better model, understanding of system be-

haviour, manifold objectives, etc. For instance, in

(Srikantaiah et al., 2008) it is demonstrated that Uti-

lization, Poor Performance, and Resource Contention

as the main performance metrics and High Energy

Consumption as the main energy efﬁcient metric are

directly inter-related to each other.

Resource Contention is widely recognized as hav-

ing a major impact on the performance of computing

systems, distributed algorithms, etc. Nonetheless, the

performance metrics that are commonly used to eval-

uate scheduling algorithms do not take into account

resource contention since researchers are interested in

improving the conventional well known performance

metrics such as utilization, slowdown and wait time.

On the other hand, in simple terms addressing Re-

source Contention issue will implicitly address Poor

Performance, and High Energy Consumption issues

as well as slowdown and wait time metrics; therefore

in some environments Energy Consumption optimiza-

tion and performance issues could be modeled as an

approximate function of Resource Contention resolu-

tion. In fact, the following approximate function rep-

resents some portion of energy consumption in terms

of Resource Contention and Poor Performance:

EnergyConsumption ' f (ResCont, PoorPer f ) (1)

and the following approximate equation represents

poor performance:

PoorPer f ' g(ResCont) (2)

so that, we simply model energy consumption as a

relative approximate function of resource contention:

EnergyConsumption ' f (ResCont) (3)

REVISINGRESOURCEMANAGEMENTANDSCHEDULINGSYSTEMS

123

Therefore, we optimize resource contention metric

(Sheikhalishahi et al., 2011b) to achieve energy

optimization.

2.2 Techniques

Emerging technologies, paradigms, and energy aware

actions provide various techniques to be exploited

in resource management. Virtualization technology

provides some heavyweight operations such as sus-

pend/resume/migrate and start/stop on virtual ma-

chines, these operations are used in many recent re-

search work to improve various metrics such as uti-

lization.

For example, in (Sotomayor et al., 2008), authors

demonstrated when using workloads that combine

best-effort and advance reservation requests, a VM-

based approach with suspend/resume can overcome

the utilization problems typically associated with the

use of advance reservations, even in the presence of

the runtime overhead resulting from using VMs.

A DVFS-enabled processor provides some en-

ergy aware operations i.e. decrease/increase fre-

quency/voltage, transitioning to an idle-state and tran-

sitioning to a performance-state. Time scaling is a

technique as a result of DVFS operations which might

be exploited in resource management to improve en-

ergy consumption and utilization metrics. For in-

stance, by combining energy aware operations we can

have some optimization in resource and energy usage

that is to ﬁll out free spaces in availability window

of a scheduler by extension of or reduction of jobs’

running times with the help of increasing/decreasing

frequency of processors.

On the other hand, many devices provide the capa-

bility to transition to one or more lower-power modes

when idle. If the transition latencies into and out of

lower power modes are negligible, energy consump-

tion can be lowered simply by exploiting these states.

Transition latencies are rarely negligible and thus the

use of low-power modes impedes performance. To

minimize this impact, it is necessary to alter the work-

load so that many small idle periods can be aggregated

into fewer large ones. This is a workload batching

technique to be exploited in such cases.

In addition, techniques for dynamically balanc-

ing MPC (Memory accesses per cycle), IPC (Instruc-

tions per cycle), utilization and also dynamically scal-

ing the frequency of processors with the help of on-

line learning algorithms (Dhiman and Rosing, 2009)

or other mechanisms are among the other techniques

within this domain.

In depth study and research on scheduling strate-

gies (Shmueli and Feitelson, 2005) in particular back-

ﬁlling mechanisms, revealed that inaccurate esti-

mates generally lead to better performance (for pure

scheduling metrics) than accurate ones. This obser-

vation proposes the development of new scheduling

techniques in HPC and Cloud paradigm according to

the differences between these paradigms.

2.3 Models

A model quantiﬁes some parameters in terms of some

other parameters such as performance, energy, power

and cost. For instance, in green computing a formal

cost model quantiﬁes the cost of a system state in

terms of power and performance. Sleep states’ power

rate and their latency i.e. the time required to tran-

sition to and from the performance/power state, are

examples of parameters in modeling cost vs. perfor-

mance. In addition, models should specify how much

energy will be saved in state transitions and how long

it takes for state transitions.

Cost/Performance, Performance/Energy, Cost/

Energy, Cost/Power, Suspension/Resumption, Migra-

tion, Turn on/off, Energy Consumption, and Power

models are examples of some emerging models.

Some models emerge as a result of some techniques

such as Suspension/Resumption.

The models we seek to design and parameterize in

green computing should relate to power consumption

and computation rate (performance) or energy con-

sumption and completion time simultaneously. These

models are exploited by the scheduling algorithms to

select the best state of a processor, memory, disk and

network.

Power management actions may affect perfor-

mance in complex ways because the overall computa-

tion rate is a net result of the speed and coordination

of multiple elements within the system. For example,

doubling the CPU speed may do little to increase the

computation rate if the memory transactions do not

move any faster. This indicates that models for the

study of energy-computation tradeoffs would need to

address more than just the CPU.

Models are also architecture and infrastructure de-

pendent e.g. internal of Multicore and NUMA sys-

tems have different features and characteristics to be

considered.

In addition, exploiting technology requires mod-

els. For instance, we can use the suspend/resume/

migrate capability of virtual machines to support ad-

vance reservation of resources efﬁciently (Sotomayor

et al., 2008), by using suspension/resumption as a pre-

emption mechanism, In (Sotomayor et al., 2009) au-

thors presented a model for predicting various run-

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

124

time overheads involved in using virtual machines in

order to support advance reservation requests. It ad-

equately models the time and resources consumed by

these operations to ensure that preemptions are com-

pleted before the start of a reservation.

2.4 Policies

We categorize policies into frontend and backend

policies. Admission control and pricing are frontend

policies whereas consolidation, host selection, map-

ping and preemption belong to backend policies.

Job requests pass through frontend policies be-

fore queueing. At the highest level in the cloud in-

terface, we have pricing strategies such as Spot Pric-

ing in Amazon (Amazon, 2010) and recent pricing ap-

proaches in Haizea (Sotomayor, 2010) or game theory

mechanisms. These mechanisms apply cloud poli-

cies that are revenue maximization or improving uti-

lization. Almost these policies have the same goals,

and they are energy efﬁcient since they keep cloud re-

sources busy by offering various prices to attract more

cloud consumers. A dynamic pricing strategy like of-

fering cheaper prices for applications that will lead

to less energy consumption (or higher performance)

based on the current cloud status (workloads and re-

sources) compared to the others is an energy efﬁcient

pricing schema. Pricing strategies implement cloud

administrators’ objective i.e. revenue maximization,

utilization improvement, etc.

Backend policies could be categorized into three

types: general-purpose policies (Dhiman et al., 2009),

architecture-speciﬁc (or

infrastructure-speciﬁc) policies (Hong et al., 1999)

application-speciﬁc (or workload-speciﬁc) policies

(Kim et al., 2007).

General-purpose policies are those that can be ap-

plied to most of the computing systems. For instance,

CPU/cache-intensive workloads should run at high

frequencies, since by increasing frequency the per-

formance scales linearly for a CPU/cache-intensive

workload. However, if a task is memory-intensive, the

performance improvement is relatively insensitive to

increase in frequency, so that a memory-bound work-

load favors by running at a lower frequency to reduce

energy consumption.

Architecture-speciﬁc policies are deﬁned based on

the architecture or the infrastructure in which compu-

tation happens. Also application-speciﬁc policies are

deﬁned around applications’ characteristic.

Technically, workload consolidation (Hermenier

et al., 2009) policy is a sort of policy at the intersec-

tion of the last two mentioned policies. Bundling var-

ious types of workloads on top of a physical machine

is called consolidation. Furthermore, consolidation-

based policies should be designed in such a way to

be an effective consolidation. In fact, effective con-

solidation (Srikantaiah et al., 2008) is not packing

the maximum workload in the smallest number of

servers, keeping each resource (CPU, disk, network

and memory) on every server at 100% utilization,

such an approach may increase the energy used per

service unit.

Placement of jobs is a critical decision to address

Resource Contention. Consolidation policies are one

of the sources of information for effective placement

of jobs in computing paradigms. In (Sheikhalishahi

et al., 2011b), we developed effective energy aware

consolidation policies. We designed consolidation

policies according to the resource contention model

(metric) and implemented them as host selection poli-

cies of a distributed system scheduler.

2.5 Algorithms

Algorithms implement techniques, models and

policies. For instance, algorithms based on

cost/performance models are part of the scheduling

to model cost vs. performance of system states. In

addition, core scheduling algorithms deal with imple-

menting various backﬁlling mechanisms, etc. to im-

prove utilization and other optimizations at core of a

scheduler.

In (Sheikhalishahi et al., 2011a) we have proposed

a novel autonomic energy efﬁcient resource manage-

ment and scheduling algorithm (architected on differ-

ent levels of resource management stack). The pro-

posed autonomic scheduling approach answers some

When questions to improve energy consumption as

it is modeled by resource contention metric. For

that, it models interaction between queue mechanism

and core scheduler information (about jobs and re-

sources), as a result according to the system state,

jobs are reordered and those jobs which satisfy nec-

essary energy aware conditions get admitted to the

wait queue and the others are delayed for the next

scheduling cycles. From an autonomic scheduling

point of view, there are some loops between queue

mechanism, scheduling function, end of a job event

and core scheduler information. This is an example of

an multi-level algorithm over various resource man-

agement components.

3 CONCLUSIONS AND FUTURE

WORKS

In this paper, we have taken into account green and

REVISINGRESOURCEMANAGEMENTANDSCHEDULINGSYSTEMS

125

performance aspects of resource management sys-

tems. First, we enumerated the main issues in

computing systems i.e. Low Utilization, Overload,

Poor Performance, Resource Contention, High En-

ergy Consumption. We highlight that these issues

should be addressed by the resource management

of computing systems. Then, resource management

components are explored in detail to seek new devel-

opments by exploiting contemporary emerging tech-

nologies, computing paradigms, energy efﬁcient op-

erations, etc. to deﬁne, design and develop new met-

rics, techniques, mechanisms, models, policies, and

algorithms.

As an example, we reduced some problems to re-

source contention problem. In addition, we modeled

energy consumption optimization and performance

resolution as an approximate function of Resource

Contention resolution.

Modeling relationships within and between vari-

ous layers are considered to present some novel ap-

proaches such as in resource contention metric design

and autonomic energy efﬁcient scheduling algorithm.

In particular, this paper lays some groundwork for

future researchers in the ﬁeld of resource management

and scheduling, as a result, a lot of future work in the

framework of this paper could be conducted.

Of some notes, although power and energy are

often used interchangeably, there are important dis-

tinctions. Transactional applications are better de-

scribed in terms of throughput and average power,

whereas completion time and energy consumed are

more meaningful for nontransactional applications.

Power consumption can often be increased over short

periods to accelerate computation or accumulate more

data and thereby minimize overall energy consump-

tion. On the other hand, in resource management sys-

tems, we are interested in energy consumption met-

ric not power metric or power average metric (as over

time power draw is variable), since this study is over a

time horizon not one second. Power metric is not re-

vealing any information about consumption, since it

is not consumption. In fact, Watt is the rate of energy

consumption within one second. In all, energy calcu-

lations become much more practical with Joules. In

other words, since in a cloud model we pay for what

we consume, so that we pay for Joules.

REFERENCES

Amazon (2010). Amazon ec2 spot instances. http://

aws.amazon.com/ec2/spot-instances/.

Dhiman, G., Marchetti, G., and Rosing, T. (2009). vgreen:

A system for energy efﬁcient computing in virtualized

environments. In the 14th IEEE/ACM International

Symposium on Low Power Electronics and Design.

ISLPED ’09.

Dhiman, G. and Rosing, T. (2009). System-level power

management using online learning. IEEE Transac-

tions on CAD’09.

Hermenier, F. et al. (2009). Entropy: a consolidation man-

ager for clusters. In VEE’09.

Hong, I., Kirovski, D., Qu, G., Potkonjak, M., and Srivas-

tava, M. B. (1999). Power optimization of variable-

voltage core-based systems. IEEE Trans. Computer-

Aided Design, 18(12):1702–1714.

Kim, K. H., Buyya, R., and Kim, J. (2007). Power aware

scheduling of bag-of-tasks applications with deadline

constraints on dvs-enabled clusters. In CCGRID,

pages 541–548.

Sheikhalishahi, M., Grandinetti, L., and Lagan

a, D.

(2011a). Autonomic energy efﬁcient scheduling.

preprint (2011), to Future Generation Computer Sys-

tems.

Sheikhalishahi, M., Llorente, I. M., and Grandinetti, L. (30

August - 2 September, 2011b). Energy aware consol-

idation policies. In International Conference on Par-

allel Computing.

Shmueli, E. and Feitelson, D. G. (2005). Backﬁlling with

lookahead to optimize the packing of parallel jobs. J.

Parallel Distrib. Comput., 65:1090–1107.

Sotomayor, B. (2010). Provisioning Computational Re-

sources Using Virtual Machines and Leases. PhD the-

sis, Department of Computer Science, University of

Chicago.

Sotomayor, B., Keahey, K., and Foster, I. (2008). Com-

bining batch execution and leasing using virtual ma-

chines. In Proceedings of the 17th international sym-

posium on High performance distributed computing,

HPDC ’08.

Sotomayor, B., Montero, R. S., Llorente, I. M., and Foster,

I. (2009). Resource leasing and the art of suspend-

ing virtual machines. In Proceedings of the 2009 11th

IEEE International Conference on High Performance

Computing and Communications.

Srikantaiah, S., Kansal, A., and Zhao, F. (2008). Energy

aware consolidation for cloud computing. In USENIX

HotPower’08: Workshop on Power Aware Computing

and Systems at OSDI.

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

126