FLEXIBLE ADVANCE-RESERVATION (FAR) FOR CLOUDS

∗

Jos´e Luis Lucas

Department of Computing Architecture, Complutense University of Madrid, Madrid, Spain

Carmen Carri´on, Blanca Caminero

Department of Computing Systems, University of Castilla-La Mancha, Albacete, Spain

Keywords:

Quality of service, Reservations in advance, Virtual machine leasing.

Abstract:

Cloud computing has gained popularity in recent times. Nowadays several companies are migrating their

software towards cloud providers. Using virtual machines as a resource provisioning mechanism offers some

beneﬁts, but depending on the applications it could be difﬁcult to preview the exact amount of resources the

company needs to provide QoS to its clients. In this paper, a new way of booking resources is proposed, in

which Cloud users can specify the minimum and maximum number of virtual resources needed, so that coping

with periods of peak load is easier and cheaper. This is supported by means of a general framework which

includes modiﬁcations and/or additions of several components at different levels of the Cloud architecture. A

microbenchmark-based proof-of-concept evaluation is included, which provides a hint on the beneﬁts of the

proposal.

1 INTRODUCTION

Cloud computing is an emerging topic in the ﬁeld

of parallel and distributed computing. A cloud is a

combination of physically and virtually connected re-

sources that aim to power the next generation data

centers by architecting them as a network of virtual

services (Buyya et al., 2009).

Virtualization is one of the key technologies be-

hind the Cloud computing paradigm. Virtualization

allows the dynamic instantiation of virtual machines

on physical resources, and to allocate them as needed.

There are several beneﬁts expected from virtualiza-

tion, such as high availability, ease of deployment,

migration, maintenance, and low power consumption

that help to establish a robust infrastructure for Cloud

computing (Barham et al., 2003).

With Cloud computing, companies can lease re-

sources on-demand from a virtually unlimited pool.

Those resources can be divided into three parts: in-

frastructure, platform, and software, which are usu-

ally made available as subscription-based services in

∗

This work was supported by the Spanish MEC and MICINN,

as well as European Commission FEDER funds, under Grants

CSD2006-00046 and TIN2009-14475-C04. It was also partly sup-

ported by JCCM under Grants PBI08-0055-2800 and PII1C09-

0101-9476.

a pay-as-you-go model to consumers. These services

are respectively referred to as Infrastructure as a Ser-

vice (IaaS), Platform as a Service (PaaS) and Soft-

ware as a Service (SaaS) in industry. The pay-as-you-

go billing model applies charges for the actually used

resources per unit time. This way, a business can op-

timize its IT investment and improve availability and

scalability.

Typical computing resources offer a static and ﬁ-

nite set of computationalcapacity to users. A resource

is initially purchased with an estimate for its peak ca-

pacity with the hope that the average load on the re-

source stays well below that estimate. However, the

demand is often dynamic, and it is then difﬁcult to

estimate the times when the demand will exceed the

capacity of the resource (e.g, social networking en-

vironment). Most often it is only realized when the

system crashes under the load and this results in user

frustration (Marshall et al., 2010). So a challenge for

Cloud providers is how to provide a solution to cope

with clients failed estimations.

Several research groups have explored the way

of taking advantage of the elasticity that provides

Cloud computing. For example, in (Dias de Assun-

cao et al., 2009) and (Marshall et al., 2010) the bene-

ﬁts of expanding a cluster with EC2 nodes are shown;

in (Llorente et al., 2009) the use of Cloud computing

610

Luis Lucas J., Carrión C. and Caminero B..

FLEXIBLE ADVANCE-RESERVATION (FAR) FOR CLOUDS.

DOI: 10.5220/0003383306100615

In Proceedings of the 1st International Conference on Cloud Computing and Services Science (CLOSER-2011), pages 610-615

ISBN: 978-989-8425-52-2

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

for Grid resource provisioning is proposed. More-

over, there are some companies which offer elastic

cloud services for an amount of money, like Elasti-

cHosts (Web, 2010b).

The target application for the proposal presented

in this paper would be one that exhibits variable loads.

The user (i.e., the application provider) can iden-

tify the range of Virtual Machines (VMs) needed to

support the minimum and maximum expected load.

Then, the system will autonomously adjust the re-

sources allocated to that user depending on the real

load, up to the speciﬁed maximum number of VMs,

and allocating always the minimum number of VMs.

More precisely, an extension to the advance-

reservation lease model implemented in the Haizea

scheduler (Sotomayor et al., 2008) named FAR (Flex-

ible Advance Reservations), is presented, aimed at

taking advantage of the elasticity provided by Cloud

computing. This extension will allow users to manage

their booked ﬂexible resources for improving their

business investment. Moreover, this new feature will

probably generate new proposals of scheduling poli-

cies (due to the new type of lease), smart pricing al-

gorithms, lease priority algorithms and so on.

The rationale behind this is similar to the one used

to accept a number of Variable Bite Rate (VBR) con-

nections into communication links. VBR connec-

tions are characterized by a mean and a peak band-

width (this is similar to the range of VMs pointed out

above). If the peak connection bandwidth is reserved,

link bandwidth will be wasted because connections

do not transmit at their peak data rate all the time.

Translated to our scenario, if the maximum number

of VMs is provisioned, it is likely that great part of

the time many of them will not be used, thus wast-

ing provider’s resources (i.e., energy, CPU power),

and maybe preventing other users to be admitted into

the system. Thus, several connections can be admit-

ted provided that it is quite unlikely that their “peaks”

will happen simultaneously (due to the Law of Large

Numbers). In the Cloud scenario, this means that it

will be unlikely that all users will need the full range

of VMs at the same time. Of course, in both scenar-

ios, some criteria must be applied in order to know

if a new connection (user) can be admitted or not,

while satisfying its requirements, and not compromis-

ing those from already admitted connections (users).

The cost of this type of reservation is intended to

be lower that the cost of reserving the maximum num-

ber of VMs in the conventional way. So the Cloud

provider would get less beneﬁts from every single

user, but as a counterpart, it could admit more users

into the system. From the users perspective, they

could opt for paying a more economic bill for a set

Figure 1: Cloud computing architecture.

of virtual resources, that will very likely be available

when needed.

The remainder of the paper is organized as fol-

lows. In Section 2 the cloud framework used for this

work is outlined. In Section 3 the proposal of ﬂexi-

ble advance-reservation leases is introduced and dis-

cussed. In Section 4 some preliminary experiments

that illustrate the usefulness of the proposal are pre-

sented. Finally, the paper ends with some conclu-

sions in Section 5, where possible guidelines for fu-

ture work are pointed out.

2 CLOUD ENVIRONMENT

Figure 1 depicts the typical Cloud Computing archi-

tecture. It includes three layers: Infrastructure Layer,

Platform Layer and Software Layer. The Infrastruc-

ture Layer can be further divided into three parts:

Physical Layer, where isolated physical resources are

placed; Hypervisor Layer, which allows each physi-

cal resource to be virtualized; and Cloud Management

Layer, which provides a general view of the system.

On top of that, the Platform layer is where virtual ma-

chines are placed. And ﬁnally, the Software layer,

which consists on applications running on those vir-

tual machines.

The authors have focused on OpenNebula

(ONE) (Sotomayor et al., 2009a) as cloud manage-

ment software, instead of other well-known frame-

works such as Eucalyptus (Nurmi et al., 2008) or

Nimbus (Keahey and Freeman, 2008), because it is

one of the most extended cloud manager all over the

world, it can be extended and integrated with third-

party developments, and its scheduler can be easily

replaced.

OpenNebula is a virtual infrastructure engine

which provides the functionality needed to deploy,

monitor and control VMs on a pool of distributed

physical resources. It is composed of three main

components, namely, OpenNebula Core, a central-

FLEXIBLE ADVANCE-RESERVATION (FAR) FOR CLOUDS

611

ized component that manages the life-cycle of a VM

(deploy, monitor, migrate, ...); Capacity Manager,

a module that governs the functionality provided by

the OpenNebula core; and Virtualizer Access Drivers,

that exposes the basic functionality of the hypervisor

to provide an abstraction of the underlying virtualiza-

tion layer. OpenNebula is able to dynamically scale-

out this infrastructure by interfacing with an external

cloud. In fact, OpenNebula already includes an Ama-

zon EC2 (Web, 2010a) virtualizer driver.

Besides, Haizea has been used as scheduling soft-

ware, instead of the OpenNebula default scheduler,

because it provides a more complete scheduling en-

vironment. Haizea is an open source lease manage-

ment architecture developed by Sotomayor et al. (So-

tomayor et al., 2009b). Haizea uses leases as a funda-

mental resource provisioning abstraction.

A lease is a “negotiated and renegotiable agree-

ment between a resource provider and a resource con-

sumer, where the former agrees to make a set of

resources available to the latter, based on a set of

lease terms presented by the resource consumer” (So-

tomayor et al., 2008).

Haizea supports three types of leases (Sotomayor

et al., 2009b), being the ﬁrst one the focus of interest

of this paper:

1. Advance reservation lease, where the resource

must be available at a speciﬁc time;

2. Best-effort leases, where resources are provi-

sioned as soon as possible;

3. Immediate leases, where resources are provi-

sioned when requested or not at all.

3 FLEXIBLE ADVANCE

RESERVATION (FAR) LEASES

The main contribution of this paper consists of ex-

tending the Cloud manager infrastructure for optimiz-

ing user cost-beneﬁt ratio and improving the use of

resources for variable load applications. Currently, an

application provider can book a number of virtual ma-

chines during a period of time. However, during the

execution, the application which runs on the allocated

VMs could get overloaded. In order to try to over-

come this, a solution consists on adding a range of

extra virtual machines which could be used in case

the application needs them.

A typical use case could be described as follows.

The clients want to get a certain level of Quality of

Service (QoS), i.e, a bounded response delay for their

application. They also want to save some money and

not to waste part of their booked resources. So, with

Figure 2: Steps for supporting ﬂexible advance-reservation

leases.

FAR leases, they could purchase enough VMs for an

average load and also a ﬂexible range for a peak load.

This possibility would be cheaper than purchasing the

full range of VMs for the whole leasing period, even

though they will be only needed for coping with a un-

predictable small peak-load period.

As a counterpart, when using FAR leases, the

Cloud provider could increase the number of Cloud

clients (thus also increasing beneﬁts) since the “ﬂexi-

ble” VMs do not be to be allocated the whole reserva-

tion time to one speciﬁc client. Instead, the scheduler

within the Cloud provider could implement strate-

gies to estimate when and how many more ﬂexible

VMs need to be allocated over time, and thus decide

whether more clients should be admitted or not.

In order to support ﬂexible advanced-reservation

leases, several changes need to be introduced in the

generic Cloud infrastructure, as depicted in Figure 2.

1. A “smart-component” needs to be created. This

component needs to be placed in the booked vir-

tual machine and it is able to connect to OpenNeb-

ula’s node. Its mission is to monitor the QoS of-

fered by the service running in that VM, triggering

the creation/deletion of VMs within the speciﬁed

ﬂexible range as needed.

2. The Haizea Metascheduler has to be extended to

consider the new type of lease. As stated ear-

lier, Haizea has three types of leases: advance-

reservation leases, best-effort leases, and immedi-

ate leases. This proposal is related to the ﬁrst one,

extending its characteristics to include the “ﬂexi-

bility” feature: a range of virtual machines. More-

over, Haizea would need some control over this

ﬂexibility feature in order to know when the sys-

CLOSER 2011 - International Conference on Cloud Computing and Services Science

612

Figure 3: Flexible advance-reservation scenario.

tem is empty of resources to provide, or if there

are still enough resources to fulﬁll a client request.

Here, a range of heuristics can be applied.

3. Support for a new SLA type should be added so

that clients can use the new type of lease. SLA are

negotiated between two parties following the spe-

ciﬁc negotiation protocols, and are generated us-

ing the SLA templates available on both the con-

sumers and the providers sides. Cloud computing

providers would need new ways to generate SLA

templates which include the new ﬂexible advance-

reservation lease.

Figure 3 shows an activity diagram which mod-

els the intended behavior of a FAR lease. First of all,

a client asks OpenNebula for a (number of) virtual

machine(s) using an Flexible Advance-Reservation

lease. Both parties negotiate an SLA, and ONE al-

locates the VM to the client. The client sets up its ser-

vice on the newly allocated VM, and monitoring on

the performance of such service starts. After that, and

depending on the action to do, if the service becomes

collapsed, the smart component would ask for a new

virtual machine to help supporting the service. On

the other hand, if the service is working better than it

was previewed, the component would tell OpenNeb-

ula to delete one existing virtual machine. This pro-

cedure works as a loop within the booked period of

time. Moreover, the smart component has the ability

to automate the dynamic provisioning of VMs taking

into account the negotiated SLA.

Of course, this basic scenario could be improved

by implementing some criteria (i.e., prediction tech-

niques) to anticipate the service collapse, and havethe

additional VM(s) ready before collapse really occurs.

Also, the time needed to set up additional VMs should

be considered, when ﬁne-tuning the heuristics inside

the smart component.

4 PRELIMINARY RESULTS

In order to have an insight into the foreseen beneﬁts of

our proposal, a proof-of-concept experiment has been

devised which shows the system behavior when VMs

are automatically provisioned and freed on workload

ﬂuctuations.

The physical infrastructure used in this work con-

sists of one central host in which OpenNebula is in-

stalled, and two more hosts where VMs will be de-

ployed. In both hosts the Xen Hypervisor (Web,

2010c) is used to manage VMs. As Figure 4 depicts,

the virtual machines have a single core 1.8 GHz AMD

Opteron processor and 256 MB of RAM, and all of

them are connected to the Internet.

It is supposed that there is a client who has booked

one virtual machine for a period of time. This VM has

an Nginx web server (ngi, 2010) with PHP installed,

and also some PHP scripts which consume CPU time.

For testing that virtual machine, a variable work-

load has been designed. Taking into account that the

arrival rate is one request per second, three types of

requests have been created: Low load requests, which

FLEXIBLE ADVANCE-RESERVATION (FAR) FOR CLOUDS

613

Figure 4: Infrastructure used in the experiment.

take less than a second to ﬁnish; High load requests,

which take more than a second to ﬁnish; Very high

load requests, which take much more than one sec-

ond to ﬁnish.

The workload submitted to the web server has

been created by mixing the three types of requests

within a period of time. As Figure5 shows, the work-

load starts with a period of low load requests, being

followed by a period of high load requests. This se-

quence of requests will load the web server. After

that, it comes another easy low load period before

introducing the highest load requests into the web

server. The test period ends with another low re-

quest period. The rationale behind this workload vari-

ations is to mimic a workload that could need more

resources than the initially foreseen.

The response time of the jobs is the metric used

in our proof-of-concept evaluation. This parameter

is a measure of the QoS provided by the web server.

More parameters will be considered in a future such

as CPU usage or energy consumption, which could be

also used to set up trade-offs between users’ service

guarantees and provider’s interests.

Next, some plots which illustrate the evolution of

the web server performance without and with FAR

leases are shown. Figure 6a shows the web server per-

formance under the speciﬁed workload without FAR

leases. It can be observed that the web server gets

overloaded during the period of high load requests,

because lots of them are being queued. When in-

troducing low load requests, the web server is most

of the time overloaded due to previously queued re-

quests. Then, while introducing very high load and

then again low load requests, the web server shows

the same performance as before.

When FAR leases are applied, performance degra-

dation is avoided just by booking a range of virtual

machines in the same period of time. As we can see

in Figure 6b when the main virtual machine gets over-

0.5

1.5

0 50 100 150 200 250 300 350 400

Time (seg)

Ellapsed time (seg)

CPU processing time Arrival rate

Figure 5: Workload proﬁle.

loaded, the smart component asks OpenNebula for

another virtual machine for helping it. In this point,

the ﬁrst peak load has been ﬁxed. Later, in the low

part of the workload, the component knows that the

web server is working ﬁne and it decides to delete the

extra VM. When the very high load part of the work-

load appears, the component asks for one extra VM

and it also immediately asks for another one for pre-

venting the web server to get overloaded. When this

part ﬁnishes and the second low part arrives, the web

server works ﬁne again and the component deletes

those extra virtual machines.

5 CONCLUSIONS AND FUTURE

WORK

In this work, a new type of Virtual Machine lease,

named Flexible Advance Reservations (FAR) lease,

is proposed in order to provide Cloud clients with a

cost-effective way to cope with periods of peak load,

while allowing the Cloud provider to make a better

utilization of resources, thus increasing beneﬁts.

The general framework for supporting FAR leases

has also been described, which includes three parts:

1) a smart-component implemented within the provi-

sioned VM, 2) extensionsto the Haizea scheduler,and

3) support from the SLA manager.

The smart-component has already been created.

Its aim is to monitor the QoS offered by the services

provided by the booked VMs, and decide whether

more VMs are needed, or the system works ﬁne.

Finally, a proof-of-conceptevaluation has been in-

cluded, showing the beneﬁts of the automatic and dy-

namic provision of resources on applications demand.

Up to now, the current implementation is an ini-

tial prototype focused speciﬁcally on the Nginx web

server. As future work, more intensive testing, over

larger environments, need to be carried out. Also, it

CLOSER 2011 - International Conference on Cloud Computing and Services Science

614

0 50 100 150 200 250 300 350 400

Response Time (seg)

Time (seg)

(a) Without FAR lease.

0 50 100 150 200 250 300 350 400 450

Response time (seg)

Time (seg)

BookedMV

Extra−VM1 Extra−VM2

(b) Using FAR lease.

Figure 6: System performance.

is intended to develop the full framework described

here. The ﬁrst step will be extending the Haizea

scheduler with the new type of lease. Finally, research

will be carried out on algorithms for improving the

smart component performance, and on strategies for

VM scheduling in a Cloud supporting FAR leases.

REFERENCES

(2010a). Amazon Elastic Compute Cloud (EC2), URL:

http://aws.amazon.com/ec2/.

(2010b). Elastic Hosts - Flexible servers in the cloud, URL:

http://www.elastichosts.com/.

(2010). Nginx web server, URL: http://nginx.org.

(2010c). Xen Hypervisor, URL: http://xen.org/.

Barham, P. et al. (2003). Xen and the Art of Virtualization.

In Proceedings of the 19th ACM Symposium on Oper-

ating Systems Principles.

Buyya, R., Yeo, C. S., Venugopal, S., Broberg, J., and

Brandic, I. (2009). Cloud computing and emerging

IT platforms: Vision, hype, and reality for delivering

computing as the 5th utility. Future Generation Com-

puter Systems.

Dias de Assuncao, M., di Costanzo, A., and Buyya, R.

(2009). Evaluating the cost-beneﬁt of using Cloud

computing to extend the capacity of clusters. In Pro-

ceedings of the 18th ACM International Symposium

on High Performance Distributed Computing.

Keahey, K. and Freeman, T. (2008). Science Clouds: Early

Experiences in Cloud Computing for Scientiﬁc Ap-

plications. In Proceedings of the Workshop on Cloud

Computing and Its Applications.

Llorente, I., Moreno-Vozmediano, R., and Montero, R.

(2009). Cloud Computing for on-Demand Grid Re-

source Provisioning. Advances in Parallel Comput-

ing, Volume 18 (2009): High Speed and Large Scale

Scientiﬁc Computing. IOS Press.

Marshall, P., Keahey, K., and Freeman, T. (2010). Using

Clouds to Elastically Extend Site Resources. Proceed-

ings of the 10th IEEE/ACM International Conference

on Cluster, Cloud, and Grid Computing (CCGrid).

Nurmi, D., Wolski, R., and Grzegoczyk, C. (2008). The Eu-

calyptus Open-source Cloud-computing System. In

Proceedings of the Workshop on Cloud Computing

and Its Applications.

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

ing Batch Execution and Leasing Using Virtual Ma-

chines. In Proc. of the 17th Intl. Symp. on High-

Performance Distributed Computing (HPDC).

Sotomayor, B., Montero, R. S., Llorente, I. M., and Fos-

ter, I. (2009a). Virtual infrastructure management in

private and hybrid clouds. IEEE Internet Computing,

13(5):14–22.

Sotomayor, B., Santiago, R., Martn, I., and Foster, I.

(2009b). Resource Leasing and the Art of Suspending

Virtual Machines. In Proceedings of the 11th IEEE

Intl. Conf. on High Performance Computing and Com-

munications.

FLEXIBLE ADVANCE-RESERVATION (FAR) FOR CLOUDS

615