Académique Documents
Professionnel Documents
Culture Documents
on High Performance Computing and Communications The 10th IEEE International Conference on High Performance
Computing and Communications
978-0-7695-3352-0/08 $25.00 © 2008 IEEE 978-0-7695-3352-0/08 $25.00 © 2008 IEEE 978-0-7695-3352-0/08 $25.00 © 2008 IEEE 5 5 5
DOI 10.1109/HPCC.2008.172 DOI 10.1109/HPCC.2008.172 DOI 10.1109/HPCC.2008.172
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
Web
Data Centres
Utility Computing
Service Computing
?}
Grid Computing +
P2P Computing
Cloud Computing
Market-Oriented
•Ubiquitous Computing -Trillion $ business
access
… - Who will own it?
•Reliability
•Scalability
•Autonomic
Paradigms
•Dynamic
discov ery
•Composability
•QoS
•SLA
•…
Attributes/Capabilities
Cloud infrastructure is very robust and will always be A Grid [3] enables the sharing, selection, and
available at any time. aggregation of a wide variety of geographically
distributed resources including sup ercomp uters,
1.1 Definition and Trends storage systems, data sources, and specialized devices
owned by different o rganizations for solving large-
A number of computing researchers and practitioners
scale resource-intensive problems in science,
have attempted to define Clouds in various ways [6].
engineering, and commerce. Inspired by the electrical
Based on our observation of the essence of what
power Grid’s pervasiveness, ease of use, and reliability
Clouds are promising to be, we propose the following
[4], the motivation of Grid computing was initially
definition:
driven by large-scale, resource (computational and
• "A Cloud is a type of parallel and distributed data)-intensive scientific applications that required
system co nsisting of a collection of inter- more resources than a single co mputer (PC,
connected and virtualised computers that are work station, supercomputer, or cluster) could have
dynamically provisioned and presented as one provided in a single administrative domain. Due to its
or more unified computing resources based on potential to make impact o n the 21st century as much
service-level agreements established through as the electric power Grid did on the 20th century, Grid
negotiation between the service provider and computing has been hailed as the next revolution after
consumers.” the Internet and the Web.
Today, the latest parad igm to emerge is that of At a cursory glance, Clouds appear to be a
Cloud computing [5] which promises reliab le services combination of clusters and Grids. However, this is not
delivered thro ugh next-generation data centers that are the case. Clouds are clearly next-generation data
built on compute and storage virtualization centers with nod es “virtualized” through hyperviso r
techno logies. Consumers will be able to access technologies such as VMs, dynamically “provisioned”
app lications and d ata from a “Cloud” anywhere in the on demand as a personalized resource collection to
world on demand. In other words, the Cloud appears to meet a specific service-level agreement, which is
be a single point of access for all the computing need s established through a “negotiation” and accessible as a
of consumers. The consumers are assured that the composable service via “Web 2.0” technologies.
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
grid
cloud
cluster cloud
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
Figure 3: High-level market-oriented cloud architecture.
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
Service Request Examiner and • support customer-driven service management
Admission Control mechanism to based on custo mer profiles and requested
improve resource allocation service requirements,
decisions.
• define computational risk management tactics
o VM Monitor : The VM Monitor to identify, assess, and manage risks involved
mechanism keeps track of the in the execution of applications with regards
availability of VMs and their to service requirements and customer needs,
resource entitlements.
• derive appropriate market-based resource
oDispatcher : The Dispatcher management strategies that encompass both
mechanism starts the execution of customer-driven service management and
accepted service requests on co mputational risk management to sustain
allocated VMs. SLA-oriented resource allocation,
o Service Request Monitor : The • incorporate autonomic resource management
Service Request Monitor mechanism models that effectively self-manage changes
keeps track o f the execution progress in service requirements to satisfy both new
of service requests. service demands and existing service
obligations, and • VMs : Multiple VMs can be started and
stopped dynamically on a single physical • leverage VM technology to dynamically
machine to meet accepted service requests, assign resource shares according to service
hence providing maximum flexibility to requirements.
configure various partitions o f resources on
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
Table 1: Co mparison of some rep resentative Cloud platforms.
System
Amaz on Sun
Google Microsoft GRIDS Lab
Elastic Compute Network.com
App Engine Live Mesh Aneka
Cloud (EC2) (Sun Grid)
Property
Software
Platform for
Focus Infrastructure Platform Infrastructure Infrastructure
enterprise
Clouds
Compute, Storage Web
Service Type Storage Compute Compute
(Amazon S3) application
Job
OS Level running Resource
Application management
Virtualisation on a Xen OS level Manager and
container system (S un
hypervisor Scheduler
Grid Engine)
SLA-based Dynamic
Negotiation of Resource
None None None None
QoS Reservation on
Aneka side. Parameters
Web-based
Job submission
Amazon EC2 Web-based Live Desktop Workbench,
User Access scripts, Sun
Command-line Administration and any devices Web-based
Interf ace Grid Web
Tools Console with Live Mesh portal
portal
installed
Web APIs Yes Yes Unknown Yes Yes
Value-added
Service Yes No No Yes No
Providers
APIs supporting
different
Customizable
Solaris OS, programming
Programming Linux-based
Python Not applicable Java, C, C++, models in C#
Amazon Machine Framework
FORTRAN and other .Net
Image (AMI)
supported
languages
Simple Storage Service (S3), before he can start, stop, data that can be accessed across required devices (such
and monitor instances o f the uploaded AMIs. Amazon as computers and mobile phones) from anywhere in the
EC2 charges the user for the time when the instance is wo rld. The user is able to access the upload ed
alive, while Amazo n S3 charges for any d ata transfer applications and data through a Web-based Live
(both upload and download). Dekstop or his own d evices with Live Mesh software
installed. Each user’s Live Mesh is password-protected
Google App Engine [14] allows a user to run Web
and authenticated via his Windows Live Login, while
app lications written using the Python programming
all file transfers are protected using Secure Socket
language. Other than supporting the Python standard
Layers (SSL).
library, Google App Engine also supports Application
Programming Interfaces (APIs) for the datastore, Sun network.com (Sun Grid) [1 6] enables the user
Google Accounts, URL fetch, image manipulation, and to run Solaris OS, Java, C, C++, and FORTRAN based
email services. Google App Engine also provides a applications. First, the user has to build and debug his
Web-based Administration Console fo r the user to applications and runtime scripts in a local development
easily manage his running Web applications. environment that is configured to be similar to that on
Currently, Google App Engine is free to use with up to the Sun Grid. Then, he needs to create a bundled zip
500MB of storage and about 5 million page views per archive (containing all the related scripts, libraries,
month. executable binaries and input data) and upload it to Sun
Grid. Finally, he can execute and monitor the
Microsoft Live Mesh [15] aims to provide a
application using the Sun Grid Web portal or API.
centralized location for a user to store applications and
10 10 10
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
After the completion of the ap plicatio n, the user will provider and sub-leasing these to the consumers. A
need to download the execution results to his local broker can accept requests fro m many users who have
development environment for viewing. a choice of submitting their requirements to different
brokers. Consumers, brokers and providers are bound
GRIDS Lab Aneka [17], which is being
to their requirements and related compensations
co mmercialized through Manjrasoft, is a .NET-based
throu gh SLAs. An SLA specifies the details of the
service-oriented platform for constructing enterprise
service to be provided in terms of metrics agreed upon
Grids. It is designed to support multiple application
by all parties, and penalties for meeting and violating
models, persistence and security solutions, and
the expectations, respectively.
co mmunication protocols such that the preferred
selection can be changed at anytime without affecting Such markets can bridge disparate Clouds allowing
an existing Aneka ecosystem. To create an enterprise consumers to choose a provider that suits their
Grid, the service provider only need s to start an requirements by either executing SLAs in advance o r
instance of the configurable Aneka container hosting by buying capacity on the sp ot. Providers can use the
required services on each selected desktop computer. markets in order to perform effective capacity
The purpose of the Aneka container is to initialize planning. A pro vider is equipped with a price-setting
services and acts as a single point for interaction with mechanism which sets the current price for the
the rest o f the enterprise Grid. Aneka provides SLA resource based on market conditions, user demand, and
support such that the user can specify QoS current level of utilization of the resource. Pricing can
requirements such as deadline (maximum time period be either fixed or variable depending o n the market
which the application needs to be completed in) and conditions. An admission-control mechanism at a
budget (maximum cost that the user is willing to pay provider’s end selects the auctions to participate in o r
for meeting the deadline). The user can access the the brokers to negotiate with, based on an initial
Aneka Enterprise Grid remotely through the Gridbus estimate of the utility. The negotiation process
broker. The Gridbus broker also enables the user to proceeds until an SLA is formed or the participants
negotiate and agree upon the QoS requirements to be decide to break off. These mechanisms interface with
provided by the service provider. the resource management systems of the provider in
order to guarantee the allocation being offered o r
negotiated can be reclaimed, so that SLA violations do 4. Global Cloud Exchange and Markets
not occur. The reso urce management system also
Enterprises currently employ Cloud services in order to provides functionalities such as advance reservations
improve the scalability of their services and to deal that enable guaranteed provisioning of resource
with bursts in resource demands. However, at present, capacity.
service providers have inflexible pricing, generally
Bro kers gain their utility through the difference
limited to flat rates or tariffs based on u sage thresholds,
between the price paid by the consu mers for gaining
and consumers are restricted to offerings from a single
resource shares and that paid to the providers for
provider at a time. Also, many providers have
leasing their resources. Therefore, a broker has to
proprietary interfaces to their services thus restricting
choose those users who se app lications can provide it
the ability of con sumers to swap one provider for
maximum utility. A broker interacts with resource
another.
providers and other brokers to gain or to trade resource
For Cloud co mputing to mature, it is required that shares. A broker is equipped with a negotiation module
the services follo w standard interfaces. This would that is informed by the current conditions of the
enable services to be commoditised and thus, would resources and the current demand to make its
pave the way for the creation of a market infrastructure decisions.
for trading in services. An example of such a market
Consumers have their own utility functions that
system, modeled on real-world exchanges, is sho wn in
cover factors such as deadlines, fidelity of results, and
Figure 4. The market directory allows participants to
turnaround time of applications. They are also
locate providers or consumers with the right offers.
constrained b y the amount of resources that they can
Auctioneers periodically clear bids and asks received
request at any time, usually b y a limited budget.
from market participants. The banking system ensures
Consumers also have their own limited IT
that financial transactions pertaining to agreements
infrastructure that is generally not completely exp osed
between participants are carried out.
to the Internet. Therefore, a consumer participates in
Brokers perform the same function in such a market the utility market through a resource management
as they do in real-world markets: they mediate between proxy that selects a set of brokers based on their
consumers and providers by buying capacity from the
11 11 11
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
Compute Cloud
Storage Cloud
Broker 1
Storage Cloud
Figure 4: Global Cloud exchange and market infrastructure for trading services.
offerings. He then forms SLAs with the brokers that obtain restitutio n in case an SLA is violated. This
bind the latter to provide the guaranteed resources. The motivates the need for a legal framework fo r
enterprise consu mer then deploys his own enviro nment agreements in such markets, a research issue that is out
on the leased resources or uses the provider’s of sco pe of themes pursued in this paper.
interfaces in order to scale his applications.
The idea of utility markets for computing resources 5. Summary and Conclusion
has been around for a long time. Recently, many
Cloud computing is a new and promising paradigm research projects such as SHARP [18], Tycoon [19],
delivering IT services as computing utilities. As Bellagio [20], and Shirako [21 ] have come up with
Clouds are designed to provide services to external market structures for trading in resource allocations.
users, providers need to be compensated for sharing These have particularly focu sed on trading in VM-
their resources and capabilities. In this paper, we have based resource slices on networked infrastructures such
proposed architecture for market-oriented allocation o f as PlanetLab. As mentioned before, the Gridbus project
resources within Clouds. We have discussed some has created a resource broker that is able to negotiate
representative platforms for Clo ud comp uting covering with resource providers. Thus, the technology for
the state-of-the-art. We have also presented a vision fo r enabling utility markets is already present and ready to
the creation of global Cloud exchange for trading be dep loyed.
services.
However, significant challenges persist in the
The state-of-the-art Cloud technologies have limited universal application of such markets. Enterprises
support for market-oriented resource management and currently employ conservative IT strategies and are
they need to be extended to support: negotiation o f unwilling to shift from the traditional controlled
QoS b etween users and providers to establish SLAs; environments. Cloud computing up take has only
mechanisms and algorithms fo r allocation of VM recently begun and many systems are in the proof-of-
resources to meet SLAs; and manage risks associated concept stage. Regu lato ry p ressures also mean that
with the violation of SLAs. Furthermore, interaction enterprises have to be careful about where their data
protocols needs to be extended to support gets processed , and therefore, are not able to emplo y
interoperability between d ifferent Cloud service Cloud services from an open market. This could be
providers. mitigated through SLAs that specify strict co nstraints
on the location of the resources. However, another As Cloud platforms become ubiq uitous, we expect
open issue is ho w the participants in such a market can
12 12 12
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.
the need for internetworking them to create a market- [10] Morgan Stanley. Technology Trends. 12 June
oriented global Cloud exchange for trading services. 2 008.
Several challenges need to be addressed to realize this http://www.morganstanley.com/institutional/tec
vision. They include: market-maker for bringing hresearch/pd fs/TechTrends062008.pdf [18 July
service providers and consumers; market registry for 2 008]
publishing and discovering Cloud service providers [11] K. Keahey, I. Foster, T. Freeman, and X.
and their services; clearing ho use and brokers for Zhang. Virtual workspaces: Achieving quality
mapping service requests to providers who can meet o f service and quality of life in the Grid.
QoS expectations; and payment management and S cientific Programming , 13(4):265-275,
accounting infrastructure for trading services. Finally, October 2005.
we need to address regulatory and legal issues, which
[12] I. Llorente, OpenNebula Project.
go b eyond technical issues.
http://www.opennebula.org/ [23 July 2008]
13 13 13
Aut hor i zed li cense d use li m it ed t o: Je ppi aar Eng ine er in g Col leg e. D own loa ded o n Augu st 25 , 20 09 at 12: 56 f r om I EEE Xp lor e . Rest r ict io ns app ly.