Csc8710 001 winter2014-mohammed_shahnawazali-ff2687_presentation_1

Opportunities and Challenges for running
Scientific Workflows on the Cloud
CSC 8710-001 – Presentation I
Mohammed Shahnawaz Ali

Executive Summary
Cloud computing has been mentioned over the recent years in
relation to services or infrastructural resources, which can be
contracted over a network, endorsing the idea of renting
infrastructure instead of buying it. Hence, cloud computing
infrastructures enables companies to cut costs by
outsourcing/offloading computations on-demand, thereby
gaining tremendous momentum in both academia and industry.
The application of cloud computing, however, has mostly
focused on Web applications and business applications; while
the recognition of using cloud computing to support large-scale
workflows, especially data intensive scientific workflows on the
cloud is still largely overlooked.

1/29/2014

CSC 8710 - Presentation I

2

Executive Summary (cont’d)
The paper coins the term “Cloud Workflow”, to refer to
the specification, execution, provenance tracking of large-scale
scientific workflows, as well as the management of data and
computing resources to enable the execution of scientific
workflows on the Cloud.
The paper analyzes:
1. Why there has been such a gap between the two
technologies,
2. What it means to bring Cloud and workflow together;
3. What are the key challenges in running Cloud workflow,
4. What are research opportunities in realizing workflows on
the Cloud.

1/29/2014


3

Introduction. Cloud – The Origin
The term “cloud” has its origins in network diagrams that
represented the internet, or various parts of it, as schematic
clouds. The term “Cloud computing” was defined for what
happens when applications and services are moved into the
internet “cloud.” Cloud computing is not something that
suddenly appeared overnight; in some form it may trace
back to a time when computer systems remotely time-shared
computing resources and applications.
More currently though, cloud computing refers to the many
different types of services and applications being delivered in the
internet cloud, and the fact that, in many cases, the devices used
to access these services and applications do not require any
special applications.
1/29/2014


4

Introduction. Cloud – The Representation
Cloud computing represents :
1. a different way to architect and remotely manage computing
resources.
2. network-based services, which appear to be provided by real
server hardware, and are in fact served up by virtual
hardware, simulated by software running on one or more
real machines
Cloud computing offerings today are suitable to:
1. host enterprise architectures and provide clear beneﬁt to
corporations by providing capabilities complementary to
what they have,
2. help elastically scale enterprise architectures.

1/29/2014


5

Introduction. Cloud – The Metaphor
Commonly used, the term "the cloud" is essentially a metaphor
for the Internet.
Marketers have further popularized the phrase "in the cloud" to
refer to software, platforms and infrastructure that are sold as a
service i.e. remotely through the Internet. Typically, the seller
has actual energy-consuming servers which host products and
services from a remote location, so end-users don't have to; they
can simply log on to the network without installing anything.
According to the field of interest, software, service or
infrastructure providers highlight different aspects.

1/29/2014


6

Introduction. Cloud – The Offerings
Some notable companies delivering services from the cloud
include:
1. Google — Has a private cloud that it uses for delivering many
different services to its users, including email access,
document applications, text translations, maps, web
analytics, and much more
2. Microsoft — Has Microsoft SharePoint online service that
allows for content and business intelligence tools to be
moved into the cloud, and Microsoft currently makes its
office applications available in a cloud.
3. Salesforce.com — Runs its application set for its customers in
a cloud, and its Force.com and Vmforce.com products
provide developers with platforms to build customized cloud
services.
1/29/2014


7

History
• 2006: August 24, 2006 conceivably goes down as the birthday
of Cloud Computing, as it was on this day that Amazon made
the test version of its Elastic Computing Cloud (EC2) public.
This offer, providing flexible IT resources (computing
capacity), marked a definitive milestone in dynamic business
relations between IT users and providers.
• 2007: The term first became popular in 2007, to which the
first entry in the English Wikipedia from March 3, 2007
attests, which, again significantly, contained a reference to
utility computing.
• 2008: In 2008, there was a glut of active parties in the
increasingly popular field of Cloud Computing.
• Today, Cloud Computing generates over 10.3 million matches
on Google.
1/29/2014


8

Cloud Computing

1/29/2014


9

Cloud Computing – The Characteristics
Cloud computing has a variety of characteristics, with the main
ones being:
1. Shared Infrastructure — Uses a virtualized software
model, enabling the sharing of physical services, storage, and
networking capabilities. The cloud infrastructure, regardless
of deployment model, seeks to make the most of the
available infrastructure across a number of users.
2. Dynamic Provisioning — Allows for the provision of services
based on current demand requirements. This is done
automatically using software automation, enabling the
expansion and contraction of service capability, as needed.
This dynamic scaling needs to be done while maintaining
high levels of reliability and security.
1/29/2014


10

Cloud Computing – The Characteristics (cont’d)
3. Network Access — Needs to be accessed across the internet
from a broad range of devices such as PCs, laptops, and
mobile devices, using standards-based APIs (for example,
ones based on HTTP). Deployments of services in the cloud
include everything from using business applications to the
latest application on the newest smartphones.
4. Managed Metering — Uses metering for managing and
optimizing the service and to provide reporting and billing
information. In this way, consumers are billed for services
according to how much they have actually used during the
billing period.

1/29/2014


11

Cloud Computing – The Service Models
Once a cloud is established, how its cloud computing services are
deployed in terms of business models can differ depending on
requirements.
The primary service models being deployed are commonly known as:
1. Software as a Service (SaaS) — Consumers purchase the ability to
access and use an application or service that is hosted in the cloud.
A benchmark example of this is Salesforce.com, as discussed
previously, where necessary information for the interaction
between the consumer and the service is hosted as part of the
service in the cloud.
2. Platform as a Service (PaaS) — Consumers purchase access to the
platforms, enabling them to deploy their own software and
applications in the cloud. The operating systems and network
access are not managed by the consumer, and there might be
constraints as to which applications can be deployed.
1/29/2014


12

Cloud Computing – The Service Models (cont’d)
3. Infrastructure as a Service (IaaS) — Consumers control and
manage the systems in terms of the operating
systems, applications, storage, and network connectivity, but
do not themselves control the cloud infrastructure.
4. Communications as a Service (CaaS) — is a model used to
describe hosted IP telephony services. Along with the move
to CaaS is a shift to more IP-centric communications and
more SIP trunking deployments. With IP and SIP in place, it
can be as easy to have the PBX in the cloud as it is to have it
on the premise. In this context, CaaS could be seen as a
subset of SaaS

1/29/2014


13

Cloud Computing – The Deployment Models
Deploying cloud computing can differ depending on
requirements, and the following four deployment models have
been identified, each with specific characteristics that support
the needs of the services and users of the clouds in particular
ways:
1. Private Cloud — The cloud infrastructure has been deployed,
and is maintained and operated for a specific organization.
The operation may be in-house or with a third party on the
premises.
2. Community Cloud — The cloud infrastructure is shared
among a number of organizations with similar interests and
requirements. This may help limit the capital expenditure
costs for its establishment as the costs are shared among the
organizations. The operation may be in-house or with a third
party on the premises.
1/29/2014


14

Cloud Computing – The Deployment Models (cont’d)
3. Public Cloud — The cloud infrastructure is available to the
public on a commercial basis by a cloud service provider. This
enables a consumer to develop and deploy a service in the
cloud with very little financial outlay compared to the capital
expenditure requirements normally associated with other
deployment options.
4. Hybrid Cloud — The cloud infrastructure consists of a
number of clouds of any type, but the clouds have the ability
through their interfaces to allow data and/or applications to
be moved from one cloud to another. This can be a
combination of private and public clouds that support the
requirement to retain some data in an organization, and also
the need to offer services in the cloud.
1/29/2014


15

Cloud Computing – The Benefits
The following are some of the possible benefits for those who
offer cloud computing-based services and applications:
1. Cost Savings — Companies can reduce their capital
expenditures and use operational expenditures for increasing
their computing capabilities. This is a lower barrier to entry
and also requires fewer in-house IT resources to provide
system support.
2. Scalability/Flexibility — Companies can start with a small
deployment and grow to a large deployment fairly
rapidly, and then scale back if necessary. Also, the flexibility
of cloud computing allows companies to use extra resources
at peak times, enabling them to satisfy consumer demands.

1/29/2014


16

Cloud Computing – The Benefits (cont’d)
3. Reliability — Services using multiple redundant sites can
support business continuity and disaster recovery.
4. Maintenance — Cloud service providers do the system
maintenance, and access is through APIs that do not require
application installations onto PCs, thus further reducing
maintenance requirements.
5. Mobile Accessible — Mobile workers have increased
productivity due to systems accessible in an infrastructure
available from anywhere

1/29/2014


17

Scientific Workflows

1/29/2014


18

Scientific Workflows
• Scientific Workflows are an amalgamation of scientific
problem-solving and traditional workflow techniques.
• These are another class of workflows, in addition to the
business workflows, that emerge in sophisticated scientific
problem-solving environments and applications viz., climate
modeling, structural biology and chemistry, medical surgery or
disaster recovery simulation.
• Compared with business workflows, scientific workflow has
special features such as computation, data or transaction
intensity, less human interaction, and a large number of
activities.
1/29/2014


19

Scientific Workflow Management System
The reference architecture for SWFMS consists of four logical
layers, seven major functional subsystems, and six interfaces:
1. Operational Layer: consists of a wide range of heterogeneous
and distributed data sources, software tools, services, and
their operational environments, including high end
computing environments.
2. Task Management Layer: consists of three subsystems: Data
Product Management, Provenance Management, and Task
Management.
3. Workflow Management Layer: consists of Workflow Engine
and Workflow Monitoring.
4. Presentation Layer: consists of the Workflow Design
subsystem and the Presentation and Visualization subsystem
1/29/2014


20

Scientific Workflows – On The Cloud Today
Cloud computing have been widely accepted and applied to Web
applications and business applications. However, the cloud
capabilities have not been successfully extended to execute and
manage workflow applications, especially data-intensive
scientific workflows.
The current state of workflow organization on the Cloud has
been either:
1. static predefined pipelines based on batch style scripts or
graphs based on the MapReduce programming model
2. ad hoc mash-up’s that are connected together
with, again, scripts that parse the output of one web
application and feed into another.
1/29/2014


21

Scientific Workflows – On The Cloud Today (cont’d)
Several scientific workflow management systems (SWFMSs) have
been successfully applied over a number of execution
environments viz., local hosts, clusters/grids, and
supercomputers.
However, Cloud computing provides a paradigm-shifting utilityoriented computing model in terms of the unprecedented size of
datacenter-level resource pool and the on-demand resource
provisioning mechanism, enabling scientific workflow solutions
capable of addressing peta-scale scientific problems.

1/29/2014


22

Cloud Workflow – Bringing them together
The term Cloud Workflow is constructed to bring the terms Cloud and Scientific Workflows
together.

It refers to the following attributes of scientific workflows
1. specification,
2. execution,
3. provenance tracking
along with management of data and computing resources to enable the running of scientific
1/29/2014


23

Opportunities

1/29/2014


24

Cloud Workflow – The Opportunities
Clouds provide a multiplicity of opportunities that are more
technological in nature, and that these opportunities
stem, primarily, from the extensive use of service-oriented
architectures
and virtualization in clouds
Scalability:
• The scale of scientific problems:
• that can be addressed by scientific workflows is now greatly
increased, which was previously limited by the size of a
dedicated resource pool.
• is reflected not only on the data sizes that scientific applications
need to handle, but also on the complexities of the applications
themselves.
• Cloud platforms can offer vast amount of storage space as well as
computing resources for applications across multiple disciplines
including physics, earth science, and medicine, allowing scientific
discoveries to be carried out in an unprecedented scale.
1/29/2014


25

Dynamic resource allocation:

• By allocating the resources only when they are needed, it
presents various advantages including:
• Optimum resource utilization.

• Improved end user experience.
• Collaborative batch based scientific workflows.

1/29/2014


26

Relinquish allocated resources:

• Cloud allows users to return resources on-demand
• Enables workflow systems to easily grow and shrink the
available resource pool as the needs of the workflow change
over time
• Closely match the needs of the application by acquiring or
releasing resources for optimal usage

1/29/2014


27

Performance to Cost Trade/off:
• Cloud computing provides a much larger room for the tradeoff between performance and cost.
• The spectrum of resource investment now ranges from:
• dedicated private resources,
• hybrid resource pool combining local resource and remote
clouds,
• full outsourcing of computing and storage to public
Clouds.

• Cloud Computing not only provides the potential of solving
larger-scale scientific problems, but also brings the
opportunity to improve the performance/cost ratio.
1/29/2014


28

Heterogeneous Applications Support:

• Clouds and their use of virtualization technology makes
different heterogeneous applications much easier to run
together.
• Virtualization enables the environment to be customized to
suit the application.
• Environment with Operating System, applications and their
configurations can be bundled up as a virtual machine image
and redeployed on a cloud to run the workflow.

1/29/2014


29

Resource Provisioning:

• Instead of delegating allocation to the resource manager, the
user directly provisions the resources required and schedules
their computations using a user-controlled scheduler.
• Provisioning model is ideal for workflows and other looselycoupled applications because it enables the application to
allocate a resource once and use it to execute many tasks.
• Reduces the total scheduling overhead which, in turn, can
dramatically improve workflow performance

1/29/2014


30

Provenance and Re-Imaging:

• Virtualization allows one to capture the exact environment
that was used to perform a computation, including all of the
software and configuration used in that environment.
• Virtual machine image can be stored along with the
provenance of the workflow.
• Redeploy the virtual machine image to create exactly the
same environment that was used to run the original
experiment.

1/29/2014


31

Challenges

1/29/2014


32

Cloud Workflow – The Challenges
Despite the advantages and opportunities we can seek in
Cloud computing for scientific workflows, there are many
major obstacles to the adaptation and running of scientific
Architectural Challenge:
The following seven are key architectural requirements for an SWFMS:
1. User interface customizability and user interaction support.
2. Reproducibility support
3. Heterogeneous and distributed services and software tools
integration.
4. Heterogeneous and distributed data product management.
5. High-end computing support.
6. Workflow monitoring and failure handling.
7. Interoperability.
1/29/2014


33

Architectural Challenge (cont’d):
There are four possible solutions for deploying the reference
architecture in a Cloud computing environment:
1. Operational-Layer-in-the-Cloud: only the Operational Layer is
deployed in the Cloud with an SWFMS running out of the Cloud.
2. Task-Management-Layer-in-the-Cloud: both the Operational Layer
and the Task Management Layer are deployed in the Cloud.
3. Workflow-Management-Layer-in-the-Cloud: the Operational Layer,
the Task Management Layer, and the Workflow Management Layer
are deployed in the Cloud with the Presentation Layer deployed at
a client machine.
4. All-in-the-Cloud: The whole SWFMS is deployed inside the Cloud
and accessible via a Web browser.
1/29/2014


34

Integration Challenge:
The integration problem includes the following:
1. In the operational-layer-in-the-Cloud approach, we treat
applications, services, and tools hosted in the Cloud as task units in
a workflow, the scheduling and management of a workflow are
mostly outside the Cloud, where these task units are invoked as
they are scheduled to execute.
2. Once we decide to get task dispatching and scheduling into the
Cloud, resource provisioning becomes the next issue.
3. The uncapped resources requested by a workflow comes at a cost.
4. Debugging, monitoring, and provenance tracking for a workflow
can be even more difficult in the Cloud, since resources are usually
dynamically assigned and based on virtual machine instances, the
environment that a task is executed on could be destroyed right
after the task is finished, and assigned to a complete different user
and task.
5. Porting an SWFMS into the Cloud is also a concern, which usually
involves wrapping up an SWFMS as a Cloud service.
1/29/2014


35

Language Challenge:
Language adopted for cloud computing include:
1. MapReduce is the “only” widely adopted computing model, and
there are a number of variations of languages based on this model
for task specification in the Cloud.
2. White-Box approach: MapReduce and its variations require
application logic to be rewritten to follow the map-reduce-merge
programming model. Thus, users need to fully understand the
applications and port the applications before they can leverage the
parallel computing infrastructure.
3. Black-Box approach: SwiftScript serves as a general purpose
coordination language, where existing applications can be invoked
without modification.
4. Mash-up’s and ad hoc scripts (Java Script, PHP, Python, etc.) have
become key technologies for developing Web applications that
dynamically integrate multiple data or service sources.
1/29/2014


36

Language Challenge (cont’d):
The language challenges includes the following:
1. Handle the mapping from input and output data into logical
structures to facilitate data integration and logical operations
on data.
2. Support large-scale parallelism via either implicit parallelism,
or explicit declaratives such as Parallel Foreach.
3. Support data partitioning and task partitioning.
4. Require a scalable, reliable, and efficient runtime system that
can support Cloud-scale task scheduling and dispatching,
provide error recovery and fault tolerance under all kinds of
hardware and service failures, and utilize a large pool of
Cloud resources efficiently.
1/29/2014


37

Computing Challenge:

The computing challenges includes the following:
1. Managing large-scale of computing resources.
2. Workflow systems may not be able to talk to Cloud resources
directly, they may still need go through middleware services
such as Nimbus and Falkon that handle resource provisioning
and task dispatching.
3. Workflow resource requirements, data dependencies, Cloud
virtualization, etc makes thing even more complicated.
4. Additional measures is needed to support large workflows
and components.
1/29/2014


38

Data Management Challenge:
The data management challenges includes the following:
1. Analyzing, visualizing, and disseminating of large data sets.
2. Management of data resources and dataflow between the storage
and resources in data intensive applications.
The following aspects of data management within a Cloud are
important from a workflow perspective:
1. Data Locality:
a. Location of the data relative to the available computational
resources. Moving data repeatedly to distant CPUs is
expensive and inefficient.
b. Data need to be distributed over many computers to achieve
good scalability.
2. Combining compute and data resource management.
3. Scalability of Clouds require scalable provenance systems to
handle storage and querying of potentially millions of tasks.
1/29/2014


39

Service Management Challenge:

The service management (Orchestrating and invoking services
via an SWFMS) challenges includes the following:
1. Service description, discovery, and composition.
2. Managing the large number of service instances.
3. Data movements across service instances involving large data
volumes.
4. For a workflow to invoke publicly available services SWFMS
also needs to handle security, interoperability and data
transformation issues.

1/29/2014


40

Storage Challenge:
The storage challenges includes the following:
1. Commercial clouds often deploy structured or object-based
storage services that can be utilized by workflow
applications.
2. In the absence of standard file system interfaces, the
application codes must either be modified to interface with
the storage services, or must be wrapped with additional
workflow components to do the translation.
3. Deploying a temporary shared file system in the cloud as part
of a virtual cluster is complex, potentially costly and requires
additional step to ensure that desired outputs are
transferred to permanent storage.
4. Storage security.
1/29/2014


41

Network and Tools Challenge:

The network and tools challenges includes the following:
1. Data-intensive workflows depend on high-performance
networks to achieve good performance.
2. Requires high-throughput, but not necessarily low
latency, and faster networks.
3. Setting up an environment to run workflows in the cloud.
4. There is some work in virtual appliances, but those are
typically designed for single nodes and not for clusters of
nodes.

1/29/2014


42

Research Directions

1/29/2014


43

Cloud Workflow – The Research Directions
The key areas for research efforts in cloud based workflows:
Architecture:
1. Implement the key components in the different layers of the
SWFMS architecture, with interoperability and reusability. This
would help us leverage existing Cloud technologies, such as
monitoring data management, resource provisioning, etc.
2. Leverage middleware technologies that bridge existing workflow
systems with the Cloud to be more cost effective.
Scripting:
1. Scripting has the advantage of being concise and flexible, yet
powerful when combined with parallel semantics and logical
operations.
2. Expect to see scripting languages that have a mixture of these
semantics, combining the coordination of applications and
services.
Cost:
1. Analyze the cost for computation and resource utilization to
estimate and optimize the ROI.
1/29/2014


44

Cloud Workflow – The Research Directions
Provenance:
1. Can adopt the SOA model making provenance less coupled with an
SWFMS than it currently does.
Security:
1. It is the first major service that needs to be provided by a Cloud
provider.
a. Access Control: Due to the dynamic nature and the large-scale
data, metadata, and service sharing nature of the Cloud,
access control is a challenging but important research
problem.
b. Information Control Flow: Since a scientific workflow might
orchestrate a large number of distributed services, data, and
applications, particularly in a large-scale Cloud environment,
the mechanism that controls mission-critical information and
intellectual property not being propagated to an unauthorized
user is important.
c. Secure electronic transaction protocol: To prevent the abuse of
Cloud accounts and double or wrong charges by a Cloud
provider further research might be needed to ensure the
security of Cloud-based transaction protocol.
1/29/2014


45

Closing Notes
• The benefit of cloud computing for science is not necessarily in its
utility computing and economic aspects, which are not new for
academic computing. The benefit of clouds is rather in its
technological features that stem from service-oriented architecture
and virtualization
• Much work is needed to bring cloud platforms up to the
performance level of the grid. This includes developing cloud
storage systems that are appropriate for workflow and other
science applications as well as tools to help scientists and workflow
engineers deploy their applications in the cloud.
• As more and more customers and applications migrate into Cloud,
the requirement to have workflow systems to manage the ever
more complex task dependencies, and to handle issues such as
large parameter space exploration, smart reruns, and provenance
tracking will become more urgent.
• Cloud needs more structured and mature workflow technologies,
and vice versa, as Cloud offers unprecedented scalability to
workflow systems, and could potentially change the way we
perceive and conduct scientific experiments.
1/29/2014


46

Thank You

1/29/2014


47

Csc8710 001 winter2014-mohammed_shahnawazali-ff2687_presentation_1

Recommended

Recommended

More Related Content

What's hot

What's hot (18)

Similar to Csc8710 001 winter2014-mohammed_shahnawazali-ff2687_presentation_1

Similar to Csc8710 001 winter2014-mohammed_shahnawazali-ff2687_presentation_1 (20)

Recently uploaded

Recently uploaded (20)

Csc8710 001 winter2014-mohammed_shahnawazali-ff2687_presentation_1

Editor's Notes