Modelling Cloud systems

What is Cloud computing?

Cloud computing is a general term for anything that involves delivering hosted services over the Internet

Clouds can be delivered according to different service and deployment models

Service models

Deployment models

Advantages

• Scalability
• Adaptability
• Accessiblity
• Maintenance
• Reduced cost

However...

Cloud systems are complex

⇨

Modelling

Queueing theory

Nonlinear programming

Cloud simulation

Cloud platform

Queueing Theory

• Mathematical method of analyzing the congestions and delays of waiting in line
• Used to develop more efficient queuing systems

Queueing Theory

A Cloud system can be transformed to...

A queueing theory model

A queueing theory model

Queueing Theory

• Transform Cloud systems to queueing theory assets
• Predict behaviours through queueing theory formulas

⇨

Mathematical formulas

Nonlinear programming

A nonlinear programming problem (NLP) deals with mathematical optimization problems where the objective function to be maximized or minimized, or some of its constraints, are nonlinear.

Nonlinear programming

A NLP can be defined as:

\begin{equation} Max(f(x_{1},x_{2}, ..., x_{n})), \end{equation}

s.t. (subject to):

\begin{equation} \begin{split} h_{i}(x) = b_{i} & \quad (i = 1, 2, ..., m) \\ x_{j} \geq 0 & \quad (j = 1, 2, ..., n) \end{split} \end{equation}

Then, it can be processed by a solver to extract an optimal solution

Cloud simulation

Why Cloud simulation?

• Timely, repeatable and controllable methodologies for evaluation of algorithms, applications and policies
• Utilization of real testbeds limits the experiments to their scale and makes the reproduction of results an extremely difficult undertaking

The utilization of simulations tools opens the possibility of evaluating the hypothesis prior to software development in an environment where one can reproduce tests

• Wide ecosystem of cloud architectures, demand... ...before actual development of cloud products.
• , alternative approaches for testing and experimentation leverage development of new Cloud technologies.

Cloud simulation tools

• CloudSim
• CloudAnalyst
• GreenCloud
• iCanCloud
• MDCSim
• NetworkCloudSim
• VirtualCloud

CloudSim is an event-driven and extensible simulation toolkit that enables Cloud systems and scenarios to be modelled and simulated

CloudSim

Modeling and simulation of...

• Large scale Cloud computing data centers
• Virtualized server hosts, with customizable policies for provisioning host resources to virtual machines
• Energy-aware computational resources
• Data center network topologies and message-passing applications
• Federated Clouds
• Dynamic insertion of simulation elements, stop and resume of simulation
• User-defined policies for allocation of hosts to VMs and policies for allocation of host resources to VMs

Cloud system as a CloudSim model

Cloud platform

Current Cloud platforms

• Amazon EC2
• Microsoft Azure
• Google Cloud Platform
• OpenStack
• Apache CloudStack
• OpenNebula

• ⇨ Commercial
• ⇨ Commercial
• ⇨ Commercial
• ⇨ Open Source
• ⇨ Open Source
• ⇨ Open Source

And much more...

OpenStack History

• Joint project with Rackspace & NASA
• Launched in June 2010
• Enable anyone to create and offer Cloud computing services
• Many corporations joined

OpenStack overview

OpenStack Components

• Nova (compute)
• Swift (object storage)
• Glance (image service)
• Keystone (identity management)
• Horizon (gui interface)

Cloud computing challenges and issues

• Reliability, availability and serviceability
• System security, user privacy and trust issues
• Performance and energy consumption issues

What is eHealth?

Union of the healthcare and the information and communications technology (ICT) areas

Telemedicine

Electronic Health Records

Patient Data Management

mHealth

Motivations for eHealth

Cloud Computing + eHealth

• Access to computing resources
• Seamles deployment
• Data sharing and analysis
• Personalized follow-up
• Effective time management
• Reduced transport costs

Research objectives

Achieve eHealth solutions through Cloud-based technologies and architectures

• Design a Cloud system by means of queueing theory
• Design of a nonlinear programming scheduling algorithm to optimize both energy consumption and SLA guarantees
• Design a Cloud architecture for an eHealth environment and analyze its behaviour with the CloudSim simulator
• Design a telemonitoring tool for smoke-quitting patients
• Design a telemonitoring tool for hypertensive patients
• Develop a scalable and power-aware Cloud-based infrastructure using the OpenStack platform

Related Work and Contributions

Publications

Journal Publications

• 2013 - Vilaplana, J. et al. The cloud paradigm applied to e-Health. BMC Medical Informatics and Decision Making
• 2014 - Vilaplana, J. et al. S-PC: An e-treatment application for management of smoke-quitting patients. Computer Methods and Programs in Biomedicine
• 2014 - Vilaplana, J. et al. A queuing theory model for cloud computing. The Journal of Supercomputing
• 2014 - Vilaplana, J. et al. Database Constraints Applied to Metabolic Pathway Reconstruction Tools. The Scientific World Journal
• 2014 - Vilaplana, J. et al. H-PC: a cloud computing tool for supervising hypertensive patients. The Journal of Supercomputing
• 2014 - Vilaplana, J. et al. An SLA and power-saving scheduling consolidation strategy for shared and heterogeneous clouds. The Journal of Supercomputing

Conference publications and attendance

• 2012 - Vilaplana, J. et al. Diseño de un Sistema Cloud Aplicado a e-Health. JP2012. Elche, Spain
• 2014 - Vilaplana, J. et al. An SLA & Power Aware Strategy for a Cloud. ITME2014. Hong Kong
• 2014 - Vilaplana, J. et al. A Green Job Scheduling Policy for Heterogeneous Clouds. CMMSE 2014. Rota, Cádiz, Spain
• 2014 - Vilaplana, J. et al. SLA-Aware Load Balancing in a Web-Based Cloud System over OpenStack. ICSOC 2013 Workshops. CCSA. Berlin, Germany
• 2014 - Vilaplana, J. et al. A Green Scheduling Policy for Cloud Computing. ARMS-CC 2014, PODC 2014. Paris, France
• 2015 - Vilaplana, J. et al. A performance model for scalable cloud computing. AusPDC 2015. Sydney, Australia

Presents the design of a Cloud platform with QoS guarantees based on response time for services.

The main contribution consists of a computer service QoS model for the Cloud architecture made up of processing servers and a data service

• In this model the Cloud is a single access point for the computing needs of the customers being served
• The service center is a collection of service resources used by a service provider to host service applications
• A service request is transmitted to the web server running a service application, which is associated with an SLA
• The SLA is a contract negotiated and agreed between a customer and a service provider
• Customers generate service requests at a given rate for processing at the service center according to negotiated QoS requirements

Obtaining $T_{ES}$

$T_{ES}$ represents the response time of the Entering Server ($ES$) which acts as a load balancer. The $ES$ node is modeled as an $M/M/1$ queue. Thus, the formula for obtaining the response time for an $M/M/1$ queue is:

\begin{equation} \label{eqTES} T_{ES}=\frac{1/L}{1-\lambda/L}, \end{equation}

where $\lambda$ is the arrival rate, and $L$ is the service rate of the node $ES$.

Obtaining $T_{PS}$

$T_{PS}$ represents the response time of the $PS$ nodes. These nodes are modeled as an $M/M/m$ queue.

\begin{equation} T_{PS}=\frac{1}{\mu}+\frac{C(m, \rho)}{m\mu-\gamma}, \end{equation}

$m$ is the number of PEs, $\gamma$ and $\mu=\mu_i, i=1..m$, are respectively the arrival and service rates of each PE respectively.

Obtaining $T_{PS}$

$C(m, \rho)$ represents Erlang's C formula, which gives the probability of a new client joining the $M/M/m$ queue:

\begin{equation} C\left ( m , \rho \right ) = \frac{\left ( \frac{(mp)^m}{m!} \right ) \left ( \frac{1}{1-\rho} \right )}{\sum_{k=0}^{m-1} \frac{(mp)^k}{k!}+\left ( \frac{(m\rho)^m}{m!} \right ) \left ( \frac{1}{1-\rho} \right )}, \end{equation}

where $\rho=\gamma/\mu$.

Obtaining $T_{DS}$

$T_{DS}$ represents the response time of the $DS$. Requests are sent to the $DS$ node with a probability $\delta$. The $DS$ node is modeled as a $M/M/1$ queue:

\begin{equation}\label{eqTDS} T_{DS}=\frac{1/D}{1-\delta\gamma/D}, \end{equation}

$\delta\gamma$ is the arrival rate to $DS$, and $D$ is the service rate.

The arrival rate at the $OS$ is the sum of the arrival rates of the two crosspoint branches entering the summatory. The one crossing the data server is the same than at its input ($\delta\gamma$). On the other branch, the $\gamma$ term must be changed by its complement to one.

\begin{equation} (1-\delta)\gamma + \delta\gamma = \gamma \end{equation}

Obtaining $T_{OS}$

$T_{OS}$ represents the $OS$ response time. Its operation is also modeled as a $M/M/1$ queue. Its service rate is defined as $O/F$, where $O$ is its average BW speed, and $F$ is the averaged size of the data responses of the system.

\begin{eqnarray}\label{eqTOS} T_{OS} &=& \frac{F/O}{1-\gamma/(O/F)} \nonumber\\ T_{OS} &=& \frac{F}{O-\gamma F} \end{eqnarray}

Obtaining $T_{CS}$

Finally, $T_{CS}$ is the $CS$ response time, that operates as an $M/M/1$ queue. The service rate is defined as $C/F$. where $C$ is the avg BW speed of the client server. $F$ is the average size of the received reply files.

\begin{eqnarray}\label{eqTCS} T_{CS} &=& \frac{F/C}{1-\gamma/(C/F)} \nonumber\\ T_{CS} &=& \frac{F}{C - \gamma F} \end{eqnarray}

Results

• Analysis of how the response time is affected by modifying some of the metrics presented in the model.
• Our purpose is to verify if our model behaves as expected when a range of parameters and system configurations are tested.
• The model was implemented using Sage 5.3 mathematical software.

Results

Parameters used in the implementation:

• $\lambda$ Arrival rate.
• $F$ Average file size.
• $O$ Server bandwidth.
• $C$ Client bandwidth.
• $\delta$ Database access probability.
• $\mu$ Service rate.

$\lambda$ is the averaged number of requests reaching the system per unit of time. $1/\lambda$ is the mean inter-arrival time. We aimed to show how the total response time ($T$) is affected by varying the parameter $\lambda$. The number of processing servers ($m$) represents the total number of processors, cores or nodes dedicated to servicing requests. Changing this parameter will show the impact of adding or removing servers from the system.

$F$ Average file size. This is the mean size of the files that are sent to clients via Internet as the service response of the overall system. This value depends on the web application that runs on top of the system, although it should be no greater than 1MB in most cases.

$O$ Server bandwidth. This is the network speed at which files are sent to clients over the Internet.

$C$ Client bandwidth. This is the average connection speed of the clients receiving the data sent from the system. This parameter will usually be outwith our control.

$\delta$ Database access probability. This is the probability that an incoming request needs to access the database node $DS$. As we are modeling a web-based system, not all requests will always require an access to the database server. Note, however, that this probability will usually be relatively high.

$\mu$ Service rate. This describes the speed at which the web servers handle the requests. \ $1/\mu$ is the mean service time. To simplify the results we have chosen the same value for all $PS_{i}$ servers. The total service rate of the system must always be greater than $\lambda$ for the system to be stable. This value is modified through the tests in order to check the performance impact. Although the service rates of $ES$, $PS_i, i=1..m$, $DS$, $OS$ and $CS$ could be different, we assumed the same value to simplify the experiments. So, in this case, $\mu=L=D=O/F=C/F$ in all the tests.

Response Time

Bottlenecks

Model validation

Implemented using the OpenStack Cloud platform

Conclusions

• Model for designing cloud computing architectures with QoS
• To provide good QoS, system bottlenecks have to be determined
• Queuing theory is a valid method to obtain performance models
• Such models can be very useful for tuning service performance

Presents a power-aware scheduling policy algorithm of VMs into nodes called GreenC

GreenC takes into account optimal assignments according to physical machine and VM heterogeneity, the current host workload and communication between the different VMs

• Based on nonlinear programming (NLP)
• Aims to minimize both energy consumption and response time
• Takes into account host and VM heterogeneity, workload and communication
• Theoretical experimentation using a solver
• Implementation using the OpenStack platform

Host heterogeneity

The relative computing power ($\Delta_{v}$) of a Host $H_{v}$ is defined as its normalized score

$\Delta_{v} = \frac{\delta_{v}}{\sum_{k=1}^{V} \delta_{k}}$

$\delta_{v}$ is a valid score of $H_{v}$

$\delta_{v}$ is a theoretical concept ➩ obtained using a benchmark (Linpack, SPEC, ...)

Virtual machine heterogeneity

Each VM has its Processing cost $P_v^i$, representing the execution time of VM, $\mho^i$, in $H_{v}$ with respect to the execution time of VM $\mho_{i}$ in the less powerful $H_{v}$

$M_v^i$ is the amount of Memory allocated to VM $\mho_{i}$ in $H_{v}$

Once the solver is executed, $\mho_v^i$ variables will inform about the assignment of VM to hosts

$\mho_v^i=1$ if $\mho^i$ is assigned to $H_v$, and $\mho_v^i=0$ otherwise

Host Workload

• If a host is underloaded, its throughput will increase as more VM are assigned to it
• When the host reaches its maximum workload capacity, its throughput starts falling asymptotically towards zero

We can model this behaviour with an Erlang distribution density function:

\begin{equation} \label{erlang_22} E(x;\alpha,\lambda) = \lambda e^{-\lambda x} \frac{(\lambda x)^{\alpha - 1}}{(\alpha - 1)!} \forall x,\lambda \geq 0 \end{equation}

Erlang distributions

Virtual Machine Communication

The Communication Cost between VMs is denoted by $C^{ij}$

Notation $C_v^{ij}$ represents the Communication Cost between $\mho^i$ residing in $H_v$ with other $\mho^j$

• Penalize (enhance) the communications performed between different hosts
• Grouping VMs inside the same host not only will depend on their respective Processing cost ($P_v^i$), but also in the relative communicating time $C_v^i$

The GreenC policy

\begin{align} \tiny max(\sum_{v=1}^{V}(\sum_{i=1}^{T}\mho_v^i&(P_v^i+\sum_{j \ st \ i>j}^T (C_v^{ij}\mho_v^jC_s + C_v^{ij}\mho_{\overline{v}}^j))) \Delta_{v} E(\sum_{j=1}^{T} P_v^j \cdot \mho_v^j;\alpha,\lambda)) \\ \small s.t. \ \:&\sum_{v=1}^{V} \mho_v^i=1 \ \ \ \forall i \leq T \\ \small &\sum_{i=1}^{T} M_v^i \leq M_v \ \ \ \forall v \leq V \end{align}

Theoretical Experimentation

Communication configurations

Erlang-shaped hosts

Host configurations

Implementation

• Four identical physical hosts
• Each host is a HP Proliant DL165 G7 with two AMD Opteron 6274 processors with 16 cores each at 2.2 GHz, 112 GB of RAM, 600 GB of SAS disk and 1 Gb Ethernet network
• CentOS release 6.5 has been used as operating system

• Energy consumption has been empirically monitored
• The following devices were used:
  • Four Individual Appliance Monitors (IAMs)
  • EnviR that transmits all the data collected from the IAMs
• These devices allow us to individually monitor the energy consumption (in Watts) of each physical host every second and transfer the data to a text file for further analysis.
• Room temperature was monitored with a Thermometer and remained constant at 27º C throughout the experimentation

Conclusions

• Cloud-based system scheduling mechanism to reduce power consumption and maintain SLA agreements
• Model increasingly complex
• First tested using a mathematical optimizer. Obtained results proved consistent over a range of scenarios
• Further experimentation in a Cloud platform was performed and the energy consumption was empirically monitored
• Initial results proved encouraging, achieving up to a 23% energy consumption reduction

Presents the computer application H-PC (Hypertension Patient Control), which allows patients with hypertension to send their readings through mobile phone SMS (Short Message Service) or e-mail to a cloud computing datacenter

Clinicians can keep track of their patients, thus facilitating monitoring

H-PC features and operation

• Designed for collecting and managing data from hypertensive patients
• Record and print/display measurement statistics
• Graphically show patients' evolution using charts
• Automatic control of risk communication by SMS messages or e-mail to aid clinicians to diagnose and generate alerts or suggestions for treatments, patient monitoring, medication, nutrition, etc.

H-PC allows target limits to be established individually from both systole and diastole blood pressures, depending on patient characteristics

If these limits are exceeded, an alert is shown in the main page of the H-PC tool so clinicians can act quickly and, if needed, perform an intervention or send an alert to the patient

H-PC readings

• Readings can be registered automatically from SMS or e-mails or can be introduced manually by the clinicians
• H-PC automatically calculates the mean values for each day, showing only one value per day only in the plot
• Individual readings can also be viewed in table format

Pulse can also be registered when manually entering the data or when sending it via e-mail

H-PC performs a data verification check in order to avoid incorrect or invalid measurements

H-PC operation

Design Principles

• Multi-platform and multi-language web-based application with a user-friendly graphic user interface (GUI)
• Provides the clinician easy function access and utilization
• Implemented using Java for the back-end and Javascript, CSS, XHTML and AJAX for the front-end
• H-PC can run on any computer, operating system and any of the major web-browsers
• Interface developed using a responsive design approach

Usability and user-friendliness criteria

Architecture

H-PC features

• Scalability
• Reliability
• Scheduler

Implementation

Results

Conclusions

Presents a new program that facilitates the management of people who want to quit smoking, implemented through an e-treatment software called S-PC (Smoker Patient Control)

Manages groups of patients, provides a bidirectional communication between patients and clinicians and offers advice and control to keep track of the patients and their status

• A total of 229 patients were enrolled in the study
• Randomly divided into two groups
• No significant differences regarding the m/f ratio, tobacco dependence, co-oximetry, avg. cigarette consumption, current age and age when smoking started
• The first group was made up of 104 patients (45.4% of total) and followed a treatment that incorporated the S-PC tool
• The second one had 125 patients without the S-PC tool

Results

• S-PC Effectiveness
• Patient Satisfaction

Patient Satisfaction

• 95% followed the treatment
• 92% of the patients were either satisfied or completely satisfied with the support given by S-PC
• 20% felt that they would still have managed without the system
• 70% strongly agreed that the S-PC system helped them remain smoke-free
• 20% disagreed or strongly disagreed that S-PC helped them remain smoke-free
• 10% had no opinion about this issue

Conclusions

• S-PC was successfully designed, implemented and used in the context of the quit smoking treatment
• Usefulness of text messaging in improving:
  • Outcome likelihood of smoking cessation interventions
  • Management of time and patients by clinicians, optimizing the health care resources and the reduction of waiting lists
  • Patient’s perception of constant psychological support by the clinician

Belfast, Northern Ireland

February - March 2015

Reader in Computer Science

• Big data analytics for satellite data exploitation
• Dempster-Shafer theory of evidence and Bayesian belief statistics
• Sensor fusion for activity recognition in smart environments
• Multiple classification systems
• Ensemble learning for text mining and sentiment analysis

Involved in the project:

Identifying cyberbullying from social media

What is cyberbullying?

The use of cell phones, instant messaging, e-mail, chat rooms or social networking sites such as Facebook and Twitter to harass, threaten or intimidate someone

Is cyberbullying a real issue?

Cyberbullying can severly affect mental and physical health and, in extreme cases, lead to suicidal behavior.

The cyberbullying project

Contributions

• Define search queries to seek potential bullying messages in Twitter
• Perform recurrent searches to Twitter using its APIs
• Connect to the existing sentiment analysis engine
• Develop a web-based application to:
  • Manage existing queries and create new ones
  • Manually review the potential bullying messages
  • Visualize the data to extract valuable information

This thesis proposed and investigated multiple Cloud architectures and models and implemented them using dierent techniques and scenarios, such as queueing theory, nonlinear programming and Cloud simulation.

A Cloud infrastructure was developed using OpenStack.

A scheduling mechanism called GreenC was developed to achieve energy eciency while maintaining the performance of the Cloud system. This policy obtained a reduction in energy consumption of up to 23% compared with the default policy

Two eHealth telemonitoring applications were implemented and deployed on the developed Cloud infrastructure.

Smoking Patient Control (S-PC) was deployed and tested in two hospitals with a total of 229 real patients, where 104 patients were treated using the S-PC tool and 125 without it.

74% of the patients treated using the S-PC tool successfully completed the treatment without relapses, while only 45.6% of those who followed the conventional treatment successfully achieved long-term abstinence.

Hypertension Patient Control (H-PC) was successfully designed and implemented, and the computing boundaries were measured, proving its good behavior in terms of performance and scalability.

These performance results are also directly applicable to the S-PC tool.

What's next?

There are a number of open research challenges to address in order to further advance in these areas.

The eHealth tools will be deployed in more hospitals and healthcare centers.

The Sant Joan de Deu hospital in Manresa has agreed to test the S-PC tool with its patients.

Healthcare centres are being contacted to expand the testing both in the number of patients and the geographic location.

The H-PC tool has entered the testing phase with real patients in three different centers: the Arnau de Vilanova and Santa Maria hospitals in Lleida and the Clinic Hospital in Barcelona.

A smartphone app is being developed for both tools.

Explore additional treatments and diseases.

Expand the number of healthcare centers, patients and treatments.

Additional performance and scalability tests to avoid performance issues.

Analysis of social networks

Internet of Things

Big Data techniques applied to eHealth

Thank you for your kind attention!