| Research & Development |
Cloud Infrastructure Resource Management Cloud Data Centres are increasingly being used as the primary computing and storage platforms for enterprises, and for consumers. Such Data Centres use large amounts of power to run the host servers, and for air conditioning. Reduction in electrical power consumption is an important goal for sustainable operation of such Data Centres. Cloud Data Centres use "host servers" to run applications, "storage" to hold data, and network access for connectivity. The host servers run Virtual Machines (VMs) which include the OS and user applications. Challenges in this problem include dealing with large amounts of log data stored by running host servers, extracting minimal feature sets to predict overloaded host servers, moving VMs with minimal user downtime, and characterizing workloads as computing paradigms change. We research and propose solutions to reduce power consumption in Cloud Data Centres by detecting overloaded host servers, and moving their VMs to underloaded host servers, and in some cases, moving VMs from underloaded host servers and powering-off those underloaded host servers. We also seek efficient ways to sample and collect VM logs and apply machine learning techniques to predict underloaded and overloaded hosts accurately from the log data. We build system models where we optimize the VM migration process, i.e. minimize SLA violations, migration time, number of migrations and at the same time reduce overall energy consumption in the data centres. We are collaborating with Concordia University (Montreal, QC, Canada.) Data Centre Energy Management using Renewable Energy The energy consumption of the Information and Communication Technology (ICT) data centers; telecommunications servers and energy storage systems have been increasing rapidly for the past decade. In order to reduce global warming and conserve depleting sources of fossil fuel, Cistech started to look into green ICT power solutions to supply clean, reliable, and sustainable energy. Data center is conventionally powered from the grid as AC power. However, since most of the electrical components and equipment in a data center, such as batteries for providing back-up power, require DC power, high voltage DC power distribution (i.e., ~ 400V) has been proposed in literature as an alternative power architecture for data center applications to achieve high power conversion efficiency. Typical power electronic interface employed in renewable energy systems with DC power distribution for ICT data centers require a separate grid-connected power converter module to provide uninterruptible power when either the renewable energy source or the backup energy storage device is not capable of providing the required load power. This type of power converter configuration faces the challenges of using high component count and multiple-power converter modules. In this project, we research into new power converter topology for renewable energy system that can reduce the component count while achieving higher efficiency. We are collaborating with York University (Toronto, ON, Canada) for this research. Anomaly Detection on Smart Device Smart devices are increasingly used as the primary computing and internet access platforms. Malware (or malicious software) is designed to disrupt, and give unauthorized access to a system such as a mobile phone. Malware running on mobile phones can be used for information theft including (account information or passwords), excessive battery-drain, unintended charges or denial of service attacks leading to slow network response. Malware running on smart devices is an anomalous behavior. Challenges in this problem include using apps downloaded from non-official download sites, "zero-day" attacks where signature-based malware detection methods are not available, and performance with low false-positives rates (classifying an app as malware when it is not malware.) By tracking changes in smart device power consumption over time ("power time-series") and machine learning algorithms, we detect anomalous behaviour on smart phones. We are using methods such as Change Point detection, Independent Component Analysis, Principal Component Analysis, Cross Correlation to extract features from benign and malicious apps and machine learning techniques such as Support Vector Machine, Logistic Regression, Naive Bayes, Random Forest for classification. We are collaborating with professors at the University of Waterloo (Waterloo, ON, Canada.) Predictive Models for Environmental Monitoring and Assessment Pollutants originating from industrial activities are adversely impacting the environment and the water quality deterioration is a serious concern. Environmental monitoring is essential for the natural resource management and for the success of mitigation plans to assess whether our efforts are having desired effects on environmental quality. Monitoring programs are both logistically difficult and prohibitively expensive. Moreover, historical and/or long-term instrumental records are often incomplete and/or do not cover sufficient timescales to quantify natural trends in environmental variability. Under these circumstances, indirect (proxy) approaches must be developed. Biological indicator based water quality monitoring offers unique potential for tracking long-term environmental trends. However, the magnitude, frequency and duration of environmental stresses on water bodies are poorly understood across spatiotemporal scales and are differently engaged among regions. Cistech aims to develop and address these critical limitations by:
|