Computer Science & Engineering
http://hdl.handle.net/10576/3093
2024-03-27T00:46:29ZOptimizing Cloud Virtual Machine Migration: Minimizing Downtime and Migration Time Using Machine Learning
http://hdl.handle.net/10576/51504
Optimizing Cloud Virtual Machine Migration: Minimizing Downtime and Migration Time Using Machine Learning
Haris, Raseena M
Cloud computing has revolutionized the way services are delivered to users, offering unparalleled flexibility and scalability. However, cloud services can become temporarily unavailable due to maintenance, resource allocation, load balancing, cyberattacks, power management, fault tolerance, and various other factors. To ensure a seamless experience for clients during these periods, live virtual machine migration (LVM) emerges as an indispensable choice. LVM involves relocating virtual machines (VMs) from source to destination with minimal disruption to client activities. While the pre-copy method is the most commonly used live migration technique due to its reliability, it faces challenges such as extended downtime and migration time caused by a large number of dirty pages generated in each iteration. To optimize these migration metrics, numerous solutions have been developed. Nevertheless, a recurring issue in most techniques is the use of static stopping conditions. If the dirty rate exceeds the network throughput, the number of dirty pages retransmitted in each iteration increases, and the hypervisor cannot complete the dirty page transfer in a specific iteration. Moreover, there is no universal stopping condition suitable for all VMs. As a result, the source VM needs to be suspended for a longer time to complete the migration, increasing downtime. Extended migration downtime causes service interruptions and affects the performance of running applications. Therefore, optimizing the memory dirty rate to minimize downtime and total migration time is the primary challenge in the pre-copy approach. To address these challenges, we conducted a thorough analysis of the critical factors influencing live migration performance. Subsequently, we devised an algorithm to identify these critical features and leveraged them to build a machine-learning model capable of intelligently predicting the optimal time to transition into the stop and copy phase, reducing reliance on static stopping conditions. Our proposed machine learning method was rigorously evaluated through experiments conducted on a dedicated testbed using KVM/QUEM technology, involving different VM sizes and memory-intensive workloads. A comparative analysis against proposed pre-copy methods and existing techniques reveals a remarkable improvement, with an average 61.24% reduction in downtime for different RAM configurations in high-write-intensive workloads, along with an average reduction in total migration time of approximately 85.81%. Furthermore, we examined the security concerns surrounding live migration, particularly in domains handling critical applications such as banking, healthcare, etc. Many organizations
in these sectors are hesitant to employ live migration due to security risks. To address this, we introduced a selective encryption approach for protecting sensitive information during migration. Our experimental results highlight that the selective encryption method enhances our proposed machine learning model to reduce downtime and total migration time while preserving the privacy of sensitive data.
2024-01-01T00:00:00ZPhenotaxis: Localization of The Source of Phenomena Using Mobile Searchers
http://hdl.handle.net/10576/51502
Phenotaxis: Localization of The Source of Phenomena Using Mobile Searchers
Abughanam, Nada Abdelelah Nazmi
Over the years, the use of robotic searchers in times of disaster and dangerous incidents,
such as toxic gas leakage, has become of higher significance. Using robot searchers
instead of humans or animals in such dangerous situations significantly reduces risks.
Several algorithms have been developed for robot searchers to search for the source
of incident by following the emitted odor via Odor Source Localization (OSL). In this
thesis, we investigate the performance of several gradient-based and bio-inspired OSL
algorithms in a turbulent environment, where the performance is greatly impacted due
to reliance on the odor concentration. Additionally, we evaluate the performance of
multiple cooperating searchers performing OSL through communicating via an error
channel and observe how the error can affect the search, showing that high error in the
channel can mislead the search. Additionally, a Reinforcement Learning (RL)-based
approach to OSL is proposed, which shows an improvement in the search efficiency.
Further, the potential of a searcher that is assisted by aWireless Sensor Network (WSN)
is studied, where two cooperation strategies between the searcher and the WSN are
proposed, showing potential in improving the performance of a single searcher up to the
performance of multiple searchers.
2024-01-01T00:00:00ZEnhancing Knowledge Distillation for Text Summarization
http://hdl.handle.net/10576/51500
Enhancing Knowledge Distillation for Text Summarization
Kotit, Mohammad Basheer
In the realm of natural language processing, recent advancements have been significantly
shaped by the development of large pretrained Seq2Seq Transformer models,
including BART, PEGASUS, and T5. These models have revolutionized various text
generation applications, such as machine translation, text summarization, and chatbot
development, by offering remarkable improvements in accuracy and fluency.
However, their deployment in text summarization often encounters significant challenges
in environments with limited computational resources. This research proposes
an innovative solution: the development of compact student models. These models are
designed to emulate the capabilities of their larger pretrained counterparts (teacher models)
while ensuring reduced computational demands and increased processing speed,
thus maintaining high performance with greater efficiency.
Knowledge distillation, a popular technique in model optimization, typically employs
two primary techniques: direct knowledge distillation and the use of pseudo-labels. Our
research enhances direct knowledge distillation by introducing an effective behavior
function. This function selectively emphasizes the more certain predictions from the
teacher model, thereby addressing the exposure bias issue that arises from differences
between training and testing environments. In addition to this, we propose a novel
approach to select the most reliable predictions from the teacher model. These highconfidence
predictions are then utilized as pseudo-summaries, optimizing the student
model’s training through the pseudo-label technique. This dual approach mainly focuses
on the confidence of teacher predictions and offers a comprehensive solution to enhance
the model’s performance while maintaining computational efficiency.
We evaluated our methods using BART on the CNN/DM dataset and Pegasus on
the XSUM dataset. The findings of these assessments revealed that our approaches not
only successfully achieved the knowledge distillation objectives, but also significantly
surpassed the performance of the teacher models.
2024-01-01T00:00:00ZEnhancing Autonomous Robot Perception Via Slam Coupled With AI-Driven Selective Obstacle Object Detection
http://hdl.handle.net/10576/51496
Enhancing Autonomous Robot Perception Via Slam Coupled With AI-Driven Selective Obstacle Object Detection
Hamad, Layth Kamal
Autonomous mobile robots are changing various industries, making them more efficient
and adaptable. They are used in critical sectors such as manufacturing, healthcare, logistics,
and infrastructure to make operations smoother, increase productivity, and make
the economy more resilient. This is done through Robotics Automation Systems (RAS).
The research presented in this thesis addresses a pivotal problem in the field of mobile
robotics, specifically the limitations of conventional Simultaneous Localization and
Mapping (SLAM) algorithms in accurately recognizing and mapping various types of
obstacles and hazards. Existing methodologies often lack the computational intelligence
to differentiate between physical and non-physical obstacles, leading to suboptimal navigational
decisions and posing safety risks. To rectify these shortcomings, the proposed
solution unfolds in a structured five-stage approach: AI-based object detection, stereo
vision, disparity and depth estimation, dimension estimation, and SLAM integration.
The AI models achieved an accuracy rate of 0.975 for bottle obstacle detection and 0.906
for fire flame identification. Depth and dimension estimation stages employed stereo vision
techniques to attain an accuracy of 94% and 96.7%, respectively, for objects within
a 6-meter radius of the robot. Including a GPS-less capability reduced locational error
to 0.3 meters, outperforming traditional GPS modules with up to 5-meter errors. The
concluding stage integrates these advancements into the SLAM-generated occupancy
grid, enhancing the robot’s environmental mapping and navigational capabilities. The
research successfully proposed an autonomous mobile robot system that can visually
distinguish obstacles, pinpoint their location and dimensions, and deftly navigate around
them, enhancing its operational autonomy and safety in complex environments.
2024-01-01T00:00:00Z