Mohsen Amini Salehi

Ali Mokhtari, Md Abir Hossen, Pooyan Jamshidi et al.

Edge computing enables smart IoT-based systems via concurrent and continuous execution of latency-sensitive machine learning (ML) applications. These edge-based machine learning systems are often battery-powered (i.e., energy-limited). They use heterogeneous resources with diverse computing performance (e.g., CPU, GPU, and/or FPGAs) to fulfill the latency constraints of ML applications. The challenge is to allocate user requests for different ML applications on the Heterogeneous Edge Computing Systems (HEC) with respect to both the energy and latency constraints of these systems. To this end, we study and analyze resource allocation solutions that can increase the on-time task completion rate while considering the energy constraint. Importantly, we investigate edge-friendly (lightweight) multi-objective mapping heuristics that do not become biased toward a particular application type to achieve the objectives; instead, the heuristics consider "fairness" across the concurrent ML applications in their mapping decisions. Performance evaluations demonstrate that the proposed heuristic outperforms widely-used heuristics in heterogeneous systems in terms of the latency and energy objectives, particularly, at low to moderate request arrival rates. We observed 8.9% improvement in on-time task completion rate and 12.6% in energy-saving without imposing any significant overhead on the edge system.

SM Zobaed, Ali Mokhtari, Jaya Prakash Champati et al.

Smart IoT-based systems often desire continuous execution of multiple latency-sensitive Deep Learning (DL) applications. The edge servers serve as the cornerstone of such IoT-based systems, however, their resource limitations hamper the continuous execution of multiple (multi-tenant) DL applications. The challenge is that, DL applications function based on bulky "neural network (NN) models" that cannot be simultaneously maintained in the limited memory space of the edge. Accordingly, the main contribution of this research is to overcome the memory contention challenge, thereby, meeting the latency constraints of the DL applications without compromising their inference accuracy. We propose an efficient NN model management framework, called Edge-MultiAI, that ushers the NN models of the DL applications into the edge memory such that the degree of multi-tenancy and the number of warm-starts are maximized. Edge-MultiAI leverages NN model compression techniques, such as model quantization, and dynamically loads NN models for DL applications to stimulate multi-tenancy on the edge server. We also devise a model management heuristic for Edge-MultiAI, called iWS-BFE, that functions based on the Bayesian theory to predict the inference requests for multi-tenant applications, and uses it to choose the appropriate NN models for loading, hence, increasing the number of warm-start inferences. We evaluate the efficacy and robustness of Edge-MultiAI under various configurations. The results reveal that Edge-MultiAI can stimulate the degree of multi-tenancy on the edge by at least 2X and increase the number of warm-starts by around 60% without any major loss on the inference accuracy of the applications.

Jacob Bradshaw, Mohsen Riahi Alam, Bhanuja Ainary et al.

Despite advances in assistive technologies, Blind and Low-Vision (BLV) individuals continue to face challenges in understanding their surroundings. Delivering concise, useful, and timely scene descriptions for ambient perception remains a long-standing accessibility problem. To address this, we introduce Audo-Sight, an AI-driven assistive system across Edge-Cloud that enables BLV individuals to perceive their surroundings through voice-based conversational interaction. Audo-Sight employs a set of expert and generic AI agents, each supported by dedicated processing pipelines distributed across edge and cloud. It analyzes user queries by considering urgency and contextual information to infer the user intent and dynamically route each query, along with a scene frame, to the most suitable pipeline. In cases where users require fast responses, the system simultaneously leverages edge and cloud processing pipelines. The edge generates an initial response quickly, while the cloud provides more detailed and accurate information. To overcome the challenge of seamlessly combining these outputs, we introduce the Response Fusion Engine, which fuses the fast edge response with the more accurate cloud output, ensuring timely and high-accuracy response for the BLV users. Systematic evaluation shows that Audo-Sight delivers speech output around 80% faster for urgent tasks and generates complete responses approximately 50% faster across all tasks compared to a commercial cloud-based solution -- highlighting the effectiveness of our system across edge-cloud. Human evaluation of Audo-Sight shows that it is the preferred choice over GPT-5 for 62% of BLV participants with another 23% stating both perform comparably.

Razin Farhan Hussain, Mohsen Amini Salehi

Deep neural network (DNN) models are effective solutions for industry 4.0 applications (\eg oil spill detection, fire detection, anomaly detection). However, training a DNN network model needs a considerable amount of data collected from various sources and transferred to the central cloud server that can be expensive and sensitive to privacy. For instance, in the remote offshore oil field where network connectivity is vulnerable, a federated fog environment can be a potential computing platform. Hence it is feasible to perform computation within the federation. On the contrary, performing a DNN model training using fog systems poses a security issue that the federated learning (FL) technique can resolve. In this case, the new challenge is the class imbalance problem that can be inherited in local data sets and can degrade the performance of the global model. Therefore, FL training needs to be performed considering the class imbalance problem locally. In addition, an efficient technique to select the relevant worker model needs to be adopted at the global level to increase the robustness of the global model. Accordingly, we utilize one of the suitable loss functions addressing the class imbalance in workers at the local level. In addition, we employ a dynamic threshold mechanism with user-defined worker's weight to efficiently select workers for aggregation that improve the global model's robustness. Finally, we perform an extensive empirical evaluation to explore the benefits of our solution and find up to 3-5% performance improvement than baseline federated learning methods.

Chavit Denninnart, Mohsen Amini Salehi

Along with the rise of domain-specific computing (ASICs hardware) and domain-specific programming languages, we envision that the next step is the emergence of domain-specific cloud platforms. Developing such platforms for popular applications in the serverless manner, not only can offer a higher efficiency to both users and providers, it can also expedite the application development cycles and enable users to become solution-oriented and focus on their specific business logic. Considering multimedia streaming as one of the most trendy applications in the IT industry, the goal of this study is to develop SMSE, the first domain-specific serverless platform for multimedia streaming. SMSE democratizes multimedia service development via enabling content providers (or even end-users) to rapidly develop their desired functionalities on their multimedia contents. Upon developing SMSE, the next goal of this study is to deal with its efficiency challenges and develop a function container provisioning method that can efficiently utilize cloud resources and improve the users' QoS. In particular, we develop a dynamic method that provisions durable or ephemeral containers depending on the spatiotemporal and data-dependency characteristics of the functions. Evaluating the prototype implementation of SMSE under real-world settings demonstrates its capability to reduce both the containerization overhead, and the makespan time of serving multimedia processing functions (by up to 30%) in compare to the function provision methods that are being used in the general-purpose serverless cloud systems.

Chavit Denninnart, Mohsen Amini Salehi

Cloud-based computing systems can get oversubscribed due to the budget constraints of their users or limitations in certain resource types. The oversubscription can, in turn, degrade the users perceived Quality of Service (QoS). The approach we investigate to mitigate both the oversubscription and the incurred cost is based on smart reusing of the computation needed to process the service requests (i.e., tasks). We propose a reusing paradigm for the tasks that are waiting for execution. This paradigm can be particularly impactful in serverless platforms where multiple users can request similar services simultaneously. Our motivation is a multimedia streaming engine that processes the media segments in an on-demand manner. We propose a mechanism to identify various types of "mergeable" tasks and aggregate them to improve the QoS and mitigate the incurred cost. We develop novel approaches to determine when and how to perform task aggregation such that the QoS of other tasks is not affected. Evaluation results show that the proposed mechanism can improve the QoS by significantly reducing the percentage of tasks missing their deadlines %. In addition, it can and reduce the overall time (and subsequently the incurred cost) of utilizing cloud services by more than 9%.

Sakib M Zobaed, Mohsen Amini Salehi, Rajkumar Buyya

Cloud-based enterprise search services (e.g., AWS Kendra) have been entrancing big data owners by offering convenient and real-time search solutions to them. However, the problem is that individuals and organizations possessing confidential big data are hesitant to embrace such services due to valid data privacy concerns. In addition, to offer an intelligent search, these services access the user search history that further jeopardizes his/her privacy. To overcome the privacy problem, the main idea of this research is to separate the intelligence aspect of the search from its pattern matching aspect. According to this idea, the search intelligence is provided by an on-premises edge tier and the shared cloud tier only serves as an exhaustive pattern matching search utility. We propose Smartness At Edge (SAED mechanism that offers intelligence in the form of semantic and personalized search at the edge tier while maintaining privacy of the search on the cloud tier. At the edge tier, SAED uses a knowledge-based lexical database to expand the query and cover its semantics. SAED personalizes the search via an RNN model that can learn the user interest. A word embedding model is used to retrieve documents based on their semantic relevance to the search query. SAED is generic and can be plugged into existing enterprise search systems and enable them to offer intelligent and privacy-preserving search without enforcing any change on them. Evaluation results on two enterprise search systems under real settings and verified by human users demonstrate that SAED can improve the relevancy of the retrieved results by on average 24% for plain-text and 75% for encrypted generic datasets.

Sm Zobaed, Md Enamul Haque, Md Fazle Rabby et al.

Word sense disambiguation (WSD) methods identify the most suitable meaning of a word with respect to the usage of that word in a specific context. Neural network-based WSD approaches rely on a sense-annotated corpus since they do not utilize lexical resources. In this study, we utilize both context and related gloss information of a target word to model the semantic relationship between the word and the set of glosses. We propose SensPick, a type of stacked bidirectional Long Short Term Memory (LSTM) network to perform the WSD task. The experimental evaluation demonstrates that SensPick outperforms traditional and state-of-the-art models on most of the benchmark datasets with a relative improvement of 3.5% in F-1 score. While the improvement is not significant, incorporating semantic relationships brings SensPick in the leading position compared to others.

Shangrui Wu, Chavit Denninnart, Xiangbo Li et al.

Serverless clouds allocate multiple tasks (e.g., micro-services) from multiple users on a shared pool of computing resources. This enables serverless cloud providers to reduce their resource usage by transparently aggregate similar tasks of a certain context (e.g., video processing) that share the whole or part of their computation. To this end, it is crucial to know the amount of time-saving achieved by aggregating the tasks. Lack of such knowledge can lead to uninformed merging and scheduling decisions that, in turn, can cause deadline violation of either the merged tasks or other following tasks. Accordingly, in this paper, we study the problem of estimating execution-time saving resulted from merging tasks with the example in the context of video processing. To learn the execution-time saving in different forms of merging, we first establish a set of benchmarking videos and examine a wide variety of video processing tasks -- with and without merging in place. We observed that although merging can save up to 44% in the execution-time, the number of possible merging cases is intractable. Hence, in the second part, we leverage the benchmarking results and develop a method based on Gradient Boosting Decision Tree (GBDT) to estimate the time-saving for any given task merging case. Experimental results show that the method can estimate the time-saving with the error rate of 0.04, measured based on Root Mean Square Error (RMSE).

Xiangbo Li, Mahmoud Darwich, Magdy Bayoumi et al.

Video streaming, in various forms of video on demand (VOD), live, and 360 degree streaming, has grown dramatically during the past few years. In comparison to traditional cable broadcasters whose contents can only be watched on TVs, video streaming is ubiquitous and viewers can flexibly watch the video contents on various devices, ranging from smart-phones to laptops and large TV screens. Such ubiquity and flexibility are enabled by interweaving multiple technologies, such as video compression, cloud computing, content delivery networks, and several other technologies. As video streaming gains more popularity and dominates the Internet traffic, it is essential to understand the way it operates and the interplay of different technologies involved in it. Accordingly, the first goal of this paper is to unveil sophisticated processes to deliver a raw captured video to viewers' devices. In particular, we elaborate on the video encoding, transcoding, packaging, encryption, and delivery processes. We survey recent efforts in academia and industry to enhance these processes. As video streaming industry is increasingly becoming reliant on cloud computing, the second goal of this survey is to explore and survey the ways cloud services are utilized to enable video streaming services. The third goal of the study is to position the undertaken research works in cloud-based video streaming and identify challenges that need to be obviated in future to advance cloud-based video streaming industry to a more flexible and user-centric service.

SM Zobaed, Mohsen Amini Salehi

Cloud-based enterprise search services (e.g., Amazon Kendra) are enchanting to big data owners by providing them with convenient search solutions over their enterprise big datasets. However, individuals and businesses that deal with confidential big data (eg, credential documents) are reluctant to fully embrace such services, due to valid concerns about data privacy. Solutions based on client-side encryption have been explored to mitigate privacy concerns. Nonetheless, such solutions hinder data processing, specifically clustering, which is pivotal in dealing with different forms of big data. For instance, clustering is critical to limit the search space and perform real-time search operations on big datasets. To overcome the hindrance in clustering encrypted big data, we propose privacy-preserving clustering schemes for three forms of unstructured encrypted big datasets, namely static, semi-dynamic, and dynamic datasets. To preserve data privacy, the proposed clustering schemes function based on statistical characteristics of the data and determine (A) the suitable number of clusters and (B) appropriate content for each cluster. Experimental results obtained from evaluating the clustering schemes on three different datasets demonstrate between 30% to 60% improvement on the clusters' coherency compared to other clustering schemes for encrypted data. Employing the clustering schemes in a privacy-preserving enterprise search system decreases its search time by up to 78%, while increases the search accuracy by up to 35%.