A5006910309- Cloud Computing and Resource Management
- Privacy-Preserving Technologies in Data
- IoT and Edge/Fog Computing
- Scientific Computing and Data Management
- Distributed and Parallel Computing Systems
- Parallel Computing and Optimization Techniques
- Stochastic Gradient Optimization Techniques
- Data Stream Mining Techniques
- Mobile Crowdsensing and Crowdsourcing
- Cryptography and Data Security
- Internet Traffic Analysis and Secure E-voting
- Stock Market Forecasting Methods
- Advanced Data Storage Technologies
- Machine Learning and Data Classification
- Age of Information Optimization
- Advanced Memory and Neural Computing
- Blockchain Technology Applications and Security
- Auction Theory and Applications
- Manufacturing Process and Optimization
- Smart Grid Energy Management
- Smart Cities and Technologies
- Software System Performance and Reliability
- Opportunistic and Delay-Tolerant Networks
- Big Data and Business Intelligence
- Explainable Artificial Intelligence (XAI)
University of Chicago
2019-2024
University of Illinois Chicago
2020-2024
Université Lille Nord de France
2022
Université Paris Cité
2022
Institute for Forecasting of the Slovak Academy of Sciences
2022
University College Dublin
2022
Vienna University of Economics and Business
2022
Argonne National Laboratory
2022
With increasing connectivity to support digital services in urban areas, there is a realization that demand for offering similar capability rural communities still limited. To unlock the potential of artificial intelligence (AI) within economies, we propose AI—the mobilization serverless computing enable AI austere environments. Inspired by problems observed New Zealand, analyze major challenges agrarian and define their requirements. We demonstrate proof-of-concept system cross-field...
Precision horticulture is evolving due to scalable sensor deployment and machine learning integration. These advancements boost the operational efficiency of individual farms, balancing benefits analytics with autonomy requirements. However, given concerns that affect wide geographic regions (e.g., climate change), there a need apply models span farms. Federated Learning (FL) has emerged as potential solution. FL enables decentralized (ML) across different farms without sharing private data....
Amazon spot instances provide preemptable computing capacity at a cost that is often significantly lower than comparable on-demand or reserved instances. Spot are charged the current price: fluctuating market price based on supply and demand for instance capacity. However, inherently volatile, changes over time, can be revoked by with as little two minutes' warning. Given potential discount---up to 90% in some cases---there has been significant interest scientific cloud community leverage...
Advances in network technologies have greatly decreased barriers to accessing physically distributed computers. This newfound accessibility coincides with increasing hardware specialization, creating exciting new opportunities dispatch workloads the best resource for a specific purpose, rather than those that are closest or most easily accessible. We present Delta, service designed intelligently schedule function-based across set of heterogeneous computing resources. Delta implements an...
Federated learning (FL) is a technique for distributed machine that enables the use of siloed and data. With FL, individual models are trained separately then only model parameters (e.g., weights in neural network) shared aggregated to create global model, allowing data remain its original environment. While many applications can benefit from existing frameworks incomplete, cumbersome, environment-dependent. To address these issues, we present FLoX, an FL framework built on funcX federated...
Federated Learning (FL) is a recent approach for distributed Machine (ML) where data are never communicated to central node. Instead, an ML model (for example, deep neural network) initialized by designated (aggregation) node and shared with training nodes that have direct access of interest. These then perform small batches on their local data. Periodically, each submits parameter/weight updates the The aggregates parameters/weights create new global it re-shares nodes. This process can...
Advances in networks, accelerators, and cloud services encourage programmers to reconsider where compute---such as when fast networks make it cost-effective compute on remote accelerators despite added latency. Workflow cloud-hosted serverless computing frameworks can manage multi-step computations spanning federated collections of cloud, high-performance (HPC), edge systems, but passing data among computational steps via storage incur high costs. Here, we overcome this obstacle with a new...
Deep learning methods are transforming research, enabling new techniques, and ultimately leading to discoveries. As the demand for more capable AI models continues grow, we now entering an era of Trillion Parameter Models (TPM), or with than a trillion parameters---such as Huawei's PanGu-Σ. We describe vision ecosystem TPM users providers that caters specific needs scientific community. then outline significant technical challenges open problems in system design serving TPMs enable research...
Cloud providers continue to expand and diversify their collection of leasable resources meet the needs an increasingly wide range applications. While this flexibility is a key benefit cloud, it also creates complex landscape in which users are faced with many resource choices for given application. Suboptimal selections can both degrade performance increase costs. Given rapidly evolving pool resources, infeasible alone select instance types; instead, automated methods needed simplify guide...
The Amazon Web Services spot market sells excess computing capacity at a reduced price and with reliability guarantees. low cost nature of the has led to widespread adoption in industry science. However, one challenges using is that it intentionally opaque thus users have little understanding underlying dynamics. In late 2017, mechanisms were significantly altered-no longer are bid prices used clear as result pricing much less volatile. this paper, we revisit prior work aim analyze...
The variety of instance types available on cloud platforms offers enormous flexibility to match the requirements applications with resources. However, selecting most suitable type and configuring an application optimally execute that can be complicated time-consuming. For example, parallelism flags must cores problem sizes tuned memory. As search space configurations enormous, we propose automated approach, called ParaOpt, automatically explore tune arbitrary instances. ParaOpt supports...
Cloud computing provides on-demand access to computational resources while outsourcing infrastructure and service maintenance. Edge could extend cloud capability areas with limited resources, such as rural areas, by utilizing low-cost hardware, single-board computers. data centre hosted machine learning algorithms may violate user privacy confidentiality requirements. Federated (FL) trains models without sending a central server ensures privacy. Using FL, multiple actors can collaborate on...
Fo s t er, Ia n, S py ri d
Federated Learning (FL) is a decentralized machine learning paradigm where models are trained on distributed devices and aggregated at central server. Existing FL frameworks assume simple two-tier network topologies end directly connected to the aggregation While this practical mental model, it does not exploit inherent topology of real-world systems like Internet-of-Things. We present Flight, novel framework that supports complex hierarchical multi-tier topologies, asynchronous aggregation,...
Recent advances in networking technology and serverless architectures have enabled automated distribution of compute workloads at the function level. As heterogeneity physical computing resources increase, so too does need to effectively use those resources. This is especially true when leveraging multiple form local, distributed, cloud Adding complexity problem different notions "cost" it comes using these Tradeoffs exist due inherent difference between costs computation for end user. For...
Deep learning methods are transforming research, enabling new techniques, and ultimately leading to discoveries. As the demand for more capable AI models continues grow, we now entering an era of Trillion Parameter Models (TPM), or with than a trillion parameters -- such as Huawei's PanGu-$\Sigma$. We describe vision ecosystem TPM users providers that caters specific needs scientific community. then outline significant technical challenges open problems in system design serving TPMs enable...
As the market for cloud computing continues to grow, an increasing number of users are deploying applications as microservices. The shift introduces unique challenges in identifying and addressing performance issues, particularly within large complex infrastructures. To address this challenge, we propose a methodology that unveils temporal deviations microservices by clustering containers based on their characteristics at different time intervals. Showcasing our Alibaba dataset, found both...
Exponentially increasing data volumes, coupled with new modes of analysis have created significant opportunities for scientists. However, the stochastic nature many science techniques results in tradeoffs between costs and accuracy. For example, machine learning algorithms can be trained iteratively indefinitely diminishing returns terms In this paper we explore cost-accuracy tradeoff through three representative examples: vary number models an ensemble, epochs used to train a model, amount...
Federated learning has shown enormous promise as a way of training ML models in distributed environments while reducing communication costs and protecting data privacy. However, the rise complex cyber-physical systems, such Internet-of-Things, presents new challenges that are not met with traditional FL methods. Hierarchical Learning extends process to enable more efficient model aggregation based on application needs or characteristics deployment environment (e.g., resource capabilities...
Advances in networks, accelerators, and cloud services encourage programmers to reconsider where compute -- such as when fast networks make it cost-effective on remote accelerators despite added latency. Workflow cloud-hosted serverless computing frameworks can manage multi-step computations spanning federated collections of cloud, high-performance (HPC), edge systems, but passing data among computational steps via storage incur high costs. Here, we overcome this obstacle with a new...
Federated Learning (FL) is an enabling technology for supporting distributed machine learning across several de-vices on decentralized data. A critical challenge when FL in practice the system resource heterogeneity of worker devices that train ML model locally. workflows can be run diverse computing devices, from sensors to High Performance Computing (HPC) clusters; however, these disparities may result some being too burdened by task training and thus struggle perform robust compared more...