A5003773058- Parallel Computing and Optimization Techniques
- Remote Sensing and LiDAR Applications
- Fire effects on ecosystems
- Cloud Computing and Resource Management
- UAV Applications and Optimization
- Remote Sensing in Agriculture
- Embedded Systems Design Techniques
- Advanced Neural Network Applications
- Real-Time Systems Scheduling
- Green IT and Sustainability
- Stochastic Gradient Optimization Techniques
- Distributed and Parallel Computing Systems
- Distributed systems and fault tolerance
- Advanced Data Storage Technologies
- Advanced Software Engineering Methodologies
- Image Enhancement Techniques
- IoT and Edge/Fog Computing
- Computer Graphics and Visualization Techniques
- Domain Adaptation and Few-Shot Learning
- Software System Performance and Reliability
- Caching and Content Delivery
- Privacy-Preserving Technologies in Data
- Sparse and Compressive Sensing Techniques
- Image and Signal Denoising Methods
- Recommender Systems and Techniques
University of Southern California
2021-2024
University of Chicago
2015-2022
University of Illinois Chicago
2018
Embedded real-time systems must meet timing constraints while minimizing energy consumption. To this end, many optimizations are introduced for specific platforms or applications. These solutions not portable, however, and when the application platform change, these be redesigned. Portable techniques hard to develop due varying tradeoffs experienced with different application/platform configurations. This paper addresses problem of finding exploiting general tradeoffs, using control theory...
Many modern computing systems must provide reliable latency with minimal energy. Two central challenges arise when allocating system resources to meet these conflicting goals: (1) complexity hardware exposes diverse complicated interactions and (2) dynamics be maintained despite unpredictable changes in operating environment or input. Machine learning accurately models the of complex, interacting resources, but does not address dynamics; control theory adjusts dynamic changes, struggles...
The problem of minimizing energy for a performance constraint (e.g., Real-time deadline or quality-of-service requirement) has been widely studied, both in theory and practice. Theoretical models have indicated large potential savings, but practical concerns made these savings hard to realize. Instead, practitioners often rely on heuristic solutions, which achieve good results practice tend be system-specific efficacy. An example is the race-to-idle heuristic, makes all resources available...
Resource scheduling in high performance computing (HPC) usually aims to minimize application runtime rather than optimize for energy efficiency. Most existing research on reducing power and consumption imposes the constraint that little or no loss is allowed, which improves but still does not maximize By optimizing efficiency instead of turnaround time, we can reduce cost running scientific applications. We propose using machine learning classification, driven by low-level hardware counters,...
Deep neural networks with large model sizes achieve state-of-the-art results for tasks in computer vision and natural language processing. However, such models are too compute- or memory-intensive resource-constrained edge devices. Prior works on parallel distributed execution primarily focus training-rather than inference-using homogeneous accelerators data centers. We propose PipeEdge, a framework systems that uses pipeline parallelism to both speed up inference enable running larger, more...
This paper explores the problem of energy optimization in embedded platforms. Specifically, it studies resource allocation strategies for meeting performance constraints with minimal consumption. We present a comparison solutions both homogeneous and single-ISA heterogeneous multi-core systems. demonstrate that different hardware platforms have fundamentally performance/energy tradeoff spaces. As result, minimizing on these requires substantially strategies. Our investigations reveal one...
Dynamic voltage and frequency scaling (DVFS) has been the cornerstone of innumerable software approaches to meeting application timing requirements with minimal energy. However, recent trends in technology-e.g., moving converters on chip-favor hardware control DVFS, as can both react faster external events perform fine-grained power management across a device. We respond these CoPPer, which instead uses capping meet performance high energy efficiency. find that is more challenging than using...
Many modern computing systems must provide reliable latency with minimal energy. Two central challenges arise when allocating system resources to meet these conflicting goals: (1) complexity hardware exposes diverse complicated interactions and (2) dynamics be maintained despite unpredictable changes in operating environment or input. Machine learning accurately models the of complex, interacting resources, but does not address dynamics; control theory adjusts dynamic changes, struggles...
Our software framework, Proteus, treats adaptation as a first-class object, enabling rapid development of robust, adaptive applications. Proteus developers specify their programs' intent and adaptable components (or knobs). A control-theoretic runtime continually monitors the running application, adjusting knobs so that specified is met.
Embedded systems are subject to timing and power constraints. To support both, software currently must integrate multiple tools, resulting in additional complexity. We address this problem with a unified, portable framework called Bard which uses control theory meet the primary constraint linear programming optimize other. evaluate on two embedded platforms that exhibit different performance power/energy characteristics show it achieves less than 2% error meeting constraints while...
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
As energy consumption becomes a first class concern for computing systems, there is an increasing need application-level access to runtime power/energy measurements. To support this need, growing number of power and monitors are being developed, each with their own interfaces. In fact, the approaches extremely diverse, porting energy-aware code new platforms hardware can involve significant rewriting effort. reduce effort portable, monitoring, common interface needed. paper, we propose...
Pipeline parallelism has achieved great success in deploying large-scale transformer models cloud environments, but received less attention edge environments. Unlike scenarios with high-speed and stable network inter-connects, dynamic bandwidth systems can degrade distributed pipeline performance. We address this issue QuantPipe, a communication-efficient system that introduces post-training quantization (PTQ) to compress the communicated tensors. QuantPipe uses adaptive PTQ change bitwidths...
Deep neural networks with large model sizes achieve state-of-the-art results for tasks in computer vision (CV) and natural language processing (NLP). However, these large-scale models are too compute- or memory-intensive resource-constrained edge devices. Prior works on parallel distributed execution primarily focus training -- rather than inference using homogeneous accelerators data centers. We propose EdgePipe, a framework systems that uses pipeline parallelism to both speed up enable...
Many modern software applications have performance requirements, like mobile and embedded systems that must keep up with sensor data, or web services return results to users within an acceptable latency bound. For such applications, the goal is not run as fast possible, but meet their requirements minimal resource usage, key in most being energy. Heuristic solutions been proposed minimize energy under a constraint, recent studies show these approaches are portable - heuristics near-optimal...
Deep learning is commonly used to make personalized recommendations users for a wide variety of activities. However, deep recommendation model (DLRM) training increasingly dominated by all-to-all and many-to-many communication patterns. While there are algorithms efficiently overlap computation many collective operations, these patterns strictly limited network bottlenecks. We propose co-designing DLRM with the recently proposed Opera network, which designed avoid multiple hops using...
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.
Wildfire perimeter mapping currently relies on deferred processing of data from manned and orbital platforms using hand-tuned physics-based models. We demonstrate real-time on-board multispectral cost-efficient unmanned aerial ML-based semantic segmentation.