580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can divide-and-conquer paradigms be generalized and optimized for parallel algorithm design?
This research theme explores the formal generalization of the divide-and-conquer (D&C) programming pattern to encompass a broader class of parallel algorithms. It investigates functional specifications, potential optimizations, and equivalences to other classical parallel programming skeletons. Understanding and formalizing D&C parallelization expands its applicability and guides efficient implementation strategies.
Key finding: This paper presents a new formal functional specification generalizing the divide-and-conquer pattern, proving that many classical parallel programming patterns can be derived through parameter instantiation. It provides a... Read more
Key finding: Focusing on a parallel bag-of-tasks propagation algorithm used in image interpolation, this study develops stochastic automata network-based analytical models to predict performance of different parallel implementation... Read more
Key finding: This work examines the parallelization of computationally intensive vision algorithms—specifically data clustering and image matching—on workstation clusters. It presents design and communication strategies using PVM and... Read more
2. What frameworks and algorithmic modifications improve parallel scalability and efficiency in combinatorial and optimization problems?
This theme focuses on algorithmic design and practical parallelization techniques tailored for large, complex combinatorial optimization problems, such as vertex cover, traveling salesman, and genetic algorithm based searches. It includes decomposition strategies, load balancing, parallel tree search methodologies, and hybrid models integrating stochastic and deterministic computations on multicore and cluster architectures.
Key finding: The paper demonstrates a parallel fixed-parameter tractable approach for large-scale vertex cover problems utilizing balanced search space decomposition and kernelization through polynomial-time preprocessing steps. Careful... Read more
Key finding: This study presents a parallel branch-and-bound algorithm optimized for cluster multicore systems to solve the traveling salesman problem exactly. By employing multithreading to concurrently explore subproblems and leveraging... Read more
Key finding: Introducing a fully asynchronous parallel genetic algorithm (PGA) designed for MIMD architectures, the work shows that distributed, intelligent search with localized selection outperforms traditional genetic algorithms with... Read more
3. How can parallel computing architectures and programming models accelerate high-dimensional data processing and machine learning workflows?
This theme investigates the integration of parallel hardware (such as multicore CPUs, GPUs, and distributed clusters) with tailored algorithms and programming models to speed up data-intensive tasks including machine learning, biometric recognition, and large-scale matrix computations. It covers GPU-based acceleration, parallel statistical computing methods, and parallel stochastic gradient descent approaches for recommendation systems.
Key finding: This survey and methodological discussion emphasizes various parallel statistical computing techniques including parallel regression algorithms, parallel bootstrap, and parallel Markov Chain Monte Carlo. It underscores... Read more
Key finding: By implementing a multimodal biometric system combining face and iris recognition on Nvidia CUDA-enabled GPUs, the authors achieve more than threefold speed improvements over CPU-based processing without loss of accuracy. The... Read more
Key finding: This paper presents a momentum-incorporated parallel stochastic gradient descent (MPSGD) algorithm for training latent factor recommendation models on large-scale sparse matrices. The algorithm combines parallel data... Read more
Key finding: Targeting multicore CPU platforms, this work proposes novel parallelization of integer multiplication algorithms through a delayed carry mechanism which decouples loop iterations to enable concurrent partial sums computation.... Read more
Key finding: Demonstrating the use of workstation clusters and message passing (PVM) for parallelizing clustering and image matching, this paper highlights how decomposing data-intensive vision tasks into concurrent subproblems enables... Read more