| Session Title: |
Challenging Research Problems in Huawei Theory Lab |
| Date: | 3 Nov 2022 |
| Time: | 10:30-12:30 |
| Venue: | Room S228, Hong Kong Convention and Exhibition Centre |
| Session Chair: | Dr. Felix Lau |
Topic: |
Several Challenging Practical and Theoretical Problems in Large Scale Optimizations |
Speaker: |
Dr Fan Zhang, Theory Lab, Huawei |
Abstract:
Abstract: Large scale optimization has played an important role in practical information and communication technology (ICT) scenarios. Even though most of the practical optimizations are linear programming or mixed integer programming, the problem size could be over one million to even one billion, which makes such optimizations very challenging under the stringent requirement of solution time. In this talk, practical ICT scenarios covering optimizations in data communication and optical networks will be covered and several theoretical and practical problems will be introduced. The problems are from the following three perspectives: routing optimization under practical limitations, AI-aided optimization, and exploration of MILP lower bound. This talk serves as a platform to announce these open questions to public and we welcome anyone who is interested to work with us to conquer the challenges.
Biography:
Fan Zhang (Michael) is currently with Hong Kong Theory Lab, Huawei Hong Kong Research Center as an optimization expert. He received the BEng (first class Hons) degree from Chu Kochen Honors College, Zhejiang University (ZJU), in 2010, and the Ph.D. degree from the Hong Kong University of Science and Technology (HKUST), in 2015. He joined Huawei Future Network Theory lab as a researcher in 2015 and is an optimization expert in Theory Lab since 2022. His research interest covers a wide range including large-scale optimization techniques, operational research with applications to networking problems, low complexity stochastic optimization, and networked control theory. Michael is now leading a diversified team with people from different backgrounds working on challenging and hardcore optimization problems originated from practical scenarios such as wide area network, optical network, wireless network, and data storage systems. He and his team has gained great experiences in designing customized algorithms for efficiently solving practical large-scale LPs/ILPs with number of variables or constraints to be over 1 million to billion. Some research outputs directly translate to solving practical problems, leading to increasing revenue or reducing costs, while some contribute to the development of Huawei’s generic solver (OptVerse) to enhance its performance.
Topic: |
Future challenges for coding |
Speaker: |
Dr Ting-Yi Wu, Theory Lab, Huawei |
Abstract:
In the history of ICT, both source coding and channel coding play crucial roles in boosting technology. Source coding saves a huge amount of storage and communication costs, which enables an efficient data representation without loss of information (physically or semantically); Channel coding adds redundancy to protect data from noise, which enables reliable data storage and communication. However, there are very few fundamental breakthroughs in coding since 2000. Especially for source coding, nowadays compression greatly relies on the techniques discovered before 1980. Due to the rapid evolution of technology, revolutionary coding technologies to realize efficient and reliable communication/storage are urgent demands. In this talk, challenges for modern coding technologies are introduced and some plausible research directions are suggested from our perspective.
Biography:
Ting-Yi Wu received a Ph.D. degree in communication engineering from National Chiao Tung University, Hsinchu, Taiwan, in 2013, where he worked with the Network Technology Lab. as a Postdoctoral Fellow until 2015. I then joined the Signal-Processing and Communication Lab., HKUST, Coordinated Science Lab., UIUC, Illinois, and the School of Electronics and Communication Engineering, SYSU, Guangzhou. Now Ting-Yi Wu is a principal engineer at Theory Lab, Central Research Institute, 2012 Labs, Huawei, who mainly focuses on erasure coding for storage, network, and communication.
Topic: |
An Unified Programming Framework Towards XPU |
Speaker: |
Dr Xin Yao, Theory Lab, Huawei |
Abstract:
With the development of domain-specific processors, it is a big challenge to scale the user code to a large heterogeneous cluster with zero change. This talk will introduce a unified programming framework to address that challenge. It first expresses the used applications into a computational graph. This graph describes a directed dataflow among operators. The weight of each edge stands for the size of data traffic between operators. We propose a graph compiler to divide the graph into several subgraphs, each of which is arranged on a machine for execution. This graph compiler minimizes the total communication cost and improves the parallelism degree when running the whole graph. Then, we deploy a runtime scheduler to address the workloads imbalance at runtime. We further design three communication primitives and offload specific computation logic to the smart network devices, e.g., DPU. In-network computing reduces the data traffic and optimizes execution time via ACSI acceleration. Finally, we provide DPP language, which could describe computation, communication, and storage in a unified model. With the backend compilers, the system will translate the model to platform-specific executable codes.
Biography:
Xin Yao received his PhD degree from the Department of Computer Science, The University of Hong Kong, Hong Kong, China, in 2020. He is currently working as a researcher in Theory Lab, Hong Kong Research Center, Huawei. His research interests include distributed AI system, database system, in-network computation, and fault tolerance.