Svetozar Margenov
Professor
Bulgarian Academy of Sciences
Speech Title:Performance Analysis of Modern Supercomputers, Problem-Oriented Metrics, and Applications
Abstract:Modern computer technologies provide more and more computing power. At the same time, high-performance computing (HPC) systems are becoming more widespread. Let us note the following example, with an increasing impact on the daily life of each of us. The availability of Big Data together with artificial intelligence (AI) tools and the rapid growth of affordable supercomputing resources have posed new challenges to language and content technologies aimed at building ever more innovative products and services. The second example is a little different. Quantum computing is considered one of the most promising technological directions of strategic importance. Although they are still at an early stage of development, quantum computing systems are already available. The synergy between advanced e-infrastructure and AI in a Big Data environment creates qualitatively new opportunities for science and engineering.
The first part of the talk discusses the current development of the computing power of HPC systems. The LINPACK benchmark has been used for more than 30 years to build the TOP500 list ranking the world's most powerful supercomputers. Although the computational linear algebra forms a backbone of the whole computing, the real life HPC performance is rather different than the one achieved on any single benchmark. This necessitates additional application-oriented approaches and metrics to HPC performance evaluation. Here we analyze data from the last two editions of the TOP500. Trends in the development of HPC architectures from the point of view of AI applications are identified. In addition, the increasingly important topic of energy efficiency is discussed. In this context, the capabilities of the petascale supercomputer HEMUS are also discussed. HEMUS is a part of the Infrastructure Complex for Digital Transformation and Large-scale Computing build in the Institute of Information and Communication Technologies of the Bulgarian Academy of Sciences.
In the second part of the presentation, we look at three case studies illustrating the concept of synergy between HPC performance and the scalability of modern methods and algorithms in a Big Data environment. Our developed BURA (Best Uniform Rational Approximation) method has a fundamental role in a wide range of computational models where the process does not satisfy the Brownian motion hypothesis, thus exhibiting anomalous (fractional) diffusion. BURA has several levels of parallelism that enable efficient use of the hierarchical architecture of modern supercomputers. The second case study discusses a biomedical application showing the necessity of supercomputing performance. The process of radiofrequency hepatic tumor ablation is modelled to optimize the parameters of the electrosurgical equipment. The problem is inherently 3D in space, nonlinear, multiscale and multiphysics, with the computer model reaching O(1012) degrees of freedom. The third but equally important example concerns the improvement of scalability and accuracy of quantum annealing computations. The advantages of the developed algorithms are confirmed by results of numerical experiments on a D-Wave 2000 Q system with more than 5000 qubits. The improvement of quantum algorithms often requires the application of fundamentally new approaches. One may be surprised by the interpretation of the substantially changed concepts of parallelism in the proposed parallel quantum annealing.
At the end, concluding remarks are given, including some open problems.
Biography: Svetozar Margenov is a corresponding member of the Bulgarian Academy of Sciences (BAS) and a professor in Computational Mathematics. He graduated from the Faculty of Mathematics and Informatics of the Sofia University “St. Kliment Ohridski”, a PhD and a Doctor of Science form the BAS. The main scientific achievements of prof. Margenov are in the field of information technologies including computational complexity, numerical methods for partial differential equations, numerical solution of fractional diffusion problems, multilevel iterative methods for ill-conditioned large-scale linear systems, parallel methods and algorithms, and supercomputer applications. He has published over 200 scientific papers and two monographs. Svetozar Margenov is the director of the Institute of Information and Communication Technologies (IICT-BAS) and head of the Department of Scientific Computing with Laboratory of 3D Digitization and Microstructure Analysis. He is a co-founder of the Bulgarian section of SIAM and its chairman for the first two mandates. Prof. Margenov has made a great contribution to the development of the modern Bulgarian HPC infrastructure and a high level of expertise in HPC applications. He currently leads the Center of Excellence in Informatics and ICT, within which the last Bulgarian petascale supercomputer HEMUS was opened in 2023. Prof. Svetozar Margenov is awarded the highest distinction of the Bulgarian Academy of Sciences - The Badge of Honor “Marin Drinov” on a Ribbon. He is recipient of the 2024 Walter Scott Award in Information Technology and has also been elected as a Fellow of the International Academy of Information Technology and Quantity Management.
Chang-Yun Lin
Professor
National Chung Hsing University
Speech Title:Multiresponse surface methodology for hyperparameter tuning to optimize multiple performance measures of statistical and machine learning algorithms
Abstract: Hyperparameter tuning is an important task in machine learning for controlling model complexity and improving prediction performance. Most methods in the literature can only be used to tune hyperparameters to optimize a single performance measure. In practice, participants may want to optimize multiple measures of the model; however, optimizing one measure may worsen another. Therefore, a hyperparameter tuning method is proposed using the multiresponse surface methodology to solve the tradeoff problem between multiple measures. A search algorithm that requires fewer tuning runs is developed to find the optimal hyperparameter settings systematically based on the preferences of participants for multiple measures of the model. An example using the random forest algorithm is provided to demonstrate the application of the proposed method to improve the prediction performance of the model on unbalanced data.
Huawei Huang
Associate Professor
Sun Yat-Sen University, China
ORCID: https://orcid.org/0000-0002-7035-6446
Speech Title:BrokerFi, Broker2Earn, and BlockEmulator: Building the Ecosystem of BrokerChain
Abstract: This talk presents the ecosystem built on top of BrokerChain, which is a sharding blockchain developed by Prof. Huawei Huang’s research group (i.e., HuangLab: http://xintelligence.pro/ ).
BrokerChain was originally designed as a cross-shard transaction protocol, and was published in INFOCOM 2022. Starting from BrokerChain protocol, HuangLab members have kept building a sharding blockchain, also named BrokerChain. To keep improving the performance of BrokerChain, we also proposed Broker2Earn protocol (published in INFOCOM 2024), which serves as an incentive mechanism for BrokerChain because it can help recruit enough broker accounts and offer token-liquidity service for other BrokerChain clients. We then developed a DeFi product, named BrokerFi, by integrating the blockchain BrokerChain and the Broker2Earn protocol. Via participating in BrokerFi, brokers can earn native revenues when they stake their idle tokens in the Broker2Earn protocol. On the other hand, Broker2Earn can also benefit the sharded blockchain BrokerChain because it can efficiently spend each staked token on improving BrokerChain’s performance, such as TPS (transaction per second), TCL (transaction confirmation latency), and the queueing length in transaction pools. In the near future, we shall try our best to promote BrokerFi as a real DeFi product in the Hong Kong Web3 market.
BlockEmulator is a blockchain experimental tool open-sourced by HuangLab. Its homepage is https://www.blockemulator.com/. We are currently building a DAO for BlockEmulator on top of BrokerChain, aiming to incentivize blockchain researchers to contribute to and participate into the ecosystem of BlockEmulator.
Biography: Huawei Huang earned his Ph.D. in Computer Science and Engineering from the University of Aizu, Japan. He is currently an Associate Professor at Sun Yat-Sen University, China. His research interests include Blockchains, Web3, and DeFi protocol. He has served as a JSPS research fellow, a visiting scholar at Hong Kong Polytechnic University, and an Assistant Professor at Kyoto University, Japan (日本京都大学). He is a senior member of IEEE. He has been recognized as the world-ranking Top 2% Scientist since 2023. Lab page: xintelligence.pro
Google scholar: https://scholar.google.com/citations?user=DlKPajcAAAAJ&hl=en
Minxian XU
Associate Professor
Chinese Academy of Sciences
ORCID: https://orcid.org/0000-0002-0046-5153
Speech Title:Cloud-native System Support for Efficient Large Language Models Inference Serving
Abstract:Large Language Models (LLMs) are revolutionizing artificial intelligence by enabling advanced natural language processing applications. However, the computational demands of LLM inference require scalable, efficient, and robust system support. This talk investigates cloud-native techniques to optimize LLM inference serving. Specifically, it explores (1) batching inference requests to optimize key-value (KV) cache utilization for enhanced GPU performance and throughput, (2) leveraging containerization for fine-grained resource control to achieve lightweight scheduling and replication, and (3) dynamically balancing workloads through adaptive resource allocation and scaling. Collaborating with industry leaders, this research will integrate first-hand workload data into system design and validation. The outcomes aim to advance system-level support for LLMs, contributing to improved performance and reduced costs for cloud-native AI services.
Biography:Dr. Minxian Xu is an Associate Professor at Shenzhen Institute of Advanced Technology (SIAT), Chinese Academy of Sciences. He received his Ph.D. in Computing and Information Systems from the University of Melbourne in 2019. Dr. Xu has authored over 70 research papers in leading journals and conferences, including CSUR, TSC, TMC, TAAS, and TOIT. His work has garnered more than 3,800 citations (Google Scholar), which focuses on the efficient management of distributed and cloud computing systems, with expertise in areas such as load balancing and energy efficiency for cloud platforms, microservice management in cloud-native environments. Dr. Xu's research achievements include winning the 2019 IEEE TCSC Outstanding PhD Dissertation Award and the 2023 IEEE TCSC Award for Excellence (Early Career Award). He has also been recognized as one of the World’s Top 2% Scientists in 2023 and 2024 by Stanford University. He is a Senior Member of both IEEE and CCF.
Google scholar:https://scholar.google.com/citations?user=KeJL7gEAAAAJ&hl=zh-CN&oi=ao
Ahmad Ali
Dr.
Shanghai Jiao Tong University
Speech Title:Exploiting Dynamic Spatio-Temporal Graph Convolutional Neural Networks for Citywide Traffic Flow Prediction
Abstract:In this talk, I will present recent advancements in using spatio-temporal graph convolutional neural networks for predicting citywide traffic flows. These models capture both spatial dependencies across road networks and temporal patterns of traffic dynamics, enabling accurate forecasting of traffic congestion. I will discuss key challenges, model design, and real-world applications that can significantly improve traffic management in smart cities.
Biography:Dr. Ahmad Ali completed his PhD at Shanghai Jiao Tong University. He is currently a postdoctoral researcher at Shenzhen University. Dr. Ali has reviewed over 1150 articles for well-reputed journals, including Applied Soft Computing, Information Fusion, Information Sciences, Neural Networks, Applied Intelligence, and IEEE Transactions on Industrial Informatics.