21 February 2024

Abstract

Numerical software is being reinvented to provide opportunities to tune dynamically the accuracy of computation to the requirements of the application, resulting in savings of memory, time, and energy.  Floating point computation in science and engineering has a history of “oversolving” relative to expectations for many models. So often are real datatypes defaulted to double precision that GPUs did not gain wide acceptance until they provided in hardware operations not required in their original domain of graphics.  Computational science is now reverting to employ lower precision arithmetic where possible. Many matrix operations allow for lower precision considered at a blockwise level without loss of accuracy, adapting to the magnitude of the norm of the block. Furthermore, many blocks can be approximated with low-rank near equivalents to a prescribed accuracy, adapting to the smoothness of the coefficients of the block.  This leads to smaller memory footprint, which implies higher residency on memory hierarchies, leading in turn to less time and energy spent on data copying, which may even dwarf the savings from fewer and cheaper flops.  We provide examples from several application domains, including Gordon Bell Prize-nominated research in 2022 in environmental statistics and in 2023 in seismic processing.

Bio

David Keyes directs the Extreme Computing Research Center at the King Abdullah University of Science and Technology (KAUST), where he was a founding Dean in 2009 and currently serves in the Office of the President. He is a professor in the programs of Applied Mathematics, Computer Science, and Mechanical Engineering. He is also an Adjunct Professor of Applied Mathematics and Applied Physics at Columbia University, where he formerly held the Fu Foundation Chair. He works at the interface between parallel computing and PDEs and statistics, with a focus on scalable algorithms that exploit data sparsity. Before joining KAUST, Keyes led multi-institutional scalable solver software projects in the SciDAC and ASCI programs of the US Department of Energy (DoE), ran university collaboration programs at US DoE and NASA institutes, and taught at Columbia, Old Dominion, and Yale Universities. He is a Fellow of SIAM, the AMS, and the AAAS. He has been awarded the Gordon Bell Prize from the ACM, the Sidney Fernbach Award from the IEEE Computer Society, and the SIAM Prize for Distinguished Service to the Profession. He earned a B.S.E. in Aerospace and Mechanical Sciences from Princeton in 1978 and a Ph.D. in Applied Mathematics from Harvard in 1984.

9 May 2025

Abstract

As computing power demands continue to grow, achieving energy efficiency in high-performance systems has become a key challenge. One of the most promising software techniques for energy efficiency is Dynamic Voltage and Frequency Scaling (DVFS) which optimize the energy-performance trade-off by changing hardware frequencies. 

This presentation introduces two complementary approaches that advance the state-of-the-art in energy-efficient heterogeneous computing through fine-grained and phase-aware frequency tuning.

The first approach, SYnergy, leverages a novel compiler- and runtime-integrated methodology built upon the SYCL programming model to enable fine-grained frequency scaling on heterogeneous hardware. SYnergy allows developers to specify energy goals for each individual kernel such as minimizing Energy-Delay Product (EDP) or achieving predefined energy-performance tradeoffs. Through compiler integration and a machine learning model, the frequency of each kernel is statically optimized based on the specific energy goal. To extend this fine-grained control to large-scale systems, SYnergy includes a custom SLURM plugin that enables execution across all available devices in a cluster, ensuring scalable energy savings.

While fine-grained frequency scaling at the kernel level can significantly improve energy efficiency, it also introduces overhead due to frequent frequency changes—an overhead that can, in some cases, outweigh the potential benefits. To address this, we propose a novel Phase-aware method that detects different phases through application profiling and DAG analysis and sets an optimal frequency for each phase. Our methodology also considers MPI programs, where the overhead can be hidden by overlapping frequency-change with communication. 

Bio

Lorenzo Carpentieri received his master’s degrees from the University of Salerno, Italy in 2022. He is now a PhD student in the Department of Computer Science at University of Salerno, Italy, under the supervision of Prof. Biagio Cosenza. His research interests include high-performance computing, compiler technology, and programming models having a particular interest in energy efficient and approximate computing.

28 August 2025

Abstract

Constraint programming is a declarative way of solving hard combinatorial, scheduling, resource allocation, and logistics problems. We specify a problem in a high-level language, give it to a solver, and the solver thinks for a while and then gives us the optimal answer. Unfortunately, even the best commercial and academic solvers contain bugs, and will occasionally give a wrong answer, potentially with devastating effects. One way of avoiding this situation is through proof logging, where solvers are modified to output a mathematical proof of correctness alongside their solution. This proof can then be independently audited by a very simple (and potentially even formally verified) proof checking tool, giving us complete confidence in the correctness of solutions (although not the solvers themselves). I’ll explain how proof logging works in general, and give an overview of the challenges and fun involved in bringing it to constraint programming. Ultimately, the aim here is to make algorithms something people can trust with their lives and livelihoods, just as engineers have already done with bridges, planes, and lifts.

Bio

Ciaran McCreesh is a Royal Academy of Engineering Research Fellow working in the Formal Analysis, Theory and Algorithms group in the School of Computing Science at the University of Glasgow. His research looks at practical parallel algorithms, particularly in relation to hard subgraph problems. His publications cover combinatorial search, parallel algorithms, and constraint programming.