Challenges in Automatic Software Optimisation: the energy efficiency case

Scientific Conference (2023)

ith the advent of the Exascale capability allowing supercomputers to perform at least 1018 IEEE 754 Double Precision (64 bits) operations per second, many concerns have been raised regarding the energy ... [more ▼]

ith the advent of the Exascale capability allowing supercomputers to perform at least 1018 IEEE 754 Double Precision (64 bits) operations per second, many concerns have been raised regarding the energy consumption of high-performance computing code. Recently, Frontier operated by the Oak Ridge National Laboratory, has become the first supercomputer to break the exascale barrier [1]. In total, Frontier contains 9,408 CPUs, 37,632 GPUs, and 8,730,112 cores. This world-leading supercomputer consumes about 21 megawatts which is truly remarkable as Frontier was also ranked first on the Green500 list before being recently replaced. The previous top Green500 machine, MN-3 in Japan, provided 39.38 gigaflops per watt, while the Frontier delivered 62.68 gigaflops per watt. All these infrastructure and hardware improvements are just the tip of the Iceberg. Energy-aware code is now required to minimize the energy consumption of distributed and/or multi-threaded software. For example, the data movement bottleneck is responsible for 35 − 60% of a system’s energy consumption during intra-node communication. In an HPC environment, additional energy is consumed through inter-node communication. This position paper aims to introduce future research directions to enter now in the age of energy-aware software. The paper is organized as follows. First, we introduce related works regarding measurement and energy optimisation. Then we propose to focus on the two different levels of granularity in energy optimisation. [less ▲]

Management of an Academic HPC Research Computing Facility: The ULHPC Experience 2.0

in 6th High Performance Computing and Cluster Technologies Conference (HPCCT 2022) (2022, July)

With the advent of the technological revolution and the digital transformation that made all scientific disciplines becoming computational, the need for High Performance Computing (HPC) has become and a ... [more ▼]

With the advent of the technological revolution and the digital transformation that made all scientific disciplines becoming computational, the need for High Performance Computing (HPC) has become and a strategic and critical asset to leverage new research and business in all domains requiring computing and storage performance. Since 2007, the University of Luxembourg operates a large academic HPC facility which remains the reference implementation within the country. This paper provides a general description of the current platform implementation as well as its operational management choices which have been adapted to the integration of a new liquid-cooled supercomputer, named Aion, released in 2021. The administration of a HPC facility to provide state-of-art computing systems, storage and software is indeed a complex and dynamic enterprise with the soul purpose to offer an enhanced user experience for intensive research computing and large-scale analytic workflows. Most design choices and feedback described in this work have been motivated by several years of experience in addressing in a flexible and convenient way the heterogeneous needs inherent to an academic environment towards research excellence. The different layers and stacks used within the operated facilities are reviewed, in particular with regards the user software management, or the adaptation of the Slurm Resource and Job Management System (RJMS) configuration with novel incentives mechanisms. In practice, the described and implemented environment brought concrete and measurable improvements with regards the platform utilization (+12,64%), jobs efficiency (average Wall-time Request Accuracy improved by 110,81%), the management and funding (increased by 10%). Thorough performance evaluation of the facility is also presented in this paper through reference benchmarks such as HPL, HPCG, Graph500, IOR or IO500. It reveals sustainable and scalable performance comparable to the most powerful supercomputers in the world, including for energy-efficient metrics (for instance, 5,19 GFlops/W (resp. 6,14 MTEPS/W) were demonstrated for full HPL (resp. Graph500) runs across all Aion nodes). [less ▲]

A RNN-Based Hyper-Heuristic for Combinatorial Problems

in A RNN-Based Hyper-Heuristic for Combinatorial Problems (2022)

Designing efficient heuristics is a laborious and tedious task that generally requires a full understanding and knowledge of a given optimization problem. Hyper-heuristics have been mainly introduced to ... [more ▼]

Designing efficient heuristics is a laborious and tedious task that generally requires a full understanding and knowledge of a given optimization problem. Hyper-heuristics have been mainly introduced to tackle this issue and are mostly relying on Genetic Programming and its variants. Many attempts in the literature have shown that an automatic training mechanism for heuristic learning is possible and can challenge human-based heuristics in terms of gap to optimality. In this work, we introduce a novel approach based on a recent work on Deep Symbolic Regression. We demonstrate that scoring functions can be trained using Recurrent Neural Networks to tackle a well-know combinatorial problem, i.e., the Multi-dimensional Knapsack. Experiments have been conducted on instances from the OR-Library and results show that the proposed modus operandi is an alternative and promising approach to human- based heuristics and classical heuristic generation approaches. [less ▲]

Proximal Policy Optimisation for a Private Equity Recommitment System

in Springer CCIS series (2021)

Recommitments are essential for limited partner investors to maintain a target exposure to private equity. However, recommitting to new funds is irrevocable and expose investors to cashflow uncertainty ... [more ▼]

Recommitments are essential for limited partner investors to maintain a target exposure to private equity. However, recommitting to new funds is irrevocable and expose investors to cashflow uncertainty and illiquidity. Maintaining a specific target allocation is therefore a tedious and critical task. Unfortunately, recommitment strategies are still manually designed and few works in the literature have endeavored to develop a recommitment system balancing opportunity cost and risk of default. Due to its strong similarities to a control system, we propose to “learn how to recommit” with Reinforcement Learning (RL) and, more specifically, using Proximal Policy Optimisation (PPO). To the best of our knowledge, this is the first attempt a RL algorithm is applied to private equity with the aim to solve the recommitment problematic. After training the RL model on simulated portfolios, the resulting recommitment policy is compared to state-of-the-art strategies. Numerical results suggest that the trained policy can achieve high target allocation while bounding the risk of being overinvested. [less ▲]

Predicting near-optimal skin distance in Verlet buffer approach for Discrete Element Method

in 10th IEEE Workshop on Parallel / Distributed Combinatorics and Optimization (2020, June)

The Verlet list method is a well-known bookkeeping technique of the interaction list used both in Molecular Dynamic (MD) and Discrete Element Method (DEM). The Verlet buffer technique is an enhancement of ... [more ▼]

The Verlet list method is a well-known bookkeeping technique of the interaction list used both in Molecular Dynamic (MD) and Discrete Element Method (DEM). The Verlet buffer technique is an enhancement of the Verlet list that consists of extending the interaction radius of each particle by an extra margin to take into account more particles in the interaction list. The extra margin is based on the local flow regime of each particle to account for the different flow regimes that can coexist in the domain. However, the choice of the near-optimal extra margin (which ensures the best performance) for each particle and the related parameters remains unexplored in DEM unlike in MD. In this study, we demonstrate that the near-optimal extra margin can fairly be characterized by four parameters that describe each particle local flow regime: the particle velocity, the ratio of the containing cell size to particle size, the containing cell solid fraction, and the total number of particles in the system. For this purpose, we model the near-optimal extra margin as a function of these parameters using a quadratic polynomial function. We use the DAKOTA SOFTWARE to carry out the Design and Analysis of Computer Experiments (DACE) and the sampling of the parameters for the simulations. For a given instance of the set of parameters, a global optimization method is considered to find the near-optimal extra margin. The latter is required for the construction of the quadratic polynomial model. The numerous simulations generated by the sampling of the parameter were performed on a High-Performance Computing (HPC) environment granting parallel and concurrent executions. This work provides a better understanding of the Verlet buffer method in DEM simulations by analyzing its performances and behavior in various configurations. The near-optimal extra margin can reasonably be predicted by two out of the four chosen parameters using the quadratic polynomial model. This model has been integrated into XDEM in order to automatically choose the extra margin without any input from the user. Evaluations on real industrial-level test cases show up to a 26% reduction of the execution time. [less ▲]

Bayesian optimisation to select Rössler system parameters used in Chaotic Ant Colony Optimisation for Coverage

in Journal of Computational Science (2020), 41

Automatic Software Tuning of Parallel Programs for Energy-Aware Executions

in Proc. of 13th Intl. Conf. on Parallel Processing and Applied Mathematics (PPAM 2019) (2019, December)

For large scale systems, such as data centers, energy efficiency has proven to be key for reducing capital, operational expenses and environmental impact. Power drainage of a system is closely related to ... [more ▼]

For large scale systems, such as data centers, energy efficiency has proven to be key for reducing capital, operational expenses and environmental impact. Power drainage of a system is closely related to the type and characteristics of workload that the device is running. For this reason, this paper presents an automatic software tuning method for parallel program generation able to adapt and exploit the hardware features available on a target computing system such as an HPC facility or a cloud system in a better way than traditional compiler infrastructures. We propose a search based approach combining both exact methods and approximated heuristics evolving programs in order to find optimized configurations relying on an ever-increasing number of tunable knobs i.e., code transformation and execution options (such as the num- ber of OpenMP threads and/or the CPU frequency settings). The main objective is to outperform the configurations generated by traditional compiling infrastructures for selected KPIs i.e., performance, energy and power usage (for both for the CPU and DRAM), as well as the runtime. First experimental results tied to the local optimization phase of the proposed framework are encouraging, demonstrating between 8% and 41% improvement for all considered metrics on a reference benchmark- ing application (i.e., Linpack). This brings novel perspectives for the global optimization step currently under investigation within the presented framework, with the ambition to pave the way toward automatic tuning of energy-aware applications beyond the performance of the current state-of-the-art compiler infrastructures. [less ▲]

Tackling Large-Scale and Combinatorial Bi-level Problems with a Genetic Programming Hyper-heuristic

in IEEE Transactions on Evolutionary Computation (2019)

Combinatorial bi-level optimization remains a challenging topic, especially when the lower-level is a NP-hard problem. In this work, we tackle large-scale and combinatorial bi-level problems using GP ... [more ▼]

Combinatorial bi-level optimization remains a challenging topic, especially when the lower-level is a NP-hard problem. In this work, we tackle large-scale and combinatorial bi-level problems using GP Hyper-heuristics, i.e., an approach that permits to train heuristics like a machine learning model. Our contribution aims at targeting the intensive and complex lower-level optimizations that occur when solving a large-scale and combinatorial bi-level problem. For this purpose, we consider hyper-heuristics through heuristic generation. Using a GP hyper-heuristic approach, we train greedy heuristics in order to make them more reliable when encountering unseen lower-level instances that could be generated during bi-level optimization. To validate our approach referred to as GA+AGH, we tackle instances from the Bi-level Cloud Pricing Optimization Problem (BCPOP) that model the trading interactions between a cloud service provider and cloud service customers. Numerical results demonstrate the abilities of the trained heuristics to cope with the inherent nested structure that makes bi-level optimization problems so hard. Furthermore, it has been shown that training heuristics for lower-level optimization permits to outperform human-based heuristics and metaheuristics which constitute an excellent outcome for bi-level optimization. [less ▲]

Co-evolutionary Hybrid Bi-level Optimization

Doctoral thesis (2019)

Multi-level optimization stems from the need to tackle complex problems involving multiple decision makers. Two-level optimization, referred as ``Bi-level optimization'', occurs when two decision makers ... [more ▼]

Multi-level optimization stems from the need to tackle complex problems involving multiple decision makers. Two-level optimization, referred as ``Bi-level optimization'', occurs when two decision makers only control part of the decision variables but impact each other (e.g., objective value, feasibility). Bi-level problems are sequential by nature and can be represented as nested optimization problems in which one problem (the ``upper-level'') is constrained by another one (the ``lower-level''). The nested structure is a real obstacle that can be highly time consuming when the lower-level is $\mathcal{NP}-hard$. Consequently, classical nested optimization should be avoided. Some surrogate-based approaches have been proposed to approximate the lower-level objective value function (or variables) to reduce the number of times the lower-level is globally optimized. Unfortunately, such a methodology is not applicable for large-scale and combinatorial bi-level problems. After a deep study of theoretical properties and a survey of the existing applications being bi-level by nature, problems which can benefit from a bi-level reformulation are investigated. A first contribution of this work has been to propose a novel bi-level clustering approach. Extending the well-know ``uncapacitated k-median problem'', it has been shown that clustering can be easily modeled as a two-level optimization problem using decomposition techniques. The resulting two-level problem is then turned into a bi-level problem offering the possibility to combine distance metrics in a hierarchical manner. The novel bi-level clustering problem has a very interesting property that enable us to tackle it with classical nested approaches. Indeed, its lower-level problem can be solved in polynomial time. In cooperation with the Luxembourg Centre for Systems Biomedicine (LCSB), this new clustering model has been applied on real datasets such as disease maps (e.g. Parkinson, Alzheimer). Using a novel hybrid and parallel genetic algorithm as optimization approach, the results obtained after a campaign of experiments have the ability to produce new knowledge compared to classical clustering techniques combining distance metrics in a classical manner. The previous bi-level clustering model has the advantage that the lower-level can be solved in polynomial time although the global problem is by definition $\mathcal{NP}$-hard. Therefore, next investigations have been undertaken to tackle more general bi-level problems in which the lower-level problem does not present any specific advantageous properties. Since the lower-level problem can be very expensive to solve, the focus has been turned to surrogate-based approaches and hyper-parameter optimization techniques with the aim of approximating the lower-level problem and reduce the number of global lower-level optimizations. Adapting the well-know bayesian optimization algorithm to solve general bi-level problems, the expensive lower-level optimizations have been dramatically reduced while obtaining very accurate solutions. The resulting solutions and the number of spared lower-level optimizations have been compared to the bi-level evolutionary algorithm based on quadratic approximations (BLEAQ) results after a campaign of experiments on official bi-level benchmarks. Although both approaches are very accurate, the bi-level bayesian version required less lower-level objective function calls. Surrogate-based approaches are restricted to small-scale and continuous bi-level problems although many real applications are combinatorial by nature. As for continuous problems, a study has been performed to apply some machine learning strategies. Instead of approximating the lower-level solution value, new approximation algorithms for the discrete/combinatorial case have been designed. Using the principle employed in GP hyper-heuristics, heuristics are trained in order to tackle efficiently the $\mathcal{NP}-hard$ lower-level of bi-level problems. This automatic generation of heuristics permits to break the nested structure into two separated phases: \emph{training lower-level heuristics} and \emph{solving the upper-level problem with the new heuristics}. At this occasion, a second modeling contribution has been introduced through a novel large-scale and mixed-integer bi-level problem dealing with pricing in the cloud, i.e., the Bi-level Cloud Pricing Optimization Problem (BCPOP). After a series of experiments that consisted in training heuristics on various lower-level instances of the BCPOP and using them to tackle the bi-level problem itself, the obtained results are compared to the ``cooperative coevolutionary algorithm for bi-level optimization'' (COBRA). Although training heuristics enables to \emph{break the nested structure}, a two phase optimization is still required. Therefore, the emphasis has been put on training heuristics while optimizing the upper-level problem using competitive co-evolution. Instead of adopting the classical decomposition scheme as done by COBRA which suffers from the strong epistatic links between lower-level and upper-level variables, co-evolving the solution and the mean to get to it can cope with these epistatic link issues. The ``CARBON'' algorithm developed in this thesis is a competitive and hybrid co-evolutionary algorithm designed for this purpose. In order to validate the potential of CARBON, numerical experiments have been designed and results have been compared to state-of-the-art algorithms. These results demonstrate that ``CARBON'' makes possible to address nested optimization efficiently. [less ▲]

A Competitive Approach for Bi-Level Co-Evolution

in 2018 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) (2018, May 25)