Dense linear algebra operations appear very frequently in high-performance computing (HPC) applications, rendering their performance crucial to achieve optimal scalability. As many modern HPC clusters contain multi-GPU nodes, BLAS operations are frequently offloaded on GPUs, necessitating the use of optimized libraries to ensure good performance. Unfortunately, multi-GPU systems are accompanied by two significant optimization challenges: data transfer bottlenecks as well as problem splitting and scheduling in multiple workers (GPUs) with distinct memories. We demonstrate that the current multi-GPU BLAS methods for tackling these challenges target very specific problem and data characteristics, resulting in serious performance degradation for any slightly deviating workload. Additionally, an even more critical decision is omitted because it cannot be addressed using current scheduler-based approaches: the determination of which devices should be used for a certain routine invocation. To address these issues we propose a model-based approach: using performance estimation to provide problem-specific autotuning during runtime. We integrate this autotuning into an end-to-end BLAS framework named PARALiA. This framework couples autotuning with an optimized task scheduler, leading to near-optimal data distribution and performance-aware resource utilization. We evaluate PARALiA in an HPC testbed with 8 NVIDIA-V100 GPUs, improving the average performance of GEMM by 1.7× and energy efficiency by 2.5× over the state-of-the-art in a large and diverse dataset and demonstrating the adaptability of our performance-aware approach to future heterogeneous systems.

Nowadays, machine learning (ML), especially deep learning(DL) methods, provide ever more real-life solutions. However, the lack of training data is often a crucial issue for these learning algorithms, the performance accuracy of which relies on the amount and the quality of the available data. This is particularly true when applying ML/DL based methods for specific areas e.g. material characteristics identification, as it requires huge cost of time and manual power getting observational data from real life. In the mean while, simulations on HPC have already been commonly used in computational science due to the fact that it has the ability of generating sufficient and noise free data, which can be used for training the ML/DL based models. However, in order to achieve accurate simulation results the input parameters usually have to be determined and validated by a large number of tests. Furthermore, the evaluation and validation of such input parameters for the simulation often require a deep understanding of the domain specific knowledge, software and programming skills, which can in turn be solved by ML/DL based methods. In this paper, a novel hybrid workflow combining a multi-task neural network and the simulation on high performance computers(HPC) is proposed, which can address the problem of data sparsity and reduce the demand for expertise, resources, and time in determining the validated parameters for simulation. This work is demonstrated through experiments on determination of material behaviors, and the results prove a promising performance (MSE = 0.0386) through this workflow.

Due to the increasing demand for fish consumption, sustainable fishery become more and more challenging. To prevent from overfishing, massive data in open sea fishing have been collected and analyzed to achieve efficient management of fishery. Still, it is extremely difficult for fishers and fishery managers to exploit available data for accurate prediction, because of their limited data processing capacities, and the overall lack of adequate database systems [1].The goal of this work is therefore to analyze the relationship between data collected from all sensors installed on-board fishing vessels and catch weight, to better support generating a map showing likely fishing effort allocation. To do so, we train neural networks to predict catch weight using all available data from sensors on fishing vessels. The raw data are pre-processed using random sampling techniques to be fed into a neural network for training. A multi-layer perceptron (MLP) neural network is proposed as the baseline. We propose a data augmentation method and a training strategy in order to optimize the prediction accuracy of the model. Our data augmentation method conducts random sampling of the original data multiple times, which reduces the root mean square error (RMSE) by 15.8%, as compared with the results obtained by the model trained without data augmentation. Our training strategy works well to further optimize the prediction accuracy of the model trained with an augmented dataset, which significantly decreased the RMSE by 11. 2%. To the best of our knowledge, this is the first study on the catch weight prediction using neural networks.

Containerisation demonstrates its efficiency in application deployment in cloud computing. Containers can encapsulate complex programs with their dependencies in isolated environments, hence are being adopted in HPC clusters. HPC workload managers lack micro-services support and deeply integrated container management, as opposed to container orchestrators (e.g. Kubernetes). We introduce Torque-Operator (a plugin) which serves as a bridge between HPC workload managers and container Orchestrators.