High Performance Computing (HPC) is nowadays a strategic asset required to sustain the surging demands for massive processing and data-analytic capabilities. In practice, the effective management of such large scale and distributed computing infrastructures is left to a Resource and Job Management System (RJMS). This essential middleware component is responsible for managing the computing resources, handling user requests to allocate resources while providing an optimized framework for starting, executing and monitoring jobs on the allocated resources. The University of Luxembourg has been operating for 15 years a large academic HPC facility which relies since 2017 on the Slurm RJMS introduced on top of the flagship cluster Iris. The acquisition of a new liquid-cooled supercomputer named Aion which was released in 2021 was the occasion to deeply review and optimize the seminal Slurm configuration, the resource limits defined and the sustaining fairsharing algorithm.This paper presents the outcomes of this study and details the implemented RJMS policy. The impact of the decisions made over the supercomputers workloads is also described. In particular, the performance evaluation conducted highlights that when compared to the seminal configuration, the described and implemented environment brought concrete and measurable improvements with regards the platform utilization (+12.64%), the jobs efficiency (as measured by the average Wall-time Request Accuracy, improved by 110.81%) or the management and funding (increased by 10%). The systems demonstrated sustainable and scalable HPC performances, and this effort has led to a negligible penalty on the average slowdown metric (response time normalized by runtime), which was increased by 0.59% for job workloads covering a complete year of exercise. Overall, this new setup has been in production for 18 months on both supercomputers and the updated model proves to bring a fairer and more satisfying experience to the end users. The proposed configurations and policies may help other HPC centres when designing or improving the RJMS sustaining the job scheduling strategy at the advent of computing capacity expansions.

peer reviewed

With renewed global interest for Artificial Intelligence (AI) methods, the past decade has seen a myriad of new programming models and tools that enable better and faster Machine Learning (ML). More recently, a subset of ML known as Deep Learning (DL) raised an increased interest due to its inherent ability to tackle efficiently novel cognitive computing applications. DL allows computational models that are composed of multiple processing layers to learn in an automated way representations of data with multiple levels of abstraction, and can deliver higher predictive accuracy when trained on larger data sets. Based on Artificial Neural Networks (ANN), DL is now at the core of state of the art voice recognition systems (which enable easy control over e.g. Internet-of- Things (IoT) smart home appliances for instance), self-driving car engine, online recommendation systems. The ecosystem of DL frameworks is fast evolving, as well as the DL architectures that are shown to perform well on specialized tasks and to exploit GPU accelerators. For this reason, the frequent performance evaluation of the DL ecosystem is required, especially since the advent of novel distributed training frameworks such as Horovod allowing for scalable training across multiple computing resources.In this paper, the scalability evaluation of the reference DL frameworks (Tensorflow, Keras, MXNet, and PyTorch) is performed over up-to-date High Performance Computing (HPC) resources to compare the efficiency of different implementations across several hardware architectures (CPU and GPU). Experimental results demonstrate that the DistributedDataParallel features in the Pytorch library seem to be the most efficient framework for distributing the training process across many devices, allowing to reach a throughput speedup of 10.11 when using 12 NVidia Tesla V100 GPUs when training Resnet44 on the CIFAR10 dataset.