In this bimonthly feature, HPCwire highlights newly published research in the high-performance computing community and related domains. From parallel programming to exascale to quantum computing, the details are here.

Developing a performance model-based predictor for parallel applications on the cloud

As cloud computing becomes an increasingly viable alternative to on-premises HPC, researchers are turning their eyes to addressing latency and unreliability issues in cloud HPC environments. These researchers – a duo from the Egypt-Japan University of Science and Technology and Benha University – propose a predictor for the execution time of MPI-based cloud HPC applications, finding an 88% accuracy on ten benchmarks.

Authors: Abdallah Saad and Ahmed El-Mahdy.

Investigating portability, performance and maintenance tradeoffs in exascale systems

As the exascale era swiftly approaches, researchers are increasingly grappling with the difficult tradeoffs between major system priorities that will be demanded by such massive systems. These researchers – a team from the University of Macedonia – explore these tradeoffs through a case study measuring the effect of runtime optimizations on code maintainability.

Authors: Elvira-Maria Arvanitou, Apostolos Ampatzoglou, Nikolaos Nikolaidis, Aggeliki-Agathi Tzintzira, Areti Ampatzoglou and Alexander Chatzigeorgiou.

Moving toward a globally acknowledged HPC certification

Skillsets are incredibly important in the HPC world, but certification is far from uniform. This paper, written by a team from four universities in the UK and Germany, describes the HPC Certification Forum: an effort to “categorize, define and examine competencies expected from proficient HPC practitioners.” The authors describe the first two years of the community-led forum and outline plans for the first officially supported certificate in the second half of 2020.

Authors: Julian Kunkel, Weronika Filinger, Christian Meesters and Anja Gerbes.

Uncovering the hidden cityscape of ancient Hermione with HPC

In this paper, a team of researchers from the Digital Archaeology Laboratory at Lund University describe how they used a combination of HPC and integrated digital methods to uncover the ancient cityscape of Hermione, Greece. Using drones, laser scanning and modeling techniques, they fed their inputs into an HPC system, where they rendered a fully 3D representation of the city’s landscape. 

Authors: Giacomo Landeschi, Stefan Lindgren, Henrik Gerding, Alcestis Papadimitriou and Jenny Wallensten.

Examining thermal neutrons’ threat to supercomputers

Off-the-shelf devices are performant, efficient and cheap, making them popular choices for HPC and other compute-intensive fields. However, the cheap boron used in these devices makes them susceptible to thermal neutrons, which these authors (a team from Brazil, the UK and Los Alamos National Laboratory) contend pose a serious threat to those devices’ reliability. The authors examine RAM, GPUs, accelerators, an FPGA and more, tinkering with variables that affect the thermal neutron flux and measuring the threat posed by the neutrons under various conditions.

Authors: Daniel Oliveira, Sean Blanchard, Nathan DeBardeleben, Fernando Fernandes dos Santos, Gabriel Piscoya Dávila, Philippe Navaux, Andrea Favalli, Opale Schappert, Stephen Wender, Carlo Cazzaniga, Christopher Frost and Paolo Rech.

Deploying scientific AI networks at petaflop scale on HPC systems with containers

The computational demands of AI and ML systems are rapidly increasing in the scientific research sphere. These authors – a duo from LRZ and CERN – discuss the complications surrounding the deployment of ML frameworks on large-scale, secure HPC systems. They highlight a case study deployment of a convolutional neural network with petaflop performance on an HPC system.

Authors: David Brayford and Sofia Vallecorsa.

Running a high-performance simulation of a spiking neural network on GPUs

Spiking neural networks (SNNs) are the most commonly used computational model for neuroscience and neuromorphic computing, but simulations of SNNs on GPUs have imperfectly represented the networks, leading to performance and behavior shortfalls. These authors from Tsinghua University propose a series of technical approaches to more accurately representing SNNs on GPUs, including a code generation framework for high-performance simulations.

Authors: Peng Qu, Youhui Zhang, Xiang Fei and Weimin Zheng.

Do you know about research that should be included in next month’s list? If so, send us an email at [email protected]. We look forward to hearing from you.