As NVMe bandwidth increases and device latency falls, the software overhead of the user-kernel storage path has become a first-order bottleneck for fine-grained I/O. Although operating systems provide scatter-gather interfaces such as preadv and pwritev to amortise this cost, integrating these interfaces into mature storage engines typically requires invasive refactoring across buffer management, logging, and caching layers. As a result, logically sequential writes are often emitted as long streams of scalar system calls.

Most research depends to some extent on technologies and computational infrastructures including, and perhaps especially, HCI.Despite the noted environmental impacts associated with information communication technology (ICT) globally, to date little consideration has been given as to how to limit the impact of research and innovation processes themselves.Working to understand the technical and cultural drivers of this impact within the specific but resource-intensive domain of High Performance Computing (HPC), we conducted 25 interviews with academic researchers, providers, funders, and commissioners of HPC.We find intersecting sociocultural and technical dimensions that link to research institutions like conferences, funders, and universities that reinforce and embed, rather than challenge, expectations of growth and waste.At a time when large scale cloud systems, generative AI and ever larger models are multiplying, we argue to de-escalate demand for computing, aiming for more moderate, responsible and meaningful use of computational infrastructures-including within HCI itself.

We propose a new approach for large-scale high-dynamic range computational imaging. Deep Neural Networks (DNNs) trained end-to-end can solve linear inverse imaging problems almost instantaneously. While unfolded architectures provide robustness to measurement setting variations, embedding large-scale measurement operators in DNN architectures is impractical. Alternative Plug-and-Play (PnP) approaches, where the denoising DNNs are blind to the measurement setting, have proven effective to address scalability and high-dynamic range challenges, but rely on highly iterative algorithms. We propose a residual DNN series approach, also interpretable as a learned version of matching pursuit, where the reconstructed image is a sum of residual images progressively increasing the dynamic range, and estimated iteratively by DNNs taking the back-projected data residual of the previous iteration as input. We demonstrate on radio-astronomical imaging simulations that a series of only few terms provides a reconstruction quality competitive with PnP, at a fraction of the cost.

Object storage solutions potentially address long-standing performance issues with POSIX file systems for certain I/O workloads, and new storage technologies offer promising performance characteristics for data-intensive use cases.In this work, we present a preliminary assessment of Intel’s Distributed Asynchronous Object Store (DAOS), an emerging high-performance object store, in conjunction with non-volatile storage and evaluate its potential use for HPC storage. We demonstrate DAOS can provide the required performance, with bandwidth scaling linearly with additional DAOS server nodes in most cases, although choices in configuration and application design can impact achievable bandwidth. We describe a new I/O benchmark and associated metrics that address object storage performance from application-derived workloads.

The A64FX processor from Fujitsu, being designed for computational simulation and machine learning applications, has the potential for unprecedented performance in HPC systems. In this paper, we evaluate the A64FX by benchmarking against a range of production HPC platforms that cover a number of processor technologies. We investigate the performance of complex scientific applications across multiple nodes, as well as single node and mini-kernel benchmarks. This paper finds that the performance of the A64FX processor across our chosen benchmarks often significantly exceeds other platforms, even without specific application optimisations for the processor instruction set or hardware. However, this is not true for all the benchmarks we have undertaken. Furthermore, the specific configuration of applications can have an impact on the runtime and performance experienced.

As HPC systems move into the Exascale era, parallel file systems are struggling to keep up with the I/O requirements from data-intensive problems. While the inclusion of burst buffers has helped to alleviate this by improving I/O performance, it has also increased the complexity of the I/O hierarchy by adding additional storage layers each with its own semantics. This forces users to explicitly manage data movement between the different storage layers, which, coupled with the lack of interfaces to communicate data dependencies between jobs in a data-driven workflow, prevents resource schedulers from optimizing these transfers to benefit the cluster's overall performance. This paper proposes several extensions to job schedulers, prototyped using the Slurm scheduling system, to enable users to appropriately express the data dependencies between the different phases in their processing workflows. It also introduces a new service for asynchronous data staging called NORNS that coordinates with the job scheduler to orchestrate data transfers to achieve better resource utilization. Our evaluation shows that a workflow-aware Slurm exploits node-local storage more effectively, reducing the filesystem I/O contention and improving job running times.

Memory and I/O performance bottlenecks in supercomputing simulations are two key challenges that must be addressed on the road to Exascale. The new byte-addressable persistent non-volatile memory technology from Intel, DCPMM, promises to be an exciting opportunity to break with the status quo, with unprecedented levels of capacity at near-DRAM speeds. Here, we explore the potential of DCPMM in the context of two high-performance scientific applications in terms of outright performance, efficiency and usability for both its Memory and App Direct modes. In Memory mode, we show equivalent performance and better efficiency for a CASTEP simulation that is limited by memory capacity on conventional DRAM-only systems without any changes to the application. For IFS, we demonstrate that a distributed object-store over NVRAM reduces the data contention created in weather forecasting data producer-consumer workflows. In addition, we also present the achievable memory bandwidth performance using STREAM.

When significant communication costs arise in the solution of multidimensional problems on parallel computers, optimal performance cannot always be achieved by perfectly balancing the computational load across cores. Modest sacrifices in the computational load balance may facilitate substantial overall performance improvements by achieving large savings in the costs associated with communications. This general approach is illustrated by application to GS2, an initial value gyrokinetic simulation code developed to study low-frequency turbulence in magnetized plasma. GS2 is parallelised using MPI with the simulation domain decomposed across tasks. The optimal domain decomposition is non-trivial, and is complicated by the fact that several domain decompositions are needed and that these do not all optimise at the chosen task count. Application to GS2, of the novel approach outlined in this paper, has improved performance by up to 17 percent for a representative simulation. Similar strategies may be beneficial in a broader class of problems.

We have previously documented the on-going work in the EUFORIA project to parallelise and optimise European fusion simulation codes, see. This involves working with a wide range of codes to try and address any performance and scaling issues that these codes have. However, as no two simulation codes are exactly the same, it is very hard to apply exactly the same approach to optimising a disparate range of codes. Indeed, it can be seen from that the codes investigated range in terms of performance and ability from well-optimised, highly parallelised codes, to serial or poorly performing codes. After analysing, optimising, and parallelising a range of codes it is, actually, possible to discern a number of distinct optimisation techniques or approaches/strategies that can be used to improve the performance or scaling of a parallel simulation code. This paper outlines the distinct approaches that we have identified, highlighting their benefits and drawbacks, giving an overview of the type of work that is often attempted for fusion simulation code optimisation.

Parallelisation, serial optimisation, compiler tuning, and many more techniques are used to optimise and improve the performance scaling of parallel programs. One area which is frequently not optimised is file I/O. This is because it is often not considered to be key to the performance of a program and also because it is traditionally difficult to optimise and very machine specific. However, in the current era of Peta- and Exascale computing it is no longer possible to ignore I/O performance as it can significantly limit the scaling of many codes when executing on very large numbers of processors or cores. Furthermore, as producing data is the main purpose of most simulation codes any work that can be undertaken to provide improved performance of I/0 can be applicable to a very large range of simulation codes, and provide them with improved functionality (i.e. the ability to produce more data).This paper describes some of the issues surrounding I/O, the technology that is commonly deployed to provide I/O on HPC machines and the software libraries available to programmers to undertake I/O. The performance of all these aspects of I/O on a range of HPC systems were investigated by the authors and a represented in this paper to motivate the discussions in the paper.

One of the central tasks of EUFORIA is to port, parallelise, and optimise fusion simulation codes, developed at individual research institutes in Europe. There are three supercomputer centres involved in the project located at Barcelona, Edinburgh, and Helsinki. For some of the fusion codes simply porting them to one of the supercomputers represents a major advancement in the use of the codes, as they until now have mainly been used by a small user community, or even exclusively by the author of the code. Also, where codes currently can only use one processor (i. e. are serial) providing any parallel functionality can be of major benefit to the code and the code owner(s). Many of the simulation codes for edge and core transport modelling of fusion plasma using high performance computing are estimated to currently require weeks or months of execution time to simulate science at a scale required to model the new fusion reactor ITER, and therefore these codes have to be optimised to run as fast as possible and parallelised in such a way that computer resources are used as effectively as possible. During the first fifteen month of the project, we have successfully ported eleven fusion codes to the supercomputers in Barcelona, Edinburgh and Helsinki. The installation procedure, library requirements and runtime scripts have been documented for each code, and deposited in the EUFORIA software repository and code revision system. Following this a number of these codes have been chosen for code optimisation and improvements in parallelisation and this paper outlines the experience that we have had with some of these codes, the performance improvements achieved, and the techniques used.

The Integrated Tokamak Modelling Task Force (ITM-TF) is developing an infrastructure where the validation needs, as being formulated in terms of multi-device data access and detailed physics comparisons aiming for inclusion of synthetic diagnostics in the simulation chain, are key components. A device independent approach to data transport and a standardized approach to data management (data structures, naming, and access) is being developed in order to allow cross validation between different fusion devices using a single toolset. The effort is focused on ITER plasmas and ITER scenario development on current fusion device. The modeling tools are, however, aimed for general use and can be promoted in other areas of modelling as well. Extensive work has already gone into the development of standardized descriptions of the data (Consistent Physical Objects) providing initial steps towards a complete fusion modelling ontology. The longer term aim is a complete simulation platform which is expected to last and be extended in different ways for the coming 30 years. The technical underpinning is therefore of vital importance. In particular, the platform needs to be extensible and open-ended to be able to take full advantage of not only today's most advanced technologies but also be able to marshal future developments. A full level comprehensive prediction of ITER physics rapidly becomes expensive in terms of computing resources and may cover a range of computing paradigms. The simulation framework therefore needs to be able to use both grid and HPC computing facilities. Hence, data access and code coupling technologies are required to be available for a heterogeneous, possibly distributed, environment. The developments in this area are pursued in a separate project - EUFORIA (EU Fusion for ITER Applications). The current status of ITM-TF and EUFORIA is presented and discussed.