Text

Research Area 3: Computing continuum and digital infrastructure

Goal

This research area aims to facilitate seamless core-to-edge computing, handling computationally heavy workloads using combined cloud, edge, and local computing resources. Targeting algorithms with high computation demands and large datasets, we intend to provide a framework for developing optimised parallel software, adaptive allocation to heterogeneous hardware platforms, and automatic hardware acceleration. This area leverages collaborations across research groups, focusing on het erogeneous computing in networked environments, and builds upon previous research initiatives like DPAC and HERO.

Research Directions

Software development

Develop a model-based software development approach for heterogeneous com puting environments, focusing on the modelling of (i) computational heavy workloads, (ii) computing en vironments combining core and edge resources, and (iii) requirements for performance, responsiveness, energy consumption, etc.

Design, analysis, and estimation

Create methods and tools for quick, approximate resource estimation (like execution time or energy consumption) for workloads. These estimates aid in decisions about software allocation, hardware resource reservation, and dynamic reallocation across core and edge platforms

Robust middlewares

Extend existing middlewares for distributed computing with automated tools for opti mised deployment and orchestration of workloads. This should also support autonomous reconfiguration of deployments based on changes in the available computation elements and/or their connectivity.

Background

The computing infrastructure of the 5/6G network can be extremely complex and het erogeneous with many types of computational elements (e.g., CPUs, GPUs, DSPs, Al-accelerators, crypto-accelerators, and FPGAs) and connectors between them (e.g., 5/6G radio, high-speed busses, GB Ethernet, Wi-Fi, and Bluetooth). Furthermore, the accessible computation- and networking capacity can be expected to vary over time as nodes enter, leave, or move within the network. Some systems will have powerful onboard computing devices to, e.g., handle data-intensive sensors like stereo cameras, lidars, and radars.

We consider a scenario where a system can host one or more edge nodes. Programming and efficiently using the computational capacity offered in a 5/6G network will likely be highly challenging. Add to the challenges that many systems are timing-sensitive, energy-constrained, and/or safety-critical and need high computing capacity and predictable access to computing resources.

To facilitate seamless core-to-edge computing, we will address the need for systems to execute computationally heavy workloads using the combined capacity of core, edge, and local computing resources. The types of workloads that we target are algorithms with large computation demands and/or large datasets.

This includes algorithms for Model Predictive Control (MPC), situation awareness based on data-intensive sensors, machine learning, and Al decision-making, to mention a few. The overall goal of this RA is to provide a framework that enables the development of optimised parallel software for computationally heavy workloads, adaptive allocation of software to heterogeneous hardware platforms, and a capability for automatic hardware acceleration for the software.

Methods

The research methods in this area follow the design science methodology for software engineering.

Each project contribution is achieved through a 4-step cycle of 1. theoretical solution (including investigation of relevant literature), 2. formalization, 3. prototyping, and 4. evaluation. Evaluation and validation are carried out through mixed-method studies. i.e., a mix of qualitative and quantitative studies. In the case of unsatisfactory outcomes of any step, the research method allows going back to any previous step to perform the desired changes or improvements. Moreover, the whole process can be iterated until satisfaction. We leverage our network of industrial collaborations to assess results on industrial applications and benchmarks.

Expected results

Modelling technology: The research area will significantly advance the technology for modeling computationally heavy workloads in a platform-independent way. Modelling techniques are specifically designed to handle computationally heavy workloads in mixed-core and edge-computing environments, addressing both structural and functional aspects of parallel software. Probabilistic modelling is investigated as a solution to model more complex systems.

Innovative a/location mechanisms: Introducing new concepts for parametric software allocation across diverse computing nodes, independent of specific target languages or platforms. This en hances the flexibility and applicability of the modelling language in heterogeneous environments.

Dynamic re-allocation strategies: The development of model transformations and mechanisms that enable dynamic re-allocation of workloads based on extra-functional requirements and constraints, such as energy consumption and resource limitations. This addresses critical challenges in manag ing runtime variability and adapting software development to heterogeneous systems.

Resource estimation techniques: Developing novel methods for estimating resources like worst case execution times, response times, end-to-end delays, and energy consumption at the code and software architecture levels. These techniques offer a detailed understanding of resource demands in heterogeneous computing platforms. Probabilistic estimation techniques will be explored as a solution to get accurate estimations.

Model-based analysis of executable models: Implementing model-based analysis for executable models and intermediate code of workloads, which aids in optimising their performance and effi ciency in varied computing environments.

Optimisation of system performance: By enhancing the analysis and estimation of critical resources, the research area contributes to improving the dynamic allocation, performance efficiency, and overall system effectiveness in core-to-edge computing.

These scientific results are poised to significantly impact software development and deployment for computationally heavy workloads in heterogeneous computing environments, addressing key chal lenges in efficiency, adaptability, and resource management. In addition, during the first four years, the research area will produce two licentiate theses and one complete PhD thesis. We expect one of our senior lecturers to be promoted from docent to full professor and one of our postdocs to assume the assistant lecturer position. Furthermore, the partner companies in RA3 (Ericsson, Hitachi, ABB Process Automation, and Volvo CE) are expected to contribute with new insights into the emerging problems of designing, deploying, and executing workloads in the computing continuum.