The performance of (industrial) digital equipment relies often on the quality of the employed components and how they are used. When these components are (multi-)processors, this refers to what kind of software is running (criticality, functional and non-functional requirements, etc.), how are the resources of the processors used by the applications, how applications interact with other applications, or with the environment, etc.
Measurable qualities that we think about are non-functional properties (NFPs) such as the load of the processors, execution times, power consumption, number of schedulable applications running concurrently without a significant decrease in quality of other NFPs, etc. While some of these are established by the provider of the equipment or application, in many cases the end-users have a certain degree of freedom in arranging the system according to their own knowledge and preferences. Such situations can very often lead to loss of performance, due to the decrease in the respective quality of one or more associated NFPs.
Artificial intelligence (AI) and machine learning (ML) techniques have been used to find solutions for application mapping, real-time scheduling, etc. However, most of the approaches are not considering resource sharing and the impact on performance of the respective solutions, nor are they transparent w.r.t. the decision-making process, lacking a formal analysis of the output.
PerFlex intends to build and employ a novel ML/ AI approach to characterize activities on multicore processors to preserve the intended performance aspects considered for individual tasks and activities. The approach is supported by formal methods (FM) based validation and verification (V&V) solutions, to ensure the correctness of the applied algorithms and of the obtained results. The outcomes are considered both at design time and run time, including potential correction actions, given predictions generated by ML/AI solutions[PB1] . Overall, PerFlex aims to: 1) Extract the necessary information on the complex relations between NFPs, building the novel ML/AI algorithms on this knowledge; 2) Provide online predictions and measures such that the equipment and the associated applications can be used efficiently; 3) Develop a verifiable AI approach that will assist both the solution provider and the end beneficiary to deploy and maintain a highly performant systems, with respect to key performance indicators (KPIs).
The project brings together two large companies operating in distinct domains (power distribution and management, and telecommunications) but converging with similar stringing concerns on performance related to processor operations. The world-leading competencies in performant embedded and cyber-physical systems of Hitachi Energy AB, in performant and time-critical systems of Ericsson AB are united with the expertise brought by Mälardalen University in the areas of ML/AI, FM, V&V, and multicore/distributed systems.
|First Name||Last Name||Title|
|Peter||Backeman||Associated Senior Lecturer|