Synthetic Test Environment. The virtual training environment (VTE) or synthetic test environment simulates the framework for the DECICE model, which consists of the digital twin and AI-scheduler. By providing artificially generated or curated data accumulated through provided datasets, the VTE is a digital construct that resembles the actual DECICE environment and reflects the compute plane and metric collection without having the need to actually setup a whole compute continuum. This simulated environment allows for a faster, cost efficient and resource-saving training. The VTE constantly provides metrics to the digital twin and AI-scheduler to enable the training process with different test scenarios such as restricted network connections or hardware restrictions. Having a digitized real-world environment enables the capability to to optimize the scheduler for different metrics such as performance, location constraints or energy efficiency. Depending on the use case feature set such as historic information, node architecture or node capabilities can be setup without having the need to actually provide these features via real hardware. Depending on the simulated hardware and environment this can result in vast cost savings during the training process. The overall training process of the AI scheduler is depicted in Figure 5a. The virtual training environ- ment loads a predefined scenario via a JSON file which is the foundation for the configuration of the digital twin. While the VTE takes care of progressing the training and stepping through the scenarios and providing constant metrics to the DT at the same time, the DT in return adapts to the newly given metrics which can be accessed by the scheduler to perform scheduling decisions [Kun+22]. In contrast to the training cycle Figure 5b shows a real scheduling scenario without a VTE. In this case the scheduling decisions from the AI-scheduler will not be used as a metric for analysis and evaluation of the decisions made by the model, but rather forwarded by the control manager (CM) to the underlying compute plane for a new scheduling task. Since the prediction quality of the AI-model relies on training data to come up with plausible and valid scheduling decisions, datasets need to be collected, processed and provided to the AI. For this case the DECICE project will rely on several sources of data such as publicly available datasets provided by Alibaba [Ali23] or Google [Goo19] and generated data by data centers from the GWDG.
Synthetic Test Environment. This project has received funding from the European Union’s Horizon Europe Research and Innovation Programme under Grant Agreement No 101092582. Document Information Deliverable number: D3.1 Deliverable title: Synthetic Test Environment Deliverable version: 1.0 Work Package number: WP3 Work Package title: Open Framework and Virtual Training Environment Responsible partner GWDG Due Date of delivery: 2023-11-30 Actual date of delivery: 2023-11-30 Dissemination level: PU Type: C Editor(s): Xxxxx Xxxxx (UGOE) Xxxxx Xxxxx (GWDG) Contributor(s): Reviewer(s): Xxxx Xxxxxxx (KTH) Xxxxxx Xxxxxxx (TOP-IX) Project name: Device-Edge-Cloud Intelligent Collaboration framEwork Project Acronym: DECICE Project starting date: 2022-12-01 Project duration: 36 months Rights: DECICE Consortium Document History Version Date Partner Description 0.1 2023-11-15 GWDG/UGOE First draft of deliverable 0.2 2023-11-17 GWDG/UGOE Add additional sections 0.3 2023-11-24 KTH/TOP-IX Internal review 1.0 2023-11-28 GWDG/UGOE Final version Disclaimer: The content of this publication is the sole responsibility of the authors, and in no way represents the view of the European Com- mission or its services.
