Page 19 - Sirius_Annual_Report_2021

Håkon K. Stensland
Xing Cai
Geir Horn
terms. There are already several paying customers
of MELODIC, and it currently has support for managing applications across all the big Cloud providers, Amazon Web Services (AWS3). Google Cloud4, and the Azure Cloud5 from Microsoft. Additionally, there is support for some smaller European Cloud providers, and the open- source Cloud infrastructure management platform OpenStack6, which is used by
• The MORPHEMIC project developing proactive and polymorphic Cross-Cloud application management has successfully passed its mid-term evaluation with a real-world application demonstration.
• Ongoing PhD project in utility optimising Cross-Cloud autonomic application management.
Objective 3:
• Interactions with Equinor on the reservoir simulations through two ongoing PhD projects.
• Identification of SIRIUS partner applications in need for scalability and the projects of Scalable Computing have been presented to the industrial partners on multiple occasions.
Cross-Cloud application management
The Cloud activities of Scalable Computing in SIRIUS was for the first three years focused on contribution to the MELODIC1 Horizon 2020 project. MELODIC supports Cross- Cloud application management through the Application Programming Interface (API) offered by the various Cloud providers and it is thereby able to deploy Cloud computing instances of the application’s components in the form of virtual machines (VMs), containers, and serverless functions. The result of the MELODIC project is a multi-Cloud application management platform. This is available as open source from the main European open-source community OW22, or as a supported and installed package on standard commercial
most academic Cloud installations worldwide, among them the he Norwegian Research and Education Cloud (NREC7).
However, artificial intelligence (AI) applications may benefit from using specialized hardware when training the algorithms. These are accelerators like Graphical Processing Units (GPU), Field Programmable Gate Arrays (FPGA) or Tensor Processing Units (TPUs) tailored for TensorFlow8 processing. Hence, an application component may come as a standard Central Processing Unit (CPU) artefact, which is currently being deployed by MELODIC, or as artefacts compiled for one or more of the hardware accelerators. Furthermore,
the accelerated version will only be beneficial during the training of the AI components, and more costly to use in the Cloud than standard CPUs. It is therefore a need to switch between different component artefacts depending on the application’s need, and the SIRIUS Scalable Computing team leads the effort in the Horizon 2020 project MORPHEMIC9 to add this support.
The optimization of the application’s Cloud deployment configuration in MELODIC is largely reactive as it is based on measured changes in the managed application’s execution context. Acquiring Cloud resources may unfortunately take several minutes, and the execution context may therefore manage to change significantly before the reconfigured appli- cation is up and running on the new resources. MORPHEMIC aims at remedying this by proactive adaptation based on real time series prediction of the measurements of the running application’s execution context and perform the optimization and reconfiguration early so that the Cloud resources will be available when they are needed by the application.
SIRIUS ANNUAL REPORT 2021 | 19