Resource Accounting and Allocation for GPUs

• Applicant: Professor Dr.-Ing. Frank Bellosa
• Subject Area: Security and Dependability, Operating-, Communication- and Distributed Systems
• Term: from 2015 to 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 270449430

Running high performance computing applications in the cloud has become a viable alternative to buying and maintaining dedicated compute clusters. At the same time, GPUs have started to deliver unprecedented performance for compute intensive applications. Cloud providers are thus facing the challenge of integrating GPUs into their platforms, which typically make extensive use of virtualization. Although first approaches to efficiently virtualize GPUs exist, the infrastructure for GPU resource accounting (1), and the mechanisms for GPU resource assignment (2) are still missing in todays system architectures. We address these challenges for virtualized environments by a sampling based approach for nested resource accounting and a time and space partitioning of GPU resources. Accounting as well as resource assignment can be applied both on the level of the hypervisor as well as on the level of the guest operating system.

• DFG Programme: Research Grants