Virtual Open Systems Scientific Publications
The 13th IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB 2018), Valencia, Spain.
Contributed slides presentation
The slides presented at this conference are made publicly available.
FPGA virtualization, vFPGAmanager, Network Function Virtualization, FPGA, virtualization, hardware acceleration, 5G cloud.
This research work has been supported by the European research and innovation program Horizon H2020 Next Generation Platform as a Service (NGPaaS), under the Grant Agreement number 761557. The work presented in this paper reflects only authors' view and the European Commission is not responsible for any use that may be made of the information it contains.
Network operators are actively pushing towards the new 5G era and a crucial part to accomplish this is the Network Functions Virtualization (NFV). FPGAs and their hardware accelerators are a promising solution for NFV and 5G cloud environments because of their fast turnaround time and great speed up potential through application parallelism mapping on the reconfigurable fabric.
Recently, consolidation reached a plateau in this field with lightweight virtualization techniques, that require a high over-commitment of FPGA accelerator resources to cope with numerous demands of guests. Although FPGAs can play an important role for the future 5G networks their capability to manage and control them from the upper layers of the software stack is inadequate. The lack of such support coupled with cloud integration and programmability issues can repel potential providers from utilizing FPGAs at their data centers.
This paper presents the communication mechanism of the vFPGAmanager, an FPGA virtualization framework which can be orchestrated, monitored and enables accelerators over-commitment with direct guest access. These are key features to allow potential adopters of FPGA technology to include them in the next generation of NFV systems. The communication mechanism architecture is detailed and then benchmarked to show that even under heavy load on the system it demonstrates a minimal overhead to orchestrate and monitor the FPGA as a resource.
Network Functions Virtualization (NFV) is a paradigm shift for telecom operators data centers, as well as, for the way virtualization is applied to high performance-low latency scenarios such as broadband multimedia systems and broadcasting. Hardware accelerators and in particular FPGAs are increasingly used in such environments, attributed to their computing power/energy efficiency ratio, increased density and reconfigurability. Thus FPGAs are a promising solution for NFV and 5G cloud environments that can be orchestrated as resources shared among multiple guests.
Recently, consolidation reached a plateau in this field with lightweight virtualization techniques, that require a high over-commitment of FPGA accelerator resources to cope with numerous demands of guests. Also, the lack of such virtualization support coupled with cloud integration and programmability issues, limits the deployment of FPGAs in 5G cloud environments.
This paper proposes to solve the aforementioned issues by exposing the virtualization framework of FPGA resources to the cloud stack. The main contribution is the design and implementation of a communication mechanism to enable orchestration for FPGA virtualization. Our benchmarks prove the feasibility of orchestrating hundreds to thousands of VMs without having a significant degradation in system performance.
The remaining structure of the paper is the following: Section II details the vFPGAmanager IP core solution, an FPGA virtualization framework that has the capability to connect multiple virtual machines to the same accelerator, which improves the efficiency and density of the system. Then in Section III, we introduce an innovative communication mechanism for FPGA virtualization framework that contains commands to perform orchestrated acceleration. Next in Section IV, we include a performance analysis of the communication mechanism under varying loads to prove the feasibility of the proposed approach. Related work is described in Section V, while Section VI presents the conclusion of the paper and future development plans for the vFPGAmanager.
Access the full content of this publication
Login or register to access full information