Figure 1. MEC accelerated with FPGATo more reduce the service time of MEC application, we build MEC server with hardware accelerators, especially FPGA that has low power characteristic. However, the hardware characteristics of the FPGA make several obstructions to the MEC service deployment. As a solution, we construct the MEC-FPGA framework, which makes it easier for MEC application service providers to deploy services using the FPGA accelerator.
- We put MEC Platform between C-RAN and vEPC
- One of the key features of the MEC Platform is virtualization: both on network and resource
- To support the virtualization, we use Docker and Kubernetes in implementing the MEC Platform.
- Especially, we have developed a Pass through engine that supports the use of FPGAs in Containers, and configure a platform that allows Kubernetes to manage them.
Figure 3. FPGA enabled MEC Platform architecture
- FPGA virtualization with four logic-locked FPGA regions
- Support Dynamic Partial Reconfiguration in each vFPGA region
- Custom MMD library and driver which are compatible with Intel OpenCL SDK
Figure 4. FPGA Virtualization via partial reconfiguration
- A reconfiguration time overhead aware scheduling mechanism lay between MMD layer and driver layer
- Select a proper vFPGA base on service characteristics and available resources
Figure 5. FPGA Hardware scheduler
2. Verifying the feasibility of using FPGA as accelerator in MEC
Figure 6. SIFT object detectionTo verify the feasibility of using an FPGA as an accelerator, we implemented an image processing (among which we selected the SIFT algorithm) that can be used for MEC applications using an FPGA.
Figure 7. Algorithm for SIFT object detectionWe assumed a scenario using the MEC for image processing in the Smart Factory, and obtained the gain at that speed as a result of processing the SIFT in the FPGA. Based on the scenario we build a PoC for the MEC with accelerators
Figure 8. CV Application on FPGA based Accelerator