Dynamic Provisioning of Network Services on Heterogeneous Resources

To reduce the cost and enhance the flexibility of network functions, Virtual Network Functions (VNFs) are deployed on commercially-of-the-shelf (COTS) resources instead of specialized hardware. This, however, downgrades the performance of network functions. To mitigate this issue, leveraging acceleration hardware such as GPU and FPGA has been suggested. This is because accelerators provide better parallelization and pipelining for compute- and network-intensive VNFs. However, studies show that using accelerators are not beneficial for all cases. One of the reason is that accelerators are expensive resources, possibly increasing the service cost. Consequently, the use of accelerators needs to be decided for each situation individually. To address that, in this study, we have extended Pishahang (i.e., a MANO framework) to support dynamic usage of accelerators in the NFV environment. Using the extended Pishahang, then, we analysed the management overhead due to such dynamic usage of accelerators.