SDN / NFV part V: flexibility or performance?

Early on in my investigations of how SDN and NFV are being implemented in mobile networks, I have found that performance remains one of the largest stumbling blocks the industry has to overcome if we want to transition to next generations networks.

Specifically, many vendors recognize behind closed doors that a virtualized environment today has many performance challenges. It explains probably why so many of the PoCs feature chipset vendors as a participant. A silicon vendor as a main proponent of virtualization is logical, as the industry seeks to transition from purpose-built proprietary hardware to open COTS platforms. It does not fully explain though the heavy involvement of the chipset vendors in these PoCs. Surely, if the technology is interoperable and open, chipset vendor integration would not be necessary?

Linux limitations

Linux as an operating system has been developed originally for single core systems. As multi-core and multithreaded architectures made their appearance, the operating system has shown great limitations in managing particularly demanding data planes applications. When one looks at virtualized network function, one has to contend with the host OS and the guest OS. 
In both cases, a major loss of performance is observed at enter and exits of a VM and the OS. These software interrupt are necessary to pull the packets from the data plane to the application layer so that they can be processed. The cost of software interrupt for kernel Linux access ends up being prohibitive and create bottlenecks and race conditions as the traffic increases and more thread are involved. Specifically, every time the application needs to access the Linux kernel, it must pause the VM, save its context and stall the application to access the kernel. For instance, a base station can have over 100k software interrupt per second.

Intel DPDK and SR-IOV

Intel Data Plane Development Kit (DPDK) developed by 6Wind is used for I/O and packet forwarding functions. A “fast path” is created between the VM and the virtual network interface card (NIC) that improves the data path processing performance. This implementation allows to effectively bypass the guest hypervisor and to provide fast processing of packets between the VM and the host.

At the host level, Single Root I/O Virtualization (SR-IOV) is also used in conjunction with DPDK to provide NIC-to-VM connectivity, bypassing the Linux host and improving packet forwarding performance. The trade off there is that each VM on SR-IOV requires to be tied to a physical network card.

Performance or Flexibility?

The implementation of DPDK and SR-IOV has a cost. While performance for VNFs implementing both techniques show results close to physical appliance, the trade off is flexibility. In implementing these mechanisms, the VMs are effectively bound to the physical hardware resource they depend on. A perfect configuration and complete identical replication of every element, at the software and physical level is necessary for migration and scaling out. While Intel is working on a virtual DPDK integrated in the hypervisor, implementation of SDN / NFV in wireless networks for data plane hungry network functions will force vendors and networks in the short to medium term to choose between performance or flexibility.

More content available here.