(Page 0 of 1)
Accelerating control plane traffic is critical to increasing datacenter network performance. Control plane traffic -- network traffic related to controlling the network including routing protocols, access protocols and exception tasks – continues to rise.
Internet-based businesses and consumer applications are driving significant growth in traffic to datacenters. Cloud datacenters where applications run on virtual machines (VM) are becoming more prevalent. Cloud datacenters allow applications to be spread across multiple servers with application data replicated across multiple physical locations. Cloud datacenters increase service resiliency and give service providers significantly more flexibility but can reduce server performance due to virtualization overheads and increased server to server traffic.
Control plane acceleration can have a dramatic impact on datacenter throughput. The control plane handles control packets and other exception packets. When the processors running the control plane functions become heavily loaded, the datacenter throughput is slowed. Offloading control plane functions to dedicated processors such as the Avago® Axxia® Enterprise Communication Processor located on a network interface card (NIC) or co-located with the host processor can significantly improve throughput for both traditional datacenters and cloud datacenters.
Datacenter traffic is expected to grow by an average of 25% per year from 2012 to 2017 as shown in the figure below from Cisco’s Global Cloud Index Forecast. Cloud datacenter traffic is expected to grow faster at 35% CAGR. This represents a 4.5-fold increase from 2012 to 2017.
Most traditional and cloud datacenters use standard servers based on x86 processors. The servers are connected using a high-speed Ethernet network. Cloud datacenters use virtualization technology to transform the datacenter by using virtualized servers, storage and networks. This allows the efficient use of compute, storage and network elements and the rapid provisioning of new services. The latest x86 processors already integrate hardware support for address, CPU and I/O virtualization.
Server virtualization allows applications to run on virtual machines (VM) rather than directly on the underlying hardware. This gives service providers maximum flexibility to move applications within a datacenter, between datacenters or onto remote servers outside the datacenter. Hypervisors are used to virtualize the underlying hardware platform, allowing multiple operating systems and VMs to run on a single server.
Datacenter Network Architectures
Network traffic in datacenters consists of data plane and control plane packets. The data plane forwards user data packets as efficiently and quickly as possible to the end destination. The control plane handles routing protocols, access protocols and exception tasks.
Network virtualization is achieved using technologies such as generic routing encapsulation (NVGRE), virtual local area networks (VLANs) and virtual extensible local area networks (VXLANs). VXLAN can support more than 16 million virtual machines. For virtualized networks it is helpful to split the control and data planes to run on separate servers. The servers executing the control plane functions can reside in the same server rack, in a separate datacenter, or in the cloud. This approach has been used for hyperscale datacenters and has been extended with the development of software-defined networking (SDN) to deliver a more comprehensive solution.
Control Plane Acceleration
Two architectures to improve control plane performance are shown below; scale up and scale out. Scale up is the traditional method of upgrading datacenter networks and is widely used in finance and e-commerce. Control plane performance is improved by adding extra or more powerful CPUs to the existing network systems or by adding function specific control processors to augment the server CPUs.
The scale out approach is used for datacenters with separated control and data plane. Control plane performance is increased by providing hardware acceleration closely coupled to the main control plane function.
Axxia® Enterprise Communication Processors implement hardware engines for several control tasks including address learning, ARP offload and Netflow/OAM as well as security functions, access control lists (ACL), policing and QoS. Axxia Enterprise Communication Processors can be co-located with the control plane CPU or located on a PCIe® adapter card or network interface card (NIC). Offloading these common, but computationally intensive tasks, can significantly improve control plane performance, leading to faster, higher-performing datacenters.