Improving operational efficiency of cloud NFV infrastructure

Mobile data traffic will grow more than 10 times from 2015 to 2021. There is a huge demand for data and new services, not only from smartphones, but also from Internet of Things (IoT) devices. Telecom operators need to enhance their competitiveness by providing faster and newer services, especially with the adoption of 5G. The network function virtualization (NFV) infrastructure that telecom operators are currently building must meet this imminent demand. A fundamental challenge facing telecom operators is to provide the most innovative services at the lowest NFV infrastructure cost.

Figure 1 – Mobile data traffic growth driven primarily by video, gaming, data, and IoT applications

Agile and efficient server infrastructure is key

Servers are the core building blocks of NFV infrastructure, hosting web, business logic, network and security applications running in virtual machines (VMs). They are the service and revenue source for operators to provide to mobile and IoT devices and users, and more than 60% of data center costs are spent on these servers and the associated power and cooling each month. Therefore, improving the efficiency of each server in the NFV infrastructure is the primary concern of telecom operators, followed by the use of cloud orchestration tools (such as OpenStack) to manage server-related operational efficiency.

Figure 2 – Server monthly costs based on 3-year server and 10-year infrastructure amortization.

Factors that affect server performance

To provide necessary services, VMs and the applications running in them need access to sufficient resources in the server, such as CPU cores, memory, storage, security policy rules, network bandwidth, and analytical processing, etc. Based on the applications they host in the NFV infrastructure, VMs have different profiles according to their resource requirements, some VMs are compute-intensive, some are memory-intensive, and others are I/O-intensive.

VMs require security and networking services, which are provided by virtual switches (such as Open vSwitch, OVS) or virtual routers (such as Contrail vRouter) managed by OpenStack Networking. They need to report real-time analytics, and in most cases, the virtual switch or virtual router collects analytics data on behalf of the VM and reports the data to a central analytics processing engine.

Providing security, networking, and analytics services using virtual switches or virtual routing can consume up to 12 CPU cores when using traditional server networking technologies such as 10/25/40 Gb Ethernet network interface cards (NICs). For Internet Mixed (IMIX) traffic, I/O throughput to VMs is limited to less than 6 Gb/s. Traffic from mobile and IoT devices transitions to connections and connection setup rates when it reaches data center servers. In these cases, connection setup rates are limited to 5,000 per second, consuming 12 additional CPU cores to handle the connection setup logic in the server software. As a result, compute-intensive VMs do not get the CPU cores they need, and network-intensive VMs do not get the I/O throughput they need. The result is poor server efficiency, where the output of each server is reduced to one-sixth of the expected value.

Server Infrastructure Operational Efficiency

There is a common sense that it is much easier to manage a lot of the same things than to manage islands of different things. Homogeneous servers managed by OpenStack are ideal for NFV infrastructure or any data center infrastructure. Using traditional NICs (or commercial NICs) in servers, the goal of homogeneity of servers across the data center can be achieved. However, the consequence of doing so is that the efficiency of each server is reduced to the minimum. When traditional network cards are used with software-based virtual switches or virtual routers, only dedicated network configurations can be used to meet the needs of specific virtual machine profiles.

For example, a VM that requires low I/O throughput can be used to implement network configuration, execute a software-based virtual switch or virtual router in its kernel, and allocate four to six CPU cores to such processing. In another case, a software-based virtual switch or virtual router is executed in user space using DPDK, and eight to ten CPU cores can be allocated to such processing to provide higher I/O throughput to the VM. In a third case, technologies such as SR-IOV can be used to transfer the highest I/O throughput to VMs of certain profiles. The advantage of this is that the server can be configured to meet the profile requirements of the VM, but the disadvantage is that we have to have different server configurations. VMs cannot be placed and moved freely, and these defects affect the operational efficiency of the NFV server infrastructure.

Figure 3 – Network configuration silos resulting in significantly reduced operational efficiencies

The era of smart NICs has arrived

Whether it is to improve the efficiency of each server or the operational efficiency of the entire data center and NFV infrastructure, SmartNIC can provide solutions to these problems. SmartNIC is a programmable network card that is optimized for COTS (common off-the-shelf) servers and open source deployments, using the latest version of Open vSwitch (OVS) and Contrail vRouter to provide a complete OpenStack management solution.

In terms of single server efficiency, SmartNICs can increase the output of each server by 6 times, resulting in the following benefits:

• Save more than 10 CPU cores
• More than 5 times the I/O throughput while maintaining the rich networking services provided by OVS and Contrail vRouter
• 20 times faster connection establishment rate
• Improves price/performance of real-time analytics by more than 10 times

In addition, SmartNICs support a unified OpenStack-managed COTS server infrastructure, enabling seamless and rapid onboarding of user virtual machines and third-party virtual network functions (VNFs), while maintaining full virtual machine mobility. Therefore, using SmartNICs can significantly improve the operational efficiency of the server infrastructure of the entire data center.