Many data centers were built for north to south traffic: information coming from outside the corporate walls to...
the inside. With data center consolidation, virtualization, hyper-converged systems and new applications gaining traction, data flows are shifting to an east to west movement inside the data center -- for example, from a server to a storage system.
In response, businesses are evaluating new network architectures, with leaf-spine architecture gaining a lot of attention. While potentially beneficial, the networking change also presents organizations with a few IT hurdles.
Moving in a different direction
New IT systems, such as hyper-converged, dramatically alter data center traffic flows, according to Brad Casemore, research director of Datacenter Networks at IDC, an analyst firm in Framingham, Mass. As the volume of east to west traffic increases, problems arise. Network bottlenecks emerge among data center devices, such as switch ports going from a server supporting a big data application to a storage system holding the data. If hosts on one access switch need to quickly communicate with systems on another access switch, uplinks among the access layer and aggregation layer become congested.
New network topologies, such as leaf-spine architecture, address these problems. Leaf switches consolidate traffic from users and then connect to the spine, comprised of the network core of servers and storage systems. This design affects how data center components are configured, with the most dramatic changes occurring in the networking layer.
Flatter is better
A leaf-spine architecture flattens enterprise networks from three tiers to two tiers, according to Scott Hogg, CTO at Global Technology Resources Inc., a consulting and technology services supplier in Denver. A three-tier model had an access layer where users enter the network, an aggregation layer where data is collected and a core where the bulk of the processing occurs. The leaf-spine topology has only two layers, and enterprises deploy fewer switches, which reduces deployment and maintenance costs.
A leaf-spine architecture places each leaf switch one hop away from another, eliminating the need for devices to search for or wait for open connections. As a result, latency improves and bottlenecks are minimized.
Leaf switches have multiple connections to the core, which enhances system redundancy and resiliency. In traditional networks, there was one link from the switch to a core. With leaf switches, a business creates a large, non-blocking fabric without single points of contention.
Mitigating potential problems
Network cabling is a mixed bag. Since there are fewer tiers of switches in a leaf-spine architecture, it initially reduces the overall number of cable runs. However, that gain can be offset because leaf switches have multiple links rather than a single connection to the core.
Companies have options for how they configure their systems, and the type of workloads supported often influence that decision. Core connections operate at different speeds -- 10 Gigabit Ethernet, 40 GbE or 100 GbE connections.
A single 100 GbE connection may be suited to a big data application. Here, the system has a lot of interactions between a server and a storage system. Putting them on a big pipe provides the bandwidth needed for such links. The bigger pipes are becoming more popular: market research firm Infonetics found that sales of 100 GbE switches quadrupled and 40 GbE more than doubled in 2015.
If a company has a more varied application mix, say a web server, an Enterprise Resource Planning system and a mobile application, it may want to take a 40 GbE or 100 GbE link and break it out into four 10 GBE or four 25 GbE connections.
In addition to application needs, the decision to split the high-speed cable depends on the number of data center devices supported. If a business has a large number of systems, then it needs dispersed connections. If it has a few high-bandwidth apps, then a big pipe makes sense.
Logically build your network
A leaf-spine architecture changes physical network connections, but also has an effect on logical networking -- the software used to move information from device to device. Companies want to automate many manual business processes. Rather than enter data manually, orchestration tools, such as Chef and Puppet, enable them to automate that process.
The next step in that process is the movement to software-defined networks (SDN). This technology has been slow to take hold because the historic network topologies were limited in the number of connections they supported and unable to easily change configurations.
"Leaf-spine designs are a good fit with SDNs," Hogg said. The fit is one reason why SDN growth is high. IDC expects SDN revenue to increase at a compound annual growth rate of 53.9% from 2014 to 2020 and reach $12.5 billion in 2020.
However, this new network topology presents data center managers with personnel challenges. Since training and new business processes are required, a firm must include time and training for employees to learn the new network design.
Explore the newest leaf addition from Arista Networks
Leaf-spine replaces three-layer designs
Spine hardware boosts capacity in big data centers