The growth witnessed in the expanse of data centers has been completely unprecedented. This has been driven by the exponential increases in cloud computing and cloud storage demand that is now being witnessed. While Gigabit switches proved more than sufficient just a few years ago, today, even 3.2 Terabit (3.2T) switches, which currently serve as the fundamental building blocks upon which data center infrastructure is constructed, are being pushed to their full capacity.
While network demands have increased, Moore's law (which effectively defines the semiconductor industry) has not been able to keep up. Instead of scaling at the silicon level, data centers have had to scale out. This has come at a cost though, with ever increasing capital, operational expenditure and greater latency all resulting. Facing this challenging environment, a different approach is going to have to be taken. In order to accommodate current expectations economically, while still also having the capacity for future growth, data centers (as we will see) need to move towards a modularized approach. Scaling out the datacenter
Data centers are destined to have to contend with demands for substantially heightened network capacity - as a greater number of services, plus more data storage, start migrating to the cloud. This increase in network capacity, in turn, results in demand for more silicon to support it.
To meet increasing networking capacity, data centers are buying ever more powerful Top-of-Rack (ToR) leaf switches. In turn these are consuming more power - which impacts on the overall power budget and means that less power is available for the data center servers. Not only does this lead to power being unnecessarily wasted, in addition it will push the associated thermal management costs and the overall Opex upwards. As these data centers scale out to meet demand, they're often having to add more complex hierarchical structures to their architecture as well - thereby increasing latencies for both north-south and east-west traffic in the process.
The price of silicon per gate is not going down either. We used to enjoy cost reductions as process sizes decreased from 90 nm, to 65 nm, to 40 nm. That is no longer strictly true however. As we see process sizes go down from 28 nm node sizes, yields are decreasing and prices are consequently going up. To address the problems of cloud-scale data centers, traditional methods will not be applicable. Instead, we need to take a modularized approach to networking.
PIPEs and Bridges
Today's data centers often run on a multi-tiered leaf and spine hierarchy. Racks with ToR switches connect to the network spine switches. These, in turn, connect to core switches, which subsequently connect to the Internet. Both the spine and the top of the rack layer elements contain full, managed switches.
By following a modularized approach, it is possible to remove the ToR switches and replace them with simple IO devices - port extenders specifically. This effectively extends the IO ports of the spine switch all the way down to the ToR. What results is a passive ToR that is unmanaged. It simply passes the packets to the spine switch. Furthermore, by taking a whole layer out of the management hierarchy, the network becomes flatter and is thus considerably easier to manage.
The spine switch now acts as the controlling bridge. It is able to manage the layer which was previously taken care of by the ToR switch. This means that, through such an arrangement, it is possible to disaggregate the IO ports of the network that were previously located at the ToR switch, from the logic at the spine switch which manages them. This innovative modularized approach is being facilitated by the increasing number of Port Extenders and Control Bridges now being made available from Marvell that are compatible with the IEEE 802.1BR bridge port extension standard.
Solving Data Center Scaling Challenges
The modularized port-extender and control bridge approach allows data centers to address the full length and breadth of scaling challenges. Port extenders solve the latency by flattening the hierarchy. Instead of having conventional ‘leaf’ and ‘spine’ tiers, the port extender acts to simply extend the IO ports of the spine switch to the ToR. Each server in the rack has a near-direct connection to the managing switch. This improves latency for north-south bound traffic.
The port extender also functions to aggregate traffic from 10 Gbit Ethernet ports into higher throughput outputs, allowing for terabit switches which only have 25, 40, or 100 Gbit Ethernet ports, to communicate directly with 10 Gbit Ethernet edge devices. The passive port extender is a greatly simplified device compared to a managed switch. This means lower up-front costs as well as lower power consumption and a simpler network management scheme are all derived. Rather than dealing with both leaf and spine switches, network administration simply needs to focus on the managed switches at the spine layer.
With no end in sight to the ongoing progression of network capacity, cloud-scale data centers will always have ever-increasing scaling challenges to attend to. The modularized approach described here makes those challenges solvable.