Hyperconvergence in the Network - Part 2
In my last article, I discussed hyperconvergence at the network edge—the concept of composable systems and a scalable set of compute, storage, and network access resources. As the networking market often follows the compute and storage market in a broad sense, it would seem some thought around what hyperconvergence in the network itself might look like. What were the components of the hyperconverged system at the edge?
- White box, or rather treating a machine as a resource based on its capabilities, rather than based on the brand on the front of the box
- Scale out, or rather adding more devices in parallel to increase capacity, rather than adding more resources to a single device
- Pooling, or rather the ability to combine multiple resources virtually to create an apparently larger “single system” that can be used as a single unit
How could these components be provided in the network?
First there is white box; the networking world is already moving in this direction. While network devices, such as firewalls, routers, and switches were once purchased in an “appliance” model, many parts of the networking world are quickly moving towards a disaggregated model, where the hardware and software are purchased as separate “things.” This enables the concept of white box—although the box might not actually be white. The terms “bright box” and “grey box” attempt to capture the reality that you might still buy boxes from brand named vendors, but rather than buying them for their software capabilities, you can now buy them for their hardware capabilities.
Second, there is scale out. The move from traditional hierarchical network designs, particularly in the data center, and towards a flatter spine and leaf design, is the equivalent scale out solution in the networking space. Rather than buying a chassis, and adding cards as needed, you buy a set of single rack unit boxes, and build a network that can be increased (or decreased!) in scope and scale by wiring more boxes in.
Third, there is pooling. Here several different trends in the networking world are working together to create the beginnings of a true pooling capability: the rise of dynamic overlay networks, software defined networks, and network function virtualization.
- To combine these three, consider the spine and leaf network built out of white box devices, with a dynamically created overlay network providing virtual sets of resources as needed. This kind of network can be—
- Scaled in resources by adding more boxes to the spine and leaf underlay, as well as adding more network based services to virtual machines connected to this underlying fabric
Pooled by building virtual networks in the overlay that can consume the services of any number of underlay devices as needed
One important question is the depth of the overlay required to build such a system; most of today’s overlay solutions are very heavy weight, full scale tunneling and based on either a “second control plane,” or a centralized control plane (rather than a more flexible hybrid distributed + centralized control plane). What will eventually be needed in this space is a lighter weight set of control planes and overlay system that will work with underlying hardware better—perhaps not even an “overlay” at all, but rather a set of services that can simply send isolated traffic through the network without the work of building an actual virtual topology. Segment Routing may provide a path to such lightweight overlay solutions.
While there are commercial solutions in this space, and custom solutions built and operated by large scale cloud providers, this is still a nascent market. The solutions available today, either based on vendor specific hardware and software and focused on the Top of Rack (ToR) switch in the data center fabric, or on the hypervisor in the server, are generally hampered by a lack of communication between the network resources—the network processors sitting on the ToR switches—and the overlay switching requirements. Further, these solutions are hampered by the amount of configuration required to simply get the system going, particularly in the underlay space.
But these markets are growing and changing; VMWare, Cumulus, and others are working on solutions that will, over time, likely develop into such a hyperconverged solution. There will always be, of course, an appliance based model; there will always be software and hardware purchased as a single system. But the disaggregation and programmable network movements are paving the way for a new kind of network, more along the lines of hyperconverged compute, storage, and network access resources.