AOS has been supporting newer leaf-spine data center architecture designs as a turnkey product. Apstra sees leaf-spine as the wave of the future. But Apstra has also done workarounds for customers running legacy L2 workloads.
Mansour Karam, Apstra’s CEO and founder, explained that VXLAN has been used in overlays as a way to run an L2 tunnel on top of an L3 infrastructure. “It solved a business problem to dynamically deploy my L2 apps over my L3 fabric,” he said. “That was a good approach at the time because silicon didn’t support VXLAN.”
However, there have been a lot of advances in silicon, and that’s given Apstra the ability to integrate the two layers. “We have VXLAN in the switches themselves,” said Karam. “And you have APIs to control those VXLANs. We don’t have to overlay anymore. We can leverage the capabilities of the switches themselves.”
The end result is that Apstra can now deliver an integrated operating system. The AOS 2.0 release gives all the automation benefits of previous releases of AOS. In addition, Apstra has written device agents for different hardware, making its software vendor agnostic.
Karam said, “We develop device agents for various devices.” The switches can come from the likes of Cisco, Juniper, and Arista, as well as white box manufacturers. “With AOS 2.0, as long as the silicon supports VXLAN we can do the work within AOS to create the device agent to make it interoperable,” he said. “We’re in the early stages of getting these devices to interoperate. We’re doing the heavy lifting. That’s the path we’re on.”
AOS 2.0 also makes troubleshooting easier.
Before the new release of AOS, the decoupling between L2 and L3 made de-bugging extremely difficult. There was a disconnect between the networking teams who manage the underlay and the cloud teams who manage the overlay. AOS 2.0 provides vendor-agnostic telemetry and correlation between the overlay and underlay, operating a network as one system.
“We’re excited about how we recreate state where all the objects are represented along with their relationships,” said Karam. “We have a model for the underlay and then we have a model for the virtual networks. It becomes very easy to track if there’s a problem.”