Google’s openness about its cloud load balancing is part of its initiative to promote and expand its new Cloud and Applications business. In November 2015, Google named Diane Greene, a founder and former CEO of VMware, to lead Cloud and Applications, where it’s going to leverage its own in-house expertise to deliver products for customers.
In March, a Google employee blogged that the company had developed its own load balancing software, which had “been handling most of the traffic to Google services since 2008.”
Then at its GCP Next16 conference later in March, Prajakta Joshi, a product manager for the Google Cloud Platform, gave a presentation about Google cloud load balancing where she said, “It’s basically software-defined, globally-distributed load balancing.”
She said other software load balancers are a good start, but they don’t have the reach provided by Google’s global network. “You can think of Google load balancing as having infinite capacity,” Joshi said.
“Where is the Google load balancer? It’s everywhere. It’s not a single box, it’s not running in a VM, and it’s not a cluster. It’s load balancing distributed across the world.”
She said it is done by distributed, software entities called Google Front Ends (GFEs). (In its March blog, these were referred to as Maglevs, but apparently that name was a little too weird, even for Google.)
Google’s global network includes more than 70 points of presence (PoPs) across 33 countries, and the GFEs live in more than 50 of these PoPs. In addition, Google connects its data centers to the edge via high-throughput, private network connections.
“Traditional cloud load balancing has to traverse the general Internet for the majority of the time and then go to the cloud provider,” she said. “With Google, GFEs are sitting very close to the user. The user request barely traverses the Internet for a short time, then [proceeds] over Google’s high-quality private network. It’s a function of proximity and capacity. It reduces latency and increases throughput for users.”
Its load balancing also includes autoscaling, which helps customers provision their load balancing needs for actual traffic as opposed to peak demand.
“If you provision for peak, you’re wasting resources. If you provision for average, you’re in big trouble. The ideal way is to provision in line with your traffic. Autoscaling does this in a matter of seconds. There is no pre-warming needed.”
In its February earnings call, Google CEO Sundar Pichai made a point of calling out Google Cloud Platform, saying, “Public cloud services are a natural place for us, so we established a business unit late last year to take full advantage of the opportunity.”