Google doesn’t mind its cloud services competing on price, but it would like to have some other competitive levers to pull. So today, the company is adding global, HTTP-based load balancing to Google Cloud Platform, along with an unrelated storage feature.
HTTP load balancing is intended for spreading workloads to other regions — other continents, really — during times of overload. If a network path heading into a North America data center starts getting saturated, a new path, possibly leading to a new virtual machine, can be set up quickly in Asia. Google claims the new path can handle 1 million requests per second without any warm-up time.
This won’t work for super-latency-sensitive applications such as high-speed trading, where the distance would actually affect performance. But for nearly all other applications, the difference is minimal, says Tom Kershaw, Google’s product management lead. “For the first time, you truly look at your network as global,” Kershaw says.
It’s already possible to load-balance between regions on any old IP network. Google’s trick is to use one IP address for the service, so that the end user’s machine doesn’t know or care which region its requests are being sent to, Kershaw says. From there, the selling point is a familiar one: Google says this will let customers provision virtual machines on a just-in-time basis, meaning they won’t have to overprovision their networks to handle peak demand.
Obviously, customers could use load balancing in other ways — based on traffic type, to make sure video is always sourced close to the customer, for instance.
HTTP load balancing is the start of what Google hopes becomes a battle of features — as opposed to the price war that’s emerging among Amazon Web Services (AWS) and its competitors, including Google, Microsoft Azure, and Rackspace. The other feature Google is introducing a storage offering called SSD Persistent Disk, which offers a flat cost for high-I/O storage. Normally, storage costs can get unpredictable as the I/O scale surges — particularly in asymmetric cases, such as when a whole bunch of read operations appear at once, Kershaw says.