Baidu cloud catches up by offloading cloudy networking and storage to SmartNICs

Also claims database breakthroughs with home-brewed MySQL-compatible code


Chinese web giant Baidu has followed other clouds by adopting SmartNICs to handle boring bits of cloud workloads and leave its Xeons free to do more heavy lifting.

The company last week announced new enhancements to its cloud, among them “Elastic Bare Metal Server V2.0 is built on Baidu’s hardware virtualization technology”, storage said to offer “40 times better performance and 80 percent lower cost per GB storage” and a database eight times faster than MySQL. The company also created many more AI services.

But those innovations were announced on stage, in Mandarin, and then described with not much more than bullet points in documentation that reached the rest of the world.

So The Register asked for more info and while we haven’t been given a lot, what has emerged is intriguing.

It turns out that hardware virtualization technology is SmartNICs – network interface cards packing an Arm CPU – inside cloud servers.

“Hardware virtualization means we use hardware to implement the virtual front-end of network IO and disk IO, and the back-end runs on the SoC of the smart network card,” a Baidu spokesperson told The Register. “Think of it as a hardware offload of the hypervisor, which is essentially the kernel-based virtual machine technology (KVM). In addition to the virtualization of CPU and MEM, KVM also virtualizes disk and net. The Taihang server does not virtualize CPU and MEM and optimizes the virtualization of disk and net while cooperating software and hardware.”

Baidu is not alone in offloading storage and networking to SmartNICs: AWS did so in 2017 with its Nitro subsystem, Alibaba uses a SmartNIC called “X-Dragon, Azure also uses the devices and Oracle does likewise.

The super-fast database is a home-brewed affair named “Gaia” and is a cloud-native concoction that is MySQL compatible, boasts eight nines reliability (That’s annual downtime of 315.58 milliseconds - Ed) and can run on clusters capable of handling 128TB of data. Baidu’s performance claims come in part from concurrent access to storage resources. The GaiaDB brand covers distributed databases, cloud-native databases and graph databases.

And the slashed storage costs? Those look to be the result of adopting faster SSDs, NvME to pool them and remote direct memory access (RDMA) to pipe data between machines at low latency.

Much of the above is tech that’s either open or already adopted elsewhere, meaning Baidu’s putting itself on a level with both western and Chinese clouds.

Which may not sound startling, but will be noticed by developers and customers across Asia, the region into which Baidu plans is next wave of expansion beyond China, partly because it's yet to make extensive translation efforts that make its offerings easier to access. Chinese clouds already do well in the region, winning up to 40 percent market share by some calculations. ®


Biting the hand that feeds IT © 1998–2020