Alibaba Cloud has revealed the hardware design it uses to run networking at its edge locations, and those devices’ reliance on Intel Tofino ASIC
The Chinese cloud champ revealed its tech in April at the 21st USENIX Symposium on Networked Systems Design and Implementation, and in a paper [PDF] titled "LuoShen: A Hyper-Converged Programmable Gateway for Multi-Tenant Multi- Service Edge Clouds" with accompanying slides.
Alibaba Cloud's material describes LuoShen as a "server switch" and revealed the devices have already been used in production for two years at "hundreds of edge sites." The boxes are 2U – an important achievement, because the Chinese cloud's edge locations feature a 42U rack and the org wants it to be filled with servers it can rent, not networking kit.
LuoShen boxes contain a processor, a field-programmable gate array (FPGA), and Tofino – an Intel Ethernet ASIC that Chipzilla suggests as ideal for improving network performance because it is programmable
Alibaba's boxes host the same network functions that run in its cloud. To run them all into a 2U box, the cloud concern created what it described as a novel pipeline that relies on Tofino to handle stateless cloud network functions. Stateful processing is "handled by the CPU and accelerated by the FPGA. All traffic will traverse through the Tofino pipeline before being forwarded to the external networks or distributed to the CPU/FPGA by the converged underlay devices," the paper explains.
"In LuoShen, the 64×100G ports of Tofino are split into different purposes," the paper states. "Some ports directly connect to servers to receive VM traffic, some ports connect to [a switching fabric named] BSW for internet access and cross-region communication, some ports connect to on-premise IDCs [internet datacenters], and some ports connect to the CPU and FPGA for fallback and stateful traffic processing." Tofino is the traffic cop: it waves traffic to the CPU and FPGA as needed, and after those devices have done their job the Intel ASIC performs further processing to ensure packets reach their destination.
LuoShen boxes have a lot of different traffic to handle, as VMs on Alibaba's edge could be chatting to machines in same rack or in virtual private clouds, VMs in a different Alibaba Cloud region, or to a resource on the internet.
Tofino's programmability is therefore key as it means Alibaba has been able to customize the ASICs to implement its network functions and the requirements of packing them into an edge device
LuoShen can be customized in other ways, too, as the FPGA is of course programmable. As the chip is pluggable, Alibaba Cloud is also open to replacing it with other devices. The CPU, however, may be less flexible. The paper notes it's attached to a network card, suggesting LuoShen uses SmartNICs/data processing unit.
Whatever is inside, results are impressive. Alibaba Cloud's paper claims LuoShen achieves 1.2Tbit/sec throughput and has reduced its upfront cost, deployment size and power usage by 75 percent, 87 percent, and 60 percent, respectively, compared with the "role-splitting architecture" it previously used which saw multiple networking devices deployed to its edge locations. ®
