Tencent Virtual switch has successfully enabled its hardware offloading function and started pilot deployment with the help of NVIDIA ConnectX-6 Dx network card and Tencent’s self-designed smart network card.
This means that with the same number of CPU cores, the throughput of TVS is increased by nearly two times, which saves resources and greatly improves product performance.
According to officials, TVS runs on Hypervisor and provides VPC (Virtual Private Cloud) network services for all virtual machines running on the server.
Currently, TVS developed based on DPDK runs on hundreds of thousands of computing servers equipped with ConnectX-6 Dx smart network cards in the Tencent Cloud data center.
With the help of the NVIDIA ConnectX-6 smart network card, TVS can offload its data plane from the CPU to the network card hardware, while keeping the control plane in the hypervisor.
This mode has two advantages:
- Since the network card hardware has higher efficiency in processing network traffic, it improves performance.
- Because TVS only needs to process the control plane, and the control plane consumes far fewer CPU resources than the data plane, thus saving CPU resources. After successfully enabling the hardware offloading function on TVS, with the same number of CPU cores, the throughput of TVS has nearly doubled, and the effect is very significant.
Nvidia said that now, Tencent Cloud plans to upgrade its network infrastructure from 25G to 100G. If TVS only uses CPU and DPDK to implement, it needs more CPU resources to obtain the expected performance required by a 100G network. With the help of the NVIDIA 100G ConnectX-6 Dx smart network card, TVS can offload the entire data plane to the network card hardware. With the same number of CPU cores, better performance can be obtained.
|VIA|
