With the development of cloud platform virtualization technology, the development of smart network cards can be basically divided into three stages ( as shown in the figure below) :
Figure Three stages of smart network card development
The first stage: basic function network card
The basic function network card ( that is, the common network card) provides 2x10G or 2x25G bandwidth throughput, has less hardware offload capabilities, mainly Checksum , LRO/LSO , etc., supports SR - IOV , and limited multi-queue capabilities. In the cloud platform virtualization network, there are three main ways for the basic function network card to provide network access to the virtual machine ( VM ): the operating system kernel driver takes over the network card and distributes network traffic to the virtual machine ( VM ); OVS-DPDK Take over the network card and distribute network traffic to the virtual machine ( VM ); and provide network access capabilities to the virtual machine ( VM ) through SR - IOV in a high-performance scenario.
Phase 2: Hardware Offload NIC
It can be considered as the first generation of smart network card, which has rich hardware offloading capabilities. Typical ones include OVS Fastpath hardware offloading, RDMA network hardware offloading based on RoCEv1 and RoCEv2 , and lossless network capabilities ( PFC , ECN , ETS, etc.) in converged networks . Hardware offloading , NVMe-oF hardware offloading in the storage field, and data plane offloading for secure transmission , etc. The smart network card in this period mainly focuses on the offloading of the data plane.
The third stage: DPU smart network card
It can be considered as the second-generation smart network card, adding CPU to the first-generation smart network card , which can be used to offload the tasks of the control plane and some flexible and complex data plane tasks. At present , the DPU smart network card is characterized by first supporting PCIe Root Complex mode and Endpoint mode. When it is configured in PCIe Root Complex mode, it can realize NVMe storage controller and build storage server together with NVMe SSD disk; in addition, due to the large-scale data center Network needs, more stringent requirements for lossless networks, need to solve the network congestion and delay problems caused by Incast traffic and " elephant " traffic in the data center network , major public cloud vendors have proposed their own solutions, such as Alibaba Cloud's High Precision Congestion Control ( HPCC , High Precision Congestion Control ), AWS ' Scalable Reliable Datagram ( SRD , Scalable Reliable Datagram ), etc. DPU smart network cards will introduce more advanced methods when solving such problems, such asFungible 's TrueFabric is a new solution on the DPU smart network card. In addition, the industry has proposed the development direction of full-stack offloading of network, storage, and security in the Hypervisor . Intel as a representative proposed the IPU , which offloads all infrastructure functions to the smart network card, which can fully release the previously used hypervisor management. CPU computing power.
Future DPU smart NIC hardware form
As more and more functions are added to the smart network card, its power will be difficult to be limited within 75 W , which requires an independent power supply system. Therefore, there may be three forms of smart NICs in the future:
( 1 ) For smart network cards with independent power supply, it is necessary to consider the low-level signal identification between the network card status and computing services. Whether the smart network card has entered the service state during or after the startup of the computing system needs to be explored and resolved.
( 2 ) The DPU smart network card without PCIe interface can form a DPU resource pool, which is specially responsible for network functions, such as load balancing, access control, firewall equipment, etc. The management software can directly define the corresponding network functions through the management interface of the smart network card, and provide the corresponding network capabilities as a virtual network function cluster without the PCIe interface.
( 3 ) DPU chips with multiple PCIe interfaces and multiple network ports . For example , the Fungible F1 chip supports 16 dual-mode PCIe controllers, which can be configured as Root Complex mode or Endpoint mode, and 8x100G network interfaces. The PCIe Gen3 x8 interface can support 8 Dual-Socket computing servers , and the network side provides 8 x1 00G bandwidth network ports.
The hardware form of the future smart network card
As a new type of special-purpose processor, DPU will surely become an important part of the future computing system as the demand side changes, and will play a vital role in supporting the next-generation data center.
Source: White Paper on Dedicated Data Processor (DPU), Institute of Computing Technology, Chinese Academy of Sciences, Yan Guihai, etc.
Full text download: https://www.yusur.tech/zkls/zkys/dpu_whitepaper.html