当前位置:网站首页>[recognizing cloud Nativity] Chapter 4 cloud network section 4.9.4.2 - Implementation of smart network card
[recognizing cloud Nativity] Chapter 4 cloud network section 4.9.4.2 - Implementation of smart network card
2022-07-18 16:16:00 【Scattered people in the river】

1 Implementation of intelligent network card
SmartNIC Is by starting from Host CPU On Offload( uninstall ) Workload to network card hardware , To improve Host CPU Processing performance of . Among them “ The workload ” not only Networking, It can also be Storage、Security etc. .

1.1 Smart NIC How to implement
According to the design of intelligent network card, adopt the form , Divide it into :
1、 Multi core smart network card , Based on the inclusion of multiple CPU Kernel ASIC( Special application integrated circuits ) chip :ASIC It has a price advantage , But flexibility is limited , Although based on ASIC Of NIC Relatively easy to configure , But the ultimate functionality will be based on ASIC The limitation of the function defined in , This may result in some workloads not being supported .
2、 be based on FPGA( Programmable gate array ) Smart network card of :FPGA It's highly programmable , And it can support almost any function relatively effectively , But as we all know ,FPGA The biggest problem is that programming is difficult and expensive .
3、FPGA Enhanced intelligent network card , take FPGA And ASIC The combination of network controllers .
meanwhile , Because different vendors have different architectures , therefore SmartNIC Can be based on ASIC、FPGA And systems on chip (SOC) Three technologies are implemented .

The cost of intelligent network cards with different implementation methods 、 There are trade-offs between the simplicity and flexibility of programming .
- ASIC Very cost effective , Can provide the best cost performance , But its flexibility is affected by ASIC The limitation of the function defined in .
- be based on FPGA The intelligent network card can be within the constraints of the available logic gates , Spend enough time and cost to support almost any function relatively effectively . but FPGA Programming is not easy and expensive .
- For the more complicated 、 Broader use cases , be based on SoC( Such as NVIDIA Of BlueField DPU) Provides the best intelligent network card implementation options : Good cost performance 、 Easy to program and highly flexible .

Performance comparison between network card and intelligent network card ( As shown in the table 2 Shown )



The above is organized from Consultant Sadie 《2021 China DPU Industry development white paper 》
1.2 Multi core smart network card
A multi-core smart NIC design may include an integration of many ASIC. these ASIC The kernel is usually better ARM processor , They process packets and send them from the main server CPU( expensive ) Uninstall task on . Multi core smart network card ASIC It can also integrate fixed function hardware engine , They can unload well-defined tasks , Such as standardized security and storage protocols .
Multi core smart network card is limited by at least two factors :
- These smart network cards are based on software programmable processors , Due to the lack of processor parallelism , They are slower when used in network processing .
- These are multicore ASIC The fixed function hardware engine in is lack of the data plane programmability and flexibility that the intelligent network card unloading function needs more and more .
The design of multi-core intelligent network card based on processor is in 10G This generation of network card has been widely used . However , As the Ethernet data rate in the data center changes from 10G Rise to 25G、40G、50G、100G Even higher , These software centric multi-core smart network cards are hard to keep up with .
These multi-core smart network cards can not achieve the required peak bandwidth under these high data rates . meanwhile , Multi core smart network card ASIC The fixed function engine in cannot be extended to handle new encryption or security algorithms , That's because they don't have enough programmability , It can only adapt to the subtle changes of the algorithm .
1.3 be based on FPGA Smart network card of
be based on FPGA Smart network card using FPGA Greater hardware programmability to build any data plane functionality required for tasks offloaded to the smart NIC . because FPGA It's reprogrammable , utilize FPGA The data plane function can be removed and reconfigured arbitrarily and in real time . All of these offloads run at hardware rather than software speed .
Compared with software based implementation , be based on FPGA The design of smart network card can speed up the network function by several orders of magnitude . In the design of intelligent network card FPGA It can provide the line speed performance and power efficiency of customized hardware , And it can create deep packets that support complex offloading tasks and improve the performance of single stream network / Network processing pipeline .
By using FPGA A lot of hardware parallelism inherent in to replicate these pipelines , Can provide huge big data performance , Enough to meet the high performance of the next generation data center architecture based on faster Ethernet network 、 High bandwidth 、 High throughput demand .
1.4 FPGA Enhanced intelligent network card
The need for backward compatibility leads to FPGA Enhanced intelligent network card , It adds FPGA function . Based on this design , The network card can be the existing multi-core smart network card , It can also be just a simple NIC (NIC ASIC).
The network cards integrated into these designs provide backward compatibility , Especially for super hypervisor compatibility . therefore , Based on the existing network card FPGA Enhanced intelligent network card design , Plus... For extended functionality FPGA, Can realize its investment income immediately , Because it automatically connects with the existing network API Compatible with interface protocol , So you can use the existing API And drivers . One FPGA On board in the design of enhanced intelligent network card FPGA Can significantly improve performance and expand functionality .
Reference link
Article to read SmartNIC - Tencent cloud developer community - Tencent cloud
from DPDK and eBPF Feel the Smart NIC
NVIDIA ASAP² Technology protects and accelerates modern data center applications - You know
FPGA Overview of intelligent network card (5): FPGA inline - You know
ASIC Overview of intelligent network card (10):5 Layer protocol uninstall L5P - You know
FPGA Overview of intelligent network card (6):HostSDN - You know
《 Re recognize the cloud native series 》 Thematic index :
- Chapter one —— Not seeking the overall situation is not enough to seek a domain
- Chapter II calculation 1 section —— Overview of computing virtualization technology
- Chapter 3 cloud storage 1 section —— Overview of distributed cloud storage
- Chapter 4 cloud network section 1 —— A brief introduction to the development of cloud network technology
- Chapter 4 cloud network 4.2 section —— Relevant basic knowledge preparation
- Chapter 4 cloud network 4.3 section —— Important network protocols
- Chapter 4 cloud network 4.3.1 section —— Brief introduction of routing technology
- Chapter 4 cloud network 4.3.2 section ——VLAN technology
- Chapter 4 cloud network 4.3.3 section ——RIP agreement
- Chapter 4 cloud network 4.3.4 section ——OSPF agreement
- Chapter 4 cloud network 4.3.5 section ——EIGRP agreement
- Chapter 4 cloud network 4.3.6 section ——IS-IS agreement
- Chapter 4 cloud network 4.3.7 section ——BGP agreement
- Chapter 4 cloud network 4.3.7.2 section ——BGP Overview of the agreement
- Chapter 4 cloud network 4.3.7.3 section ——BGP Principle of protocol implementation
- Chapter 4 cloud network 4.3.7.4 section —— Advanced features
- Chapter 4 cloud network 4.3.7.5 section —— Practice
- Chapter 4 cloud network 4.3.7.6 section ——MP-BGP agreement
- Chapter 4 cloud network 4.3.8 section —— Policy routing
- Chapter 4 cloud network 4.3.9 section ——Graceful Restart( Smooth restart ) technology
- Chapter 4 cloud network 4.3.10 section ——VXLAN technology
- Chapter 4 cloud network 4.3.10.2 section ——VXLAN Overlay Network design
- Chapter 4 cloud network 4.3.10.3 section ——VXLAN Tunnel mechanism
- Chapter 4 cloud network 4.3.10.4 section ——VXLAN Message forwarding process
- Chapter 4 cloud network 4.3.10.5 section ——VXlan Networking architecture
- Chapter 4 cloud network 4.3.10.6 section ——VXLAN Application deployment scheme
- Chapter 4 cloud network 4.4 section ——Spine-Leaf Network architecture
- Chapter 4 cloud network 4.5 section —— Layer 2 network
- Chapter 4 cloud network 4.6 section ——Underlay and Overlay Concept
- Chapter 4 cloud network 4.7.1 section —— A brief description of the evolution of network virtualization and offload Acceleration Technology
- Chapter 4 cloud network 4.7.2 section ——virtio Introduction to network semi virtualization
- Chapter 4 cloud network 4.7.3 section ——Vhost-net programme
- Chapter 4 cloud network 4.7.4 section vhost-user programme ——virtio Of DPDK Uninstall scheme
- Chapter 4 cloud network 4.7.5 section vDPA programme ——virtio Implementation of semi hardware virtualization
- Chapter 4 cloud network 4.7.6 section ——virtio-blk Storage virtualization solutions
- Chapter 4 cloud network 4.7.8 section ——SR-IOV programme
- Chapter 4 cloud network 4.7.9 section ——NFV
- Chapter 4 cloud network 4.8.1 section ——SDN executive summary
- Chapter 4 cloud network 4.8.2.1 section ——OpenFlow summary
- Chapter 4 cloud network 4.8.2.2 section ——OpenFlow protocol
- Chapter 4 cloud network 4.8.2.3 section ——OpenFlow Operating mechanism
- Chapter 4 cloud network 4.8.3.1 section ——Open vSwitch brief introduction
- Chapter 4 cloud network 4.8.3.2 section ——Open vSwitch How it works
- Chapter 4 cloud network 4.8.4 section ——OpenStack And SDN Integration of
- Chapter 4 cloud network 4.8.5 section ——OpenDayLight
- Chapter 4 cloud network 4.8.6 section ——Dragonflow
Chapter 4 cloud network 4.9.1 section —— Overview of network offload Acceleration Technology
Chapter 4 cloud network 4.9.2 section —— Traditional network offload Technology
Chapter 4 cloud network 4.9.3.1 section ——DPDK Technology review
Chapter 4 cloud network 4.9.3.2 section ——DPDK The principle,
Chapter 4 cloud network 4.9.4.1 section —— Smart network card SmartNIC Scheme overview
Chapter 4 cloud network 4.9.4.2 section —— Intelligent network card implementation
边栏推荐
- bbox center-size representation and corners representation
- #kubeadm安装Kubernetes 1.15最佳实践#
- Sword finger offer 09 Implementing queues with two stacks
- CF514B Han Solo and Lazer Gun
- Halcon 距离计算
- 可再生金融ReFi:提供对地球有利的技术和金融系统
- Sydtek Internship (I): 4K and ble profile burning
- Sword finger offer 46: Translate numbers into strings
- Elaticsearch installs Vietnamese word breaker
- 9. 说说hashCode() 和 equals() 之间的关系?
猜你喜欢

Refi heating up in summer: uniswap V3 and green asset pool are launched on CELO

精度提升方法:自适应Tokens的高效视觉Transformer框架(已开源)

1388.3n pizza dynamic planning

全局变量、局部变量、静态变量和常量的地址分配

JS variable points you don't know
![[untitled] pseudo class selector and box model](/img/4d/eb5f429d4841baf6d88ad6eb7a5611.png)
[untitled] pseudo class selector and box model

#kubeadm安装Kubernetes 1.15最佳实践#

HCIP PPP/HDLC、GRE/MGRE实验

Halcon 距离计算

uniapp扫码原生插件(Google MLKit、zxing;支持同时扫多个码)
随机推荐
Web page making (II)
FPGA 20 routines: 8 Reading and writing of any address of SD card
Refi heating up in summer: uniswap V3 and green asset pool are launched on CELO
如何利用无常损失从流动资金池中提取价值
Sword finger offer 10- I. Fibonacci sequence (4 solutions)
[binary tree] all elements in two binary search trees
Elaticsearch installs Vietnamese word breaker
halcon 图片文件夹重命名
What are the key smart contracts in defi?
Take Celsius as a negative textbook to teach you how to judge the quality of products and avoid investment risks
Sydtek Internship (I): 4K and ble profile burning
自定义类型下(枚举、联合)C语言
bbox center-size representation and corners representation
HCIP --- OSPF综合实验
OSPF的不规则区域
Which is the most suitable price system for the three cloud rendering platforms (I)
【无标题】伪类选择器和盒模型
Oracle存储过程的几种调用方式
Uniapp code scanning native plug-ins (Google mlkit, zxing; support scanning multiple codes at the same time)
重邮SYDTEK实习(一): 4k和BLE profile烧录