Summary of accelerating mobile applications at network edge with software programmable FPGA

Programmable with software FPGA Accelerate mobile applications at the edge of the network

Accelerating Mobile Applications at the Network Edge with Software-Programmable FPGAs

FPGA Significance of acceleration

Nowadays, edge computing has become a new computing paradigm for mobile applications , Can improve performance and energy consumption . say concretely , It offloads compute intensive tasks to adjacent edge nodes , Interactive applications on power limited devices . meanwhile , Field programmable gate array (FPGA) Because of its customizable hardware, it is famous for accelerating computing intensive tasks such as deep learning algorithms in a high-performance and energy-saving way , Has proven to be an attractive solution for speeding up compute intensive workloads . Besides ,FPGA It has been used for computing acceleration of cloud computing .

Edge based offload and edge based offload FPGA The advantage of acceleration , Combine these two technologies , Further improve the response performance of edge computing , Can be deployed at the edge of the network FPGA The accelerator , Accelerate mobile applications from the perspective of computing offload .

Challenge

How to unload the edge and base it on FPGA Accelerate the combination of the two technologies , How to achieve .

programme

The overall architecture of the system is shown in the figure . It consists of a mobile device 、 An edge node ( Represents the edge network ) And a cloud .

Mobile devices adopt Wi-Fi Connect to the edge node . It mainly runs the front-end part of the application , Interface with the calculation unloading module , For sending requests ( For example, raw sensor data ), And pass UI Receive application specific responses .

Edge networks By wireless router and ARM-FPGA The board is connected through Ethernet . It has two main components , That is, uninstall the manager module and the calculation module . The uninstall manager module is implemented on the router , Used to interface with front-end applications , And route the data to the unloading target ( Local computing uninstall module or remote cloud ).

cloud Used to provide traditional based on CPU Benchmark performance of cloud offload .

Based on this scheme, three interactive applications are implemented ： Recognize handwritten digits in a given image , Recognize a given image object , Face detection . Handwritten numeral recognition adopts a relatively simple deep learning model , Object recognition adopts a relatively complex deep learning model , Face does not use deep learning methods , But based on the traditional computer vision algorithm . These applications are compute intensive , It is expected to benefit from the edge unloading method .

Open questions

In our experiment , No, right FPGA And another widely used hardware GPU Compare . Generally speaking ,GPU Can achieve higher throughput , In most cases, the peak speed is usually higher than FPGA fast . However ,FPGA Can reduce the latency of a single request , And consume less energy . We can see FPGA More energy efficient in most cases . Besides ,FPGA Its reconfigurability makes it better than GPU More flexible . Considering these aspects , We think FPGA It is a better choice for edge unloading .

All the applications in our experiments are related to computer vision . Today, , Applications involving audio and voice processing are emerging , And become an important part of interactive applications . Most of the most advanced solutions for such applications are based on deep learning and machine learning algorithms . utilize FPGA Research on accelerated audio and speech processing applications has begun . therefore , We believe that based on FPGA Edge offloading can also speed up applications involving audio and voice .

Our work does not take into account much about the unique characteristics of the network edge . We try to use... At the edge of the network FPGA, Make applications run faster , And verify its effectiveness , Instead of considering the unique case of edge computing to optimize the workload .

Develop efficient FPGA Accelerators are difficult .CPU The program is familiar to most programmers , And there are many based on CPU Your work can be used in interactive applications . by comparison , Development FPGA The program requires programmers to understand the application and FPGA Have good knowledge . The development cycle is much longer , And because of the poor readability of the code , It is difficult to debug hardware programs .

Third , at present FPGA The processor frequency on the board is much lower than that of laptop or virtual machine CPU. in other words , Onboard coprocessors may become based on FPGA The bottleneck of edge unloading . Fortunately, , These problems can be solved by FPGA Development of design tools and hardware performance to solve .

opinion

yes , These problems can be solved by FPGA Development of design tools and hardware performance to solve .

opinion

Edge offloading is attractive for improving the user experience of today's interactive applications , Because of its powerful computing power and energy efficiency ,FPGA Accelerating computing intensive workloads ( Such as deep learning algorithm ) Very good performance in terms of . This paper attempts to combine edge unloading and FPGA The advantages of , Deploy at the edge of the network FPGA To speed up interactive mobile applications , A new network aided computing model is proposed , That is, based on FPGA Edge calculation of . Although based on FPGA The edge unloading of is still in its infancy , But we believe , This paper is helpful to consider using new devices and technologies to improve mobile applications in the context of edge computing .