RUNNING AND ACCELATING CLOUD APPLICATIONS ON FIELD PROGRAMABLE GRATE ARRAYS

FPGA

1. G. P. R. T. Pathirana (239342C) RUNNING AND ACCELERATING CLOUD APPLICATIONS ON FIELD-PROGRAMMABLE GATE ARRAYS (FPGA) SERVER 1

2. G. P. R. T. Pathirana (239342C) Dept. of Computer Science & Engineering, Faculty of Engineering, University of Moratuwa , Sri Lanka RUNNING AND ACCELERATING CLOUD APPLICATIONS ON FIELD-PROGRAMMABLE GATE ARRAYS (FPGA) SERVER Postgraduate Diploma Supervisor Dr. Chathuranga Hettiarachchi 2 Modelling the thermal behavior of a lithium-ion battery using Sri Lankan graphite

3. CONTENT 3

4. Cloud Computing Overview Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Cloud computing enables on-demand access to computing resources over the internet. Key Technologies 🔹 Virtualization - Separates physical hardware from software via a hypervisor . 🔹 Distributed Computing - Allows multiple systems to work together for scalability . Significance & Industry Impact 🔹 Used in AI, machine learning, IoT, finance, and healthcare . 🔹 Supports big data analytics, remote access, and real-time applications . 4 M. Armbrust et al., doi : 10.1145/1721654.1721672., M. A. Vouk , doi : 10.2498/cit.1001391., R. Periasamy , doi : 10.1109/CCEM.2012.6354621., B. Gan and S. Jin, doi : 10.26855/acc.2024.07.004., V. Betz, J. Rose, and A. Marquardt, doi : 10.1007/978-1-4615-5145-4.

5. Performance Challenges in Cloud Computing 5 Mitigation Strategies 🔹 Content Delivery Networks (CDNs) & Edge Computing reduce latency. 🔹 Resource Optimization & Auto-scaling improve efficiency. 🔹 Advanced Monitoring & Performance Analytics for early issue detection. Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server What are FPGAs? Field-Programmable Gate Arrays (FPGAs) are reconfigurable silicon devices that can be programmed for specific applications. Some key benefits of FPGAs include 🔹 Reconfigurability 🔹 Parallel Processing 🔹 Power Efficiency 🔹 High performance Figure 1: Challenges in Cloud Computing Figure 2: Overview of the FPGA Architecture B. Gan and S. Jin, doi : 10.26855/acc.2024.07.004., V. Betz, J. Rose, and A. Marquardt, doi : 10.1007/978-1-4615-5145-4.

6. Challenges in FPGA Adoption Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 6 🔹 Programming Complexity – Requires HDL (Verilog, VHDL) , limiting accessibility. 🔹 Security Risks – Bitstream vulnerabilities & side-channel attacks . 🔹 Resource Utilization Issues – Multi-tenant execution in cloud platforms is underdeveloped . FPGA Programming Technologies 🔹 SRAM-Based FPGAs – High speed & reprogrammable, but volatile. 🔹 Flash-Based FPGAs – Non-volatile, but limited reconfigurations. 🔹 Anti-Fuse FPGAs – Permanent configuration, low power, but non-reprogrammable. Cloud Integration of FPGAs 🔹 Used in AWS EC2 F1, Microsoft Azure Catapult, Alibaba Cloud FPGA ECS . 🔹 Applied in AI, cryptography, real-time video processing, & HPC applications . Figure 1: Static Memory Cell Figure 2: FPGA Software Flow R. Periasamy , doi : 10.1109/CCEM.2012.6354621., V. Betz, J. Rose, and A. Marquardt, doi : 10.1007/978-1-4615-5145-4., S. M. S. Trimberger , doi : 10.1109/MSSC.2018.2822862., Brown S. and Rose J., “FPGA and CPLD Architectures: A Tutorial,” 1996., I. Kuon and J. Rose, doi : 10.1109/TCAD.2006.884574.

7. 7 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔹 Programming Complexity Requires HDL (VHDL, Verilog) , making it difficult for cloud developers. HLS tools lack full compatibility with AI frameworks ( TensorFlow, PyTorch ). 🔹 Resource Management Limitations No standardized multi-tenant execution or dynamic workload allocation . Underutilization of FPGA resources affects cost efficiency. 🔹 Security Risks Bitstream vulnerabilities & side-channel attacks expose FPGA configurations. No standardized security framework for cloud-based FPGA deployments. 🔹 Lack of Benchmarking Standards Existing evaluations focus only on hardware metrics (power, logic utilization). No application-level benchmarking for AI & big data workloads.

8. Evaluate FPGA performance in cloud computing by benchmarking latency, throughput, power consumption, and resource utilization compared to CPUs and GPUs . Identify challenges in usability, security, and resource management by assessing HLS tools, multi-tenant execution, and workload distribution strategies . Establishing best practices for FPGA integration by proposing cloud-based configurations, deployment strategies, and benchmarking methods for performance, scalability, and security 8 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server

9. Introduction to FPGA Acceleration in Cloud Computing 9 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 Why FPGAs in Cloud Computing? 🔹 Reconfigurable hardware for workload-specific optimization. 🔹 Better performance & efficiency than CPUs & GPUs. 🔹 Growing adoption in AI inference, big data analytics, cryptography, and HPC. 🔸 Key Drivers of FPGA Adoption 🔹 High-performance computing (HPC) with low power consumption. 🔹 On-demand acceleration via FPGA-as-a-Service ( FaaS ) in AWS, Azure, Alibaba Cloud . 🔹 Ideal for AI, real-time analytics, and video processing. J. A. Galaviz-Aguilar, C. Vargas-Rosales, F. Falcone, and C. Aguilar-Avelar, doi : 10.3390/s25020584., X. Wang, Y. Niu, F. Liu, and Z. Xu, doi : 10.1109/TCC.2020.2992548, M. Vaithianathan , S. Udkar , M. Patil, and S. F. Ng, doi : 10.56472/25838628/IJACT-V1I1P107., https://www.vmaccel.com/posts/the-fpga-cloud-advantage

10. Cloud-Based FPGA Services ( FaaS ) 10 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 FPGA-as-a-Service ( FaaS ) Offerings 🔹 AWS EC2 F1 - Used in AI inference, genomics, finance modeling . 🔹 Microsoft Azure Catapult - Optimized for search, AI, and networking acceleration . 🔹 Alibaba Cloud FPGA ECS - Supports video processing, encryption, AI workloads . 🔸 How FaaS Enhances Cloud Computing 🔹 No need for physical FPGA hardware . 🔹 Scalable & cost-effective for businesses. 🔹 On-demand acceleration for compute-intensive applications. M. Shepovalov and V. Akella, doi : 10.1016/j.vlsi.2019.09.007.

11. Comparative Analysis – FPGAs vs. CPUs vs. GPUs vs. ASICs 11 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 Key Takeaways 🔹 FPGAs balance performance & power efficiency better than CPUs/GPUs. 🔹 ASICs offer maximum efficiency but lack flexibility. 🔹 FPGA reconfigurability makes it ideal for evolving cloud workloads. Table 1: Performance comparison of CPUs, GPUs, FPGAs, & TPUs in AlexNet inference processing Davide C., “Comparing hls4ml and Vitis AI for CNN Synthesis and Evaluation on FPGA: A Comprehensive Study,” 2022. Table 2: Performance comparison for deep learning inference at the edge

12. High-Level Synthesis (HLS) & FPGA Programming in Cloud 12 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 Evolution from HDL to HLS 🔹 Traditional FPGA development required VHDL/Verilog (low-level programming). 🔹 HLS tools (e.g., Vitis AI, Intel OpenCL, hls4ml ) enable C, C++, Python-based FPGA design 🔹 Simplifies FPGA programming for cloud application developers . 🔸 Security Challenges in FPGA-Based Cloud Computing 🔹 Bitstream Vulnerabilities – Risk of reverse engineering & tampering . 🔹 Side-Channel Attacks – Power/electromagnetic leakage can expose data. 🔹 Hardware Trojans – Malicious FPGA modifications affecting security. 🔸 Mitigation Strategies 🔹 Bitstream encryption & authentication protocols. 🔹 Trusted Execution Environments (TEEs) for secure FPGA workloads. Y. Lu, C. Gao, R. Saha, S. Saha, K. D. McDonald-Maier, and X. Zhai, doi : 10.1007/978-3-031-21867-5_5., C. Du and Y. Yamaguchi, doi : 10.3390/electronics9081275., R. Skhiri , V. Fresse , J.-P. Jamont , and B. Suffran , doi : 10.1109/CIVEMSA.2017.7995321.

13. FPGA Acceleration in AI & Video Processing 13 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 FPGA-Based AI Acceleration 🔹 AI inference on FPGAs is more power-efficient than GPUs . 🔹 CNN & NLP models perform faster inference with lower power consumption . 🔹 Case Studies - YOLOv3 for real-time object detection (70% power reduction vs.GPU ). - Transformer-based NLP on FPGA (50% energy savings vs. GPU). 🔸 FPGA Acceleration for Video Analytics 🔹 Real-time AI video inference with ultra-low latency. 🔹 FPGA pipelining for high-speed video processing. 🔹 Use cases - Smart surveillance, industrial automation, autonomous vehicles. Y. Lu, C. Gao, R. Saha, S. Saha, K. D. McDonald-Maier, and X. Zhai, doi : 10.1007/978-3-031-21867-5_5., J. Wang and S. Gu, doi : 10.1109/ICIST52614.2021.9440554., Davide C., “Comparing hls4ml and Vitis AI for CNN Synthesis and Evaluation on FPGA: A Comprehensive Study,”, C. Du and Y. Yamaguchi, doi : 10.3390/electronics9081275. Figure 1 : FPGA vs. GPU Comparison for AI Workloads

14. FPGA Virtualization & Dynamic Workload Allocation 14 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 Challenges in FPGA Virtualization for Cloud Computing 🔹 Multi-Tenant FPGA Reconfiguration – Requires hardware partitioning for multiple workloads. 🔹 Resource Contention – Shared FPGA resources must be allocated efficiently. 🔹 Latency Overhead – Dynamic reconfiguration introduces scheduling delays . 🔸 Solutions for Efficient FPGA Virtualization 🔹 Partial Reconfiguration (PR) – FPGA logic can be updated without full reset . 🔹 Overlay Architectures – Predefined templates allow fast reprogramming . 🔹 Time-Slicing & Scheduling – Fair resource allocation for multiple users. 🔸 Cloud Case Studies 🔹 Microsoft Azure Catapult – FPGA acceleration for Bing search & AI. 🔹 AWS F1 Instances – FPGA-as-a-Service for deep learning & encryption. 🔹 Alibaba Cloud FPGA ECS – Multi-tenant FPGA execution in AI & big data. Y. Lu, C. Gao, R. Saha, S. Saha, K. D. McDonald-Maier, and X. Zhai, doi : 10.1007/978-3-031-21867-5_5., J. Wang and S. Gu, doi : 10.1109/ICIST52614.2021.9440554., Davide C., “Comparing hls4ml and Vitis AI for CNN Synthesis and Evaluation on FPGA: A Comprehensive Study,”, C. Du and Y. Yamaguchi, doi : 10.3390/electronics9081275.

15. 15 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server

16. 16 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Application Selection for FPGA Acceleration 🔸 Objective 🔹 Identify real-world cloud computing workloads that benefit from FPGA acceleration compared to CPUs/GPUs 🔸 Selected Applications 🔹 AI Model Inference - Accelerating deep learning tasks such as image recognition and NLP 🔹 Real-Time Video Processing - Enhancing video streaming, encoding/decoding, and object detection 🔹 Data Encryption & Cryptographic Processing - Improving security with FPGA - based encryption algorithms (AES, RSA). 🔸 These applications demand low latency, high throughput, and energy efficiency, making them ideal for FPGA acceleration in cloud environments

17. 17 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Performance Benchmarking for FPGA Acceleration 🔸 Objective 🔹 Evaluate FPGA acceleration efficiency in cloud computing by measuring key performance metrics 🔸 Key Performance Metrics 🔹 Latency - Execution time for each FPGA-accelerated task. 🔹 Throughput - Data processed per second. 🔹 Power Consumption - Energy efficiency comparison between FPGA and CPU/GPU. 🔹 Resource Utilization - FPGA hardware efficiency in multi-tenant cloud settings. 🔸 Benchmarking Setup 🔹 Deploy applications on AWS EC2 F1 FPGA instances . 🔹 Compare results with CPU/GPU workloads using Xilinx Vitis & AWS CloudWatch/ Vitis Analyzer. and measure the FPGA efficiency in multi-tenant environments 🔸 Application of TOE Framework 🔹 Technology Factors - Does FPGA acceleration provide measurable performance improvements? 🔹 Organizational Factors - Are cloud providers ready for large-scale FPGA deployment? 🔹 Environmental Factors - How do market trends and regulations impact FPGA cloud adoption? 🔸 Expected Outcome - FPGA acceleration enhances latency, power efficiency, and hardware utilization, making it a scalable cloud solution

18. 18 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Usability Testing for FPGA Adoption 🔸 Objective 🔹Assess developer challenges and adoption feasibility of FPGA programming in cloud computing. 🔸 Participant Groups 🔹 Beginners - No prior FPGA experience, assessing ease of learning. 🔹 Intermediate Users - Some FPGA programming knowledge, evaluating usability. 🔹 Experts - FPGA professionals, testing advanced optimization features. 🔸 Usability Testing Tasks 🔹 Basic Task - Configure an FPGA instance using HLS tools (e.g., Vitis HLS, OpenCL). 🔹 Intermediate Task - Deploy an AI model on FPGA for inference. 🔹 Advanced Task - Optimize FPGA workload for real-time execution and energy efficiency 🔸 Application of UTAUT Model 🔹 Performance Expectancy - Does FPGA acceleration improve cloud computing workloads? 🔹 Effort Expectancy - How difficult is FPGA programming using HLS tools? 🔹 Social Influence - How do industry trends impact FPGA adoption? 🔹 Facilitating Conditions - Are training resources and support available for FPGA developers? 🔸 Expected Outcome - Improved FPGA usability and accessibility through HLS tools will enhance developer adoption

19. 19 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Security Risk Assessment 🔸 Objective 🔹 Identify security vulnerabilities in FPGA cloud deployments and propose mitigation strategies. 🔸 Key Security Risks Identified 🔹 Bitstream Vulnerabilities - Risk of unauthorized modification or reverse engineering. 🔹 Side-Channel Attacks - Exploiting power & electromagnetic leakage to extract sensitive data. 🔹 Multi-Tenant Execution Risks - Data leakage and Denial-of-Service (DoS) threats in shared FPGA environments 🔸 Security Enhancements Proposed 🔹 Bitstream Authentication & Encryption - Prevent unauthorized FPGA modification 🔹 AI-driven Anomaly Detection - Real-time monitoring of FPGA security threats. 🔹 Trusted Execution Environments (TEEs) - Secure FPGA workload execution in multi-tenant cloud platforms . 🔸 Application of TAM Model 🔹 Perceived Usefulness - How effective are FPGA security measures? 🔹 Perceived Ease of Use - Are security mechanisms manageable for cloud providers? 🔸 Expected Outcome - Secure multi-tenant FPGA execution, improving trust and adoption in cloud computing.

20. 20 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Project Timeline (Gantt Chart)

21. 21 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔸 Integrating FPGAs into Cloud Computing 🔹 Addresses resource management, performance optimization, and usability challenges. 🔹 Overcomes latency, throughput, and power efficiency bottlenecks in traditional cloud computing. 🔸 Key Focus Areas 🔹 FPGA Acceleration – Enhancing cloud application performance. 🔹 Usability Challenges – Evaluating High-Level Synthesis (HLS) tools for accessibility. 🔹 Intelligent Resource Management – Optimizing FPGA utilization in cloud environments. 🔸 Expected Contribution 🔹 Structured framework for cloud-based FPGA integration. 🔹 Best practices for cloud service providers to maximize scalability & efficiency . 🔹 Guidelines for optimizing FPGA deployments , improving cloud infrastructure affordability. 🔹 This study provides a strategic approach for accelerating cloud applications using FPGAs, ensuring better performance, accessibility, and resource utilization.

22. 22 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server Recommendations 🔸 Enhancing FPGA Acceleration for Cloud Computing 🔹 Refine FPGA frameworks for AI, real-time analytics, and high-throughput cloud services. 🔹 Extend benchmarking to new cloud architectures for better performance evaluation. 🔸 Optimizing Resource Management 🔹 Standardized multi-tenant execution for workload balancing & scalability. 🔹 Automated workload scheduling to maximize FPGA utilization. 🔸 Strengthening Security 🔹 Advanced encryption & bitstream authentication to mitigate security risks. 🔹 AI-driven anomaly detection for proactive threat identification in FPGA cloud infrastructure. 🔸 Improving Usability & Developer Accessibility 🔹 Enhanced FPGA programming tools with interactive tutorials & guided workflows. 🔹 Cloud-based training programs to support software engineers transitioning to FPGA development. Implementing these strategies will drive broader FPGA adoption in cloud computing, improving efficiency, security, and usability for next-generation cloud infrastructures.

23. 23 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server M. Armbrust et al. , “A view of cloud computing,” Apr. 01, 2010. doi : 10.1145/1721654.1721672. M. A. Vouk , “Cloud computing - Issues, research and implementations,” Journal of Computing and Information Technology , vol. 16, no. 4, pp. 235–246, 2008, doi : 10.2498/cit.1001391. A. S. Tanenbaum and M. van Steen, “Distributed Systems: Principles and Paradigms,” 2006. [Online]. Available: www.minix3.org.This R. Periasamy , “Performance Optimization in Cloud Computing Environment,” in 2012 IEEE International Conference on Cloud Computing in Emerging Markets (CCEM) , IEEE, Oct. 2012, pp. 1–6. doi : 10.1109/CCEM.2012.6354621. B. Gan and S. Jin, “Research on Optimization Strategy of Cloud Computing Resources in Big Data Environment,” Advances in Computer and Communication , vol. 5, no. 3, pp. 177–182, Aug. 2024, doi : 10.26855/acc.2024.07.004. R. Zeng, X. Hou, L. Zhang, C. Li, W. Zheng, and M. Guo, “Performance Optimization for Cloud Computing Systems in the Microservice Era: State-of-the-Art and Research Opportunities.” [Online]. Available: http://my.sjtu.edu.cn/are V. Betz, J. Rose, and A. Marquardt, Architecture and CAD for Deep-Submicron FPGAS . Boston, MA: Springer US, 1999. doi : 10.1007/978-1-4615-5145-4.

24. Acknowledgement 24 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server 🔹 Dr. Chathuranga Hettiarachchi – For his invaluable guidance, support, and expertise in cloud computing and FPGA servers. 🔹 Department of Computer Science, University of Moratuwa – For providing resources, facilities, and an enriching academic environment . 🔹 Thesis Committee Members – For their constructive feedback and valuable insights that enhanced this research. 🔹 My Family – For their unwavering love, encouragement, and support throughout this academic journey. 🔹 Everyone who contributed – Your support and motivation have been instrumental in the successful completion of this research.

25. THANK YOU 25 Running and accelerating cloud applications on Field-Programmable Gate Arrays (FPGA) server