- 快召唤伙伴们来围观吧
- 微博 QQ QQ空间 贴吧
- 视频嵌入链接 文档嵌入链接
- 复制
- 微信扫一扫分享
- 已成功复制到剪贴板
1. 联邦学习算力挑战与硬件加速-张骏雪
联邦学习已经成为了隐私计算的一个关键技术,随着进一步大规模部署,其性能问题开始受到广泛关注。本次演讲将介绍我们对联邦学习应用的性能定量分析结果并总结联邦学习在算力方面遇到的挑战。
其次,本次演讲也会介绍星云 Clustar 的硬件加速技术,介绍如何通过对联邦学习底层算子的加速,有效地提升联邦学习应用的性能。 最后,本次演讲也会展示硬件加速对联邦学习多个应用端到端性能的提升效果。
张骏雪 ,博士,FATE开源社区TSC Maintainers、开发专委会主要成员、星云 Clustar CTO。主要研究方向为高性能联邦学习系统、高性能人工智能系统、数据中心网络等领域,曾在网络系统顶级学术会议,包括SIGCOMM/CoNEXT等多次发表学术论文。担任联邦学习国际研讨会FL-IJCAL 2021 审稿人,参与了联邦学习IEEE标准工作组,并负责底层算力的标准制定。
展开查看详情
1 .Computational Challenges & Hardware Acceleration of Cross-silo Federated Learning Junxue ZHANG
2 .Outline 1. Background — Federated Learning & FATE 2. Computational Challenges of Federated Learning 3. Hardware Acceleration — Challenges & Opportunities 4. Hardware Acceleration Results
3 .Background Data Silos & Islands of Data Company A Company B Company D Company C
4 .Background Federated Learning Hospital 1 Hospital 2 Hospital 3 database database database NN layers specialized to each hospital and trained locally … … … Global NN layer trained by All data sources in a federated manner … Analyzer Cross-device Federated Learning Cross-silo Federated Learning
5 .Background Cross-silo Federated Learning Horizontal Federated Learning Vertical Federated Learning
6 .Background FATE — An Industrial Grade Federated Learning Framework
7 .Computational Challenges Understand the Performance Challenges of Federated Learning Computation Overhead Networking Overhead
8 .Computational Challenges A Fine-grained Analysis over Computational Overhead ID Crytographic Operations ID Crytographic Operations O1 Paillier Encryption w/ Obfuscation O6 Ciphertext & Cleartext Matrix Multiplication O2 Paillier Encryption w/o Obfuscation O7 Ciphertext Matrix Summation O3 Paillier Decryption O8 RSA Encryption/Decryption O4 Ciphertext Matrix Addition O9 RSA Blind Ciphertext & Cleartext Matrix Element-wise Multi- O1 O5 RSA Unblind plication. O
9 .Computational Challenges Conclusions • Cryptographic operations do cause much performance penalty for cross- silo federated learning applications • Different applications may use different cryptographic operations • Even within one application, different cryptographic operations are used at different time
10 .Hardware Acceleration Idea of Hardware Offloading CPU … Cryptographic Operation 1 … Cryptographic Operation 2 Total Computation Time CPU … … More CO1 CO2 Advanced Hardware Total Computation Time
11 .Hardware Acceleration Brief Introduction to FPGA
12 .Hardware Acceleration One Advantage of FPGA — Low Latency • Fixed structure • Fully-customized structure • High-performance cache can only • 5X on-chip cache, which is distributed evenly on the chip serve part of the ASIC • Large latency • Low latency, high throughput, high pipelining
13 .Hardware Acceleration Straw-man Approach O1: Paillier Encryption w/ Obfuscation O2: Paillier Encryption w/o Obfuscation O3: Paillier Decryption … O10: RSA Unblind Problem 1: FPGA has limited programmable resources to directly bring 10 operations to a good acceleration Problem 2: Wasted resource because not all operations are used at the same time
14 .Hardware Acceleration Microscope these Cryptographic Operations The core of these cryptographic operations is 2 basic operator: modular multiplication & exponentiation The performance of these operations largely reply on the 2 basic operators
15 .Hardware Acceleration Composable & Dynamic Cryptographic Operation Offloading
16 .Hardware Acceleration Fine-grained Pipelining Inter-engine Pipelining Intra-engine Pipelining
17 .Evaluation Results Performance of Cryptographic Operations Cryptographic Operations
18 .Evaluation Results Performance of End-to-end Applications Vertical Logistic Regression Vertical Linear Regression Horizontal Linear Regression
19 .Product FPGA Acceleration Card
20 .Product Product Roadmap
21 .FATE开源社区:全球首个隐私计算、联邦学习开源社区 FATE开源社区是面向隐私计算、联邦学习开源生态中的开发者、贡献者、用户及生态伙伴建立的学习与交流平台, 是全球首个隐私计算、联邦学习开源社区,拥有全球首个工业级安全联邦学习框架。社区以“开源开放,共力创新” 为愿景,不断汇集更多创新力量,现有3000+位来自近千家企业及科研机构的开发者参与社区生态共建。 杨强 唐杰 技术指导委员会主席 技术指导委员会名誉主席 ⚫ 香港科技大学计算机与工程系讲席教授 ⚫ 清华大学计算机系教授、副系主任 ⚫ 加拿大工程院和加拿大皇家科学院两院院士 ⚫ 清华-工程院知识智能联合实验室主任 ⚫ AAAI/ACM/CAAI/IEEE/IAPR/AAAS Fellow ⚫ 国家杰青、王选杰出青年学者,IEEE Fellow ⚫ 香港人工智能与机器人学会理事长 ⚫ 香港人工智能与机器人学会理事长
22 .FATE开源社区:全球首个隐私计算、联邦学习开源社区 开 发 者 : 用户:8000+ PR : Github Star : 3000+ 20+ 4.1K • 汇集3000余位开发者参与社区共建 • 累计实名用户3500+ • 累计发布20次版本更新 • Github 累计获得4.1K Star • 覆盖1576家企业、高校 • 社群用户超8000+ • 持续创新功能开发,充分支撑 联邦学习产业应用
23 .FATE开源社区:全球首个隐私计算、联邦学习开源社区 我们希望与各位一起, 共同打造更安全、易用的FATE开源框架! 共同构筑繁荣开放的联邦学习技术生态! 微信公众号 添加微信小助手 即可加入官方社群交流 官网:https://cn.fedai.org/ GitHub:https://github.com/FederatedAI
24 .THANK YOU