海老澤 健太郎 Arrcus, Inc 2 Research & Support Engineer Netmarks Mar 1998 ~ Jun 2001 Regional manager, Global Support Center Network Appliance Jul 2001 ~ Dec 2006 Director of Service Development Operation SGI Japan Feb 2007 ~ Sep 2008 • Product Intro and Support for Operators • Data Center Operation (energy efficiency) • Technical Team Management Network (Enterprise / Telecom / ISP) ATM, VPN(IPsec), xDSL, MPLS Content Delivery Network & Storage Web/Streaming Cache & LB, NFS/SAN Data Center & SaaS (MEX/SGI) Energy Efficiency, DesktopVPN (SaaS) ~10 years Senior Product Manager Sable Networks Apr 2008 ~ Nov 2010 Solution Architect Parallels Dec 2010 ~ Mar 2014 VP of Technology Riava Jul 2014 ~ Sep 2015 Principal Engineer Lumina Networks Nov 2018 ~ Aug 2020 Co-Founder, CTO Ponto Networks Dec 2015 ~ Jun 2018 Network Automation SDN Controller, Orchestrator OpenFlow, FPGA, WhiteBox NOS Switch Design & Development Flow Router, IPv6/v4 Translation (nat64) Service Automation & Container Operation/Business Support System • Product Design and Development • Management and Board member of Startup Companies • Support Japan Market Entry • Technical Consulting / Support • Open-Source Community 10~20 years Lagopus OF Switch Kentaro Ebisawa(海老澤 健太郎) https://www.linkedin.com/in/ebiken/ Principal Researcher TOYOTA Motor Corp. Nov 2016 ~ Dec 2023 Network Infra for Connected Cars Research Professor NTT Ltd. Dec 2020 ~ Nov 2023 Operator Network Technology TOYOTA InfoTechnology (Nov 2016 ~ Mar 2019) Merged to TOYOTA Mortor Corp on Mar 2019 ~25 years Principal Engineer | Arrcus, Inc. Jan 2024 ~ current Virtualization of Data Plane
| 海老澤 健太郎 Arrcus, Inc 13 リリース毎にスイッチASICの帯域が倍増 ファブリック階層の削減 => 遅延の削減 YouTube: “Broadcom AI Interconnect and Tomahawk AI Fabrics” https://www.youtube.com/watch?v=fFojfZpxyIM 800GbE x 16port x 4 switch
| 海老澤 健太郎 Arrcus, Inc 17 North-South Traffic 中心 アプライアンスがゲートウェイとして ネットワーク機能を提供 East-West Traffic 中心 サーバーに近い場所での ネットワーク機能提供が必要に Firewall Load Balancer Server Server Server Server Internet Server Server Server Internet micro-service 半導体(CPU)トレンドの変化による、アプリケーション構成の変化
https://dl.acm.org/doi/10.1145/2749469.2750392 • They found that the diversity of workloads would benefit from flexible architectures, and identify a “datacenter tax” in the lower layers of the software stack that comprises nearly 30% of cycles across jobs and that are prime candidates for hardware specialization and acceleration. • 2020 Facebook: Accelerometer: Understanding Acceleration Opportunities for Data Center Overheads at Hyperscale • https://dl.acm.org/doi/10.1145/3373376.3378450 • Another study by Facebook published in 2020 found that microservices spend as few as 18% of CPU cycles executing core application logic. The remaining cycles were spent in common operations not core to the application logic, including I/O processing, logging, and compression. Facebook believed that accelerating standard building blocks can significantly improve data center performance and they built a model to project possible hardware speedup in microservices. サーバーサイドのデータプレーンプログラミング 〜 NVIDIA Blue Field / DOCA 〜| 日本P4ユーザ会2024@品川 | 海老澤 健太郎 Arrcus, Inc 18 半導体(CPU)トレンドの変化による、アプリケーション構成の変化
| 海老澤 健太郎 Arrcus, Inc 35 Tesla Transport Protocol over Ethernet (TTPoE) TCPをシンプル化しハードウェアで実行 RDMAなど生成AI向けワークロード処理を(ロスレスではなく)通常のイーサネットで実現 E. Quinnell, "Tesla Transport Protocol Over Ethernet (TTPoE): A New Lossy, Exa-Scale Fabric for the Dojo AI Supercomputer," in 2024 IEEE Hot Chips 36 Symposium (HCS), Stanford, CA, USA, 2024 pp. 1-23 | doi: 10.1109/HCS61935.2024.10664947
海老澤 健太郎 Arrcus, Inc 36 Tesla Transport Protocol over Ethernet (TTPoE) E. Quinnell, "Tesla Transport Protocol Over Ethernet (TTPoE): A New Lossy, Exa-Scale Fabric for the Dojo AI Supercomputer," in 2024 IEEE Hot Chips 36 Symposium (HCS), Stanford, CA, USA, 2024 pp. 1-23 | doi: 10.1109/HCS61935.2024.10664947 TCPをシンプル化しハードウェアで実行 RDMAなど生成AI向けワークロード処理を(ロスレスではなく)通常のイーサネットで実現
海老澤 健太郎 Arrcus, Inc 37 E. Quinnell, "Tesla Transport Protocol Over Ethernet (TTPoE): A New Lossy, Exa-Scale Fabric for the Dojo AI Supercomputer," in 2024 IEEE Hot Chips 36 Symposium (HCS), Stanford, CA, USA, 2024 pp. 1-23 | doi: 10.1109/HCS61935.2024.10664947