Towards An Application Objective-Aware Network Interface

on YouTube ↗21m

" summary

Cloud network interface proposes CadentFlow as an ANI with per-flow priorities or deadlines and an explicit application objective, implemented for parameter‑server distributed deep learning and compared to TCP and coflow. Experiments with 8 workers and 8 parameter servers show up to 25% improvement in iteration time, while 16×16 setups show smaller gains because computation becomes the bottleneck. The paper also defines CCT flexibility ratio to quantify how much bandwidth scheduling can be delayed.

AI-written summary · about

" chapters (12)

1. 0:00Intro
2. 0:28Evolution of Application Network Interface 4
3. 2:27What is the ultimate goal of an ANI?
4. 2:48Understanding an Application's Objective
5. 5:03An Example Application: Distributed Deep Lea
6. 6:39Other Applications
7. 8:11Towards A Novel Application Network Inter
8. 9:40Defining CCT flexibility ratio
9. 11:53Distributed DNN Training CadentFlow
10. 15:00Quantifying benefits achievable with a better network a
11. 16:21Performance Improvement
12. 18:30Open Challenges

" related talks

• 14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...Qi Li, Jinzhu Yan, Haotian Xu, Zhuotao Liu, Ke Xu, Mingwei Xu, Jianping Wu · 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24)
• 13mContracting Wide-area Network Topologies to Solve Flow Problems QuicklyPeter Bailis, Matei Zaharia, Behnaz Arzani, Srikanth Kandula, Ishai Menache, Firas Abuzaid · NSDI ’21
• 16mBackpressure Flow ControlKevin Zhao, Thomas E. Anderson, Mohammad Alizadeh, Georgios Nikolaidis, Prateesh Goyal, Preey Shah · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 26mEvolution of Application Visibility for Telcos: From Programmable Switches to Cloud NativeThomas Graf, Himal Kumar · Cloud Native Telco Day North America 2023
• 17mPerformance Interfaces for Network FunctionsRishabh Iyer, Katerina Argyraki, George Candea · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 21mYour Coflow has Many Flows: Sampling them for Fun and SpeedAkshay Jajoo, Y. Charlie Hu, Xiaojun Lin · USENIX ATC '19
• 16mFlow Scheduling with Imprecise KnowledgeKai Chen, Guyue Liu, Wenxin Li, Xin He, Yuan Liu, Keqiu Li, Zhao Ge, Zewei Guan, Heng Qi, Song Zhang · 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24)
• 31mFlow-based Packet Processing Framework on DPDK and VPPHongjun Ni · KubeCon + CloudNativeCon + Open Source Summit China 2019