Outline n Introduction n Background n Motivation n AMRT Design n - PowerPoint PPT Presentation

The 49th International Conference on Parallel Processing (ICPP 2020) AM AMRT RT : Anti-ECN Marking to Improve Utilization of Receiver-driven Transmission in Data Center Jinbin Hu 1 , Jiawei Huang 1 , zhaoyi Li 1 , Jianxin Wang 1 , Tian He 2 1 Central South Unive versi sity, y, China 2 Unive versi sity y of Minneso sota, USA

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

Introduction n Key idea: Improve link utilization in receiver-driven transmission under multi-bottleneck and dynamic traffic scenarios . n Solution: AMRT uses anti-ECN marked packets to notify the sender of link under-utilization and correspondingly increases sending rate to grab spare bandwidth.

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

Background n Data Center (DC) Spine n Data Center Traffic 10/100G Leaf delay-sensitive flows n throughput-sensitive flows n Hosts n Transport protocols Sender-driven (DCTCP [1] , D 2 TCP [2] , pFabric [3] , DCQCN [4] ,Timely [5] ) n Receiver-driven (pHost [8] , NDP [10] , Homa [7] , Aeolus [11] ) n

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

Motivation n Multiple bottlenecks scenario

Motivation n Dynamic traffic scenario

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

AMRT: A nti-ECN M arking R eceiver-driven T ransmission n AMRT Overview n At switch n At receiver n At sender

AMRT: Design Details n At switch n Packet Interval Estimation n Anti-ECN Marking

AMRT: Design Details n At receiver n Grant Generation n Explicit Feedback n At sender n Receiver-driven Rate Adjustment

AMRT: Model Analysis

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

T estbed Results n Testbed settings 2-layer Leaf-spine topology; n 1Gbps bottleneck link; n

T estbed Results n Testbed settings 2-layer Leaf-spine topology; n 1Gbps bottleneck link; n

Larger-scale Simulations n Simulation settings NS2 simulator; 2-layer Leaf-spine topology n 10Gbps bottleneck link; 400 hosts, 10 ToR switches, 8 core switches n better Reducing the 99 th FCT up to ~56% Reducing the AFCT up to ~49%

Larger-scale Simulations n Simulation settings NS2 simulator; 2-layer Leaf-spine topology n 10Gbps bottleneck link; 400 hosts, 10 ToR switches, 8 core switches n better Improving the link utilization up to ~36%

Larger-scale Simulations n Performance in Many-to-many Communications better Improving the link utilization up to ~60% better

Outline n Introduction n Background n Motivation n AMRT Design n Evaluation n Summary

Summary n Conservative receiver-driven transmission Under-utilization in Multiple bottlenecks scenario ; n Under-utilization in Dynamic traffic scenario; n n Challenges for AMRT How to detect and feedback the under-utilization information to senders to n improve link utilization and guarantee ultra low latency simultaneously? n Key points of AMRT Packet Interval Estimation and Anti-ECN Marking at switches ； n Grant Generation and Explicit Feedback at receivers; n Receiver-driven Rate Adjustment at senders . n

Q&A