Composing SDN Controller Enhancements with Mozart Theophilus Benson Zhenyu Zhou
Google Runs SDNs @Scale
Shhhh….. App App Controller 3
CONTENTS 02 01 03 01 Evalua>on Background Mozart Design SDNApps’ Assump>ons Performance and Abstrac>ons and Interfaces Overhead 4
What are SDNs? App App SDNApps Controller Rule 1 Rule 2 Network 5
Case Study: Hedera • Improving data center performance App App by load balancing elephant flows Controller Gather Configure new network Rule 1 paths state Rule 2 Merged Rule 6
What are SDNs? Functionality App App SDNApps Controller Rule 1 Rule 2 SDNEnhancements Optimization Merged Rule Network 7
SDNApps’ Assump>ons & SDNEnhancements Rule 1 Rule 2 Merged Rule SDN- Consistent Update Fault Tolerance Path TCAM-Op>mizer Enhancements More results can be found in the paper App App App App SDNApps’ Controller Controller Assump6ons Instantaneous Infinite Hardware Unmodified Ac>ons Updates Resources 8
Shhhh….. App App Controller SDNEnhancements Optimization Merged Rule 9
Case Study: Hedera Rule 1 Rule 2 Merged Rule 10
Related Works App App 3 App 1 App 2 Novel Interface Composer Athens [CoNEXT’14] , Chopin [CoNEXT ’18] Pyre>c [NSDI’13] Requires understanding all SDNEnhancements Ignores SDNEnhancements 11
What abstrac?ons are required to systema?cally include SDNEnhancements into the SDN ecosystem? 12
CONTENTS 02 01 03 01 Evalua>on Background Mozart Design SDNApps’ Assump>ons Performance and Abstrac>ons and Interfaces Overhead 13
Another Story : Compiler Op>miza>on -fgcse-ader-reload -fipa-cp-clone -floop-interchange -floop-unroll-and-jam -fpeel-loops … Image referenced from: https://www.cs.cmu.edu/afs/cs/academic/class/15745-s02/www/lectures/lect1.pdf 14
Mozart -fgcse-ader-reload Compiler … SDN App 1 Mozart SDN Rule 1 App 2 Rule 2 Network Transac6on App 3 {IO} {AD} {LS} {PF} 15
SDN- Flags Rule 1 Merged Rule Rule 2 Op>miza>on App Transac>on Message 1 Flag 1 Message 2 Flag 2 … … {IO} Global Flag TransacAon Checker 16
SDN- Flags Match AcAon Match 1 Ac?on 1 Order Match 2 Ac?on 2 LocaAon 17
Match 1 Ac>on 1 Match Ac>on Ac>on Match 2 Ac>on 2 1 {IO} Merged Rule 1 {AD} … … ... Rule Rule SDN- Flags ... Rule 2 Match AcAon Match 1 Ac?on 1 Order Match 2 Ac?on 2 LocaAon {LS} {PF} Rule Rule 18
CONTENTS 02 01 03 01 Evalua>on Background Mozart Design SDNApps’ Assump>ons Performance and Abstrac>ons and Interfaces Overhead 19
Evalua>on • Data-Plane Mininet Fat Tree • Control-Plane • Workloads Realis>c DCN Synthe>c Best - Random Worst - Stride 20
Evalua>on Rule 1 What’s Mozart’s benefit Rule 2 Can Mozart improve for backward compa>bility? App App performance? Controller Merged Rule How much work does What’s the overhead of Mozart introduce for Mozart? App SDNApp updates? More results can be found in the paper 21
Evalua>on Proac>ve SDNApp (Hedera) Saves 24.8% reduc>on in • aggregate bandwidth introduced by TCAMOp>mizer. More results can be found in the paper 22
Evalua>on Mozart Overhead Sublinear • 1.58% to latency • More results can be found in the paper 23
Conclusion • SDN controllers are ill-equipped with poor primi>ves for suppor>ng SDNApps and abstrac>ons for enabling SDNEnhancements. • Mozart interface allows for a systema>c and principled inclusion of SDNEnhancements into the SDN ecosystem. 24
Thanks for your aMen?ons! 25
Recommend
More recommend