ProAc&ve Rou&ng In Scalable Data Centers with PARIS

Transcription

1 ProAc&ve Rou&ng In Scalable Data Centers with PARIS Theophilus Benson Duke University Joint work with Dushyant Arora + and Jennifer Rexford* + Arista Networks *Princeton University

2 Data Center Networks Must Support diverse applica8on High throughput/low latency U8lize mul8ple paths Scale to cloud size 5-10 million VMs Support flexible resource u8liza8on Support seamless VM mobility

3 Evolu8on of Data Center Networks Scalable Seamless mobility Mul8path rou8ng Layer 2: Flat Addresses Layer 3: Hierarchical Addresses Overlays: VL2/Portland PARIS

4 PARIS in a Nutshell PARIS is a scalable and flexible flat layer 3 network fabric. PARIS hierarchically par88ons addresses at the core PARIS runs on a data center of commodity switches

5 Outline Evolu8on of Data Center Networks PARIS Architecture Evalua8on and Conclusion

6 Evolu8on of Data Center Networks Not scalable Seamless mobility No Mul8path Flat layer 2: Spanning Tree Uses flooding to discover loca8on of hosts Supports seamless VM migra8on Traffic restricted to single network path

7 Evolu8on of Data Center Networks Scalable No seamless mobility Mul8path Layer 3:Hierarchical Addresses Host loca8ons are predefined During VM mobility, IP- addresses change Load balances over k shortest paths

8 Evolu8on of Data Center Networks Seamless mobility Mul8path Not scalable Overlay solu8ons: Portland/VL2 Uses two addressing schemes: hierarchical addresses: for rou8ng traffic flat addresses: for iden8fying VMs

9 Overheads introduced by Overlays Solu8ons Flat- Address Resolve Hierarchical- Address Address resolu8on infrastructure Inflated flow startups 8mes Switch CPU for encapsula8on Switch storage for caching address resolu8ons

10 Evolu8on of Data Center Networks Scalable Seamless mobility Mul8path rou8ng Layer 2: Flat Addresses Layer 3: Hierarchical Addresses Overlays: VL2/Portland

11 Challenges.. Develop data center network that supports benefits of overlay rou8ng while elimina8ng.. Overheads of caching and packet- encapsula8on Overheads of address transla8on

12 ProAc&ve Rou&ng In Scalable PARIS Architecture

13 Architectural Principles Flat layer- three network Allows for seamless VM mobility Proac8ve installa8on of forwarding state Eliminates startup latency overheads Hierarchical par88oning of network state Promotes scalability

14 Paris Architecture Network Controller Overheads eliminated Network Controller: Monitors network traffic Performs traffic engineering Tracks network topology Pro- ac8vely installs forwarding entries Pro- ac8ve rule installa8on à No start- up delay for switch rule installa8on No addresses indirec8on à No address resolu8on, encapsula8on, caching /32 network addresses à No broadcast traffic; no ARP Switches: Support ECMP Programmable devices End- Hosts: /32 addresses Default GW: edge switch

15 Evolu8on of Data Center Networks Scalable Seamless mobility Mul8path rou8ng Layer 2: Flat Addresses Layer 3: Hierarchical Addresses Overlays: VL2/Portland PARIS

16 Paris Network Controller Switches have 1 million entries But data center has 5-10 million VMs Each pod has ~100K VMs Core- Addressing Par88on IP- Address across core devices Pod- Addressing Network Controller Pod switch track addresses for all VMs in the pod

17 Pod- Addressing Module Edge & aggrega8on addressing scheme Edge: stores address for all connected end- hosts Pod: stores addresses for all end- hosts in pod

18 Pod- Addressing Module > > > > > >1 default- >(2,3) > > > > > >1 default- >(2,3) Edge & aggrega8on addressing scheme Edge: stores address for all connected end- hosts Agg: stores addresses for all end- hosts in pod

19 Core Addressing- Modules /14 Par88ons the IP- space into virtual- prefix Each core is an Appointed prefix switch (APS) Tracks all address in a virtual- prefix

20 Core Addressing- Modules / / / / /16 Par88ons the IP- space into virtual- prefix Each core is an Appointed prefix switch (APS) Tracks all address in a virtual- prefix

21 Core Addressing- Modules / / / / /16 Par88ons the IP- space into virtual- prefix Each core is an Appointed prefix switch (APS) Tracks all address in a virtual- prefix

22 DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >2 DIP: >2 DIP: >1 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- > DIP: >2 DIP: >2 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- >4 DIP: >1 DIP: >1 DIP:*.*.*.*- >{2,3} DIP: >1 DIP:*.*.*.*- >{2,3}

23 DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >2 DIP: >2 DIP: >1 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- > DIP: >2 DIP: >2 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- >4 DIP: >1 DIP: >1 DIP:*.*.*.*- >{2,3} Limita&ons No Load balancing between the core nodes Mul&- path in core is not u&lized! DIP: >1 DIP:*.*.*.*- >{2,3}

24 DIP: /16- >{1,2} DIP: /16- >{3,4} Not u8lized Highly u8lized DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >2 DIP: >2 DIP: >1 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- > DIP: >2 DIP: >2 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- >4 DIP: >1 DIP: >1 DIP:*.*.*.*- >{2,3} Limita&ons No Load balancing between the core nodes Mul&- path in core is not u&lized! DIP: >1 DIP:*.*.*.*- >{2,3}

25 DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >2 DIP: >2 DIP: >1 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /14- >{3,4} DIP: /16- >4 DIP: /16- > DIP: >2 DIP: >2 DIP: >1 DIP: /16- >3 DIP: /14- >{3,4} DIP: /16- >3 DIP: /16- >4 DIP: /16- >4 DIP: >1 DIP: >1 DIP:*.*.*.*- >{2,3} High- BW PARIS: Connect core nodes in a mesh Change rules at aggrega&on to load balance across core nodes Use Valiant Load- balancing in the core DIP: >1 DIP:*.*.*.*- >{2,3}

26 DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >2 DIP: >2 DIP: >1 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /14- >{3,4} DIP: /16- >4 DIP: /16- > DIP: >2 DIP: >2 DIP: >1 DIP: /16- >3 DIP: /14- >{3,4} DIP: /16- >3 DIP: /16- >4 DIP: /16- >4 DIP: >1 DIP: >1 DIP:*.*.*.*- >{2,3} High- BW PARIS: Connect core nodes in a mesh Change rules at aggrega&on to load balance across core nodes Use Valiant Load- balancing in the core DIP: >1 DIP:*.*.*.*- >{2,3}

27 DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >{7,8} DIP: /16- >{1,2} DIP: /16- >{3,4} DIP: >1 DIP: >2 DIP: >2 DIP: >1 DIP: >1 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- > DIP: >2 DIP: >2 DIP: /16- >3 DIP: /16- >3 DIP: /16- >4 DIP: /16- >4 DIP: >1 DIP: >1 DIP: >1 DIP:*.*.*.*- >{2,3} DIP: >1 DIP:*.*.*.*- >{2,3}

28 Evalua8on

29 Evalua8on How does PARIS scale to large data centers? Does PARIS ensure good performance? How does PARIS perform under failures? How quickly does PARIS react to VM migra8on?

30 Evalua8on How does PARIS scale to large data centers? Does PARIS ensure good performance? How does PARIS perform under failures? How quickly does PARIS react to VM migra8on?

31 TestBed Emulate data center topology using Mininet Generate traffic using IPerf Random traffic traffic matrix Implemented PARIS on NOX Data center topology 32 hosts, 16 edge, 8 aggrega8on, and 4 core No over- subscrip8on Link capacity: Server Uplinks: 1Mbps Switch- Switch: 10Mbps

32 Scaling to Large Data Centers Hosts ports* Flow table size NoviFlow has developed switches with 1 million entries [1]. [1] NoviFlow Datasheet. hpp://bit.ly/1baqd0a.

33 Does PARIS Ensure Good Performance? How low is latency? Recall: random traffic matrix. Communica&on Pa^ern Inter- pod Intra- pod Latency 61us 106us

34 Summary PARIS achieves scalability and flexibility Flat layer 3 network Pre- posi8oning forwarding state in switches Using topological knowledge to par88on forwarding state Our evalua8ons show that PARIS is prac8cal! Scales to large data- centers Can be implemented using exis8ng commodity devices

35 Ques8ons