Networking Case tudies center center Networks Enterprise Backbone Mike Freedman CO 461: Computer Networks Lectures: MW 10-10:50am in Architecture N101 Cellular hcp://www.cs.princeton.edu/courses/archive/spr13/cos461/ Wireless 2 Cloud CompuJng Cloud CompuJng ElasJc resources Expand and contract resources Pay- per- use Infrastructure on demand MulJ- tenancy MulJple independent users ecurity and resource isolajon AmorJze the cost of the (shared) infrastructure Flexible service management 3 4 1
Cloud ervice Models ouware as a ervice Provider licenses applicajons to users as a service E.g., customer relajonship management, e- mail,.. Avoid costs of installajon, maintenance, patches, Cloud ervice Models Infrastructure as a ervice Provider offers raw compujng, storage, and network E.g., Amazon s ElasJc CompuJng Cloud (EC2) Avoid buying servers and esjmajng resource needs PlaXorm as a ervice Provider offers plaxorm for building applicajons E.g., Google s App- Engine, Amazon 3 storage Avoid worrying about scalability of plaxorm 5 6 Enabling Technology: VirtualizaJon MulJ- Tier ApplicaJons ApplicaJons consist of tasks Many separate components Running on different machines MulJple virtual machines on one physical machine ApplicaJons run unmodified as on real machine VM can migrate from one computer to another Commodity computers Many general- purpose computers Not one big mainframe Easier scaling 7 8 2
ComponenJzaJon leads to different types of network traffic North- outh traffic Traffic to/from external clients (outside of datacenter) Handled by front- end (web) servers, mid- Jer applicajon servers, and back- end databases Traffic pacerns fairly stable, though diurnal variajons East- West traffic Traffic within data- parallel computajons within datacenter (e.g. ParJJon/Aggregate programs like Map Reduce) in distributed storage, parjjons transferred to compute nodes, results joined at aggregajon points, stored back into F Traffic may shiu on small Jmescales (e.g., minutes) 9 North- outh Traffic erver Front- End Proxy Router erver Front- End Proxy base erver base 10 East- West Traffic center Network Distributed torage Map Tasks Reduce Tasks Distributed torage 11 12 3
Virtual witch in erver Top- of- Rack Architecture Rack of servers Commodity servers And top- of- rack switch Modular design Preconfigured racks Power, network, and storage cabling 13 14 Aggregate to the Next Level Modularity, Modularity, Modularity Containers Many containers 15 16 4
center Network Topology Capacity Mismatch? Internet AR AR AR AR ~ 1,000 servers/pod Key = Core Router AR = Access Router = Ethernet witch A = Rack of app. servers 17 1 AR AR AR AR 3 2 OversubscripRon : Demand/upply A. 1 > 2 > 3 B. 1 < 2 < 3 C. 1 = 2 = 3 A A A 18 Capacity Mismatch! Layer 2 vs. Layer 3? AR AR AR AR ~ 40:1 ~ 5:1 ~ 200:1 ParRcularly bad for east- west traffic Ethernet switching (layer 2) Cheaper switch equipment Fixed addresses and auto- configurajon eamless mobility, migrajon, and failover IP roujng (layer 3) calability through hierarchical addressing Efficiency through shortest- path roujng MulJpath roujng through equal- cost muljpath 19 20 5
center RouJng Internet DC- Layer 3 DC- Layer 2 AR AR AR AR Outstanding datacenter networking problems remains ~ 1,000 servers/pod == IP subnet Key = Core Router (L3) AR = Access Router (L3) = Ethernet witch (L2) A = Rack of app. servers 21 22 Network Incast erver Network Incast erver Incast arises from synchronized parallel requests server sends out parallel request ( which friends of Johnny are online? Nodes reply at same Jme, cause traffic burst Replies potenjal exceed switch s buffer, causing drops 23 olujons mijgajng network incast A. Reduce TCP s min RTO (ouen use 200ms >> DC RTT) B. Increase buffer size C. Add small randomized delay at node before reply D. Use ECN with instantaneous queue size E. All of above 24 6
Full BisecJon Bandwidth Full BisecJon Bandwidth Not ufficient Eliminate oversubscripjon? Enter FatTrees Provide stajc capacity But link capacity doesn t scale- up. cale out? Build mulj- stage FatTree out of k port switches k/2 ports up, k/2 down upports k 3 /4 hosts: 48 ports, 27,648 hosts 25 Must choose good paths for full bisecjonal throughput Load- agnosjc roujng Use ECMP across muljple potenjal paths Can collide, but ephemeral? Not if long- lived, large elephants Load- aware roujng Centralized flow scheduling, end- host congesjon feedback, switch local algorithms 26 Conclusion Cloud compujng Major trend in IT industry Today s equivalent of factories center networking Regular topologies interconnecjng VMs Mix of Ethernet and IP networking Modular, mulj- Jer applicajons New ways of building applicajons New performance challenges 27 Load Balancing 28 7
Load Balancers Wide- Area Network pread load over server replicas Present a single public address (VIP) for a service Direct each request to a server replica ervers Router" centers Router" ervers Virtual IP (VIP) 192.121.10.1 10.10.10.1 10.10.10.2 DN erver Internet 10.10.10.3 29 DN- based site selecron Clients 30 Wide- Area Network: Ingress Proxies ervers Router" centers Router" ervers Proxy Proxy Clients 31 8