lecture22-networking.md

Lecture 22 Networking

Network Layers

OSI 7 Layer Model

OSI vs. TCP/IP

Life of a Packet

Layer Encapsulation

Building Larger LANs: Bridges

• Extend reach of a single shared medium
• Connect two or more segments by copying data frames between them
  • Only copy data when needed -> key difference from repeaters/hubs
  • Reduce collision (sending at the same time) domain compared with single LAN
  • Separate segments can send at once -> much greater bandwidth
• Challenge: learning which packets to copy across links

Frame Forwarding

• A machine with MAC Address lies in the direction of number port of the bridge
• For every packet, the bridge "looks up" the entry for the packet's destination MAC address and forwards the packet on that port
• Timer is used to flush old entries

Learning Bridges

• Keep track of source address of packets arriving on every link, showing what segment hosts are on
  • Fill the the forwarding table based on this information

Internet

IP to MAC Address Translation

• Address Resolution Protocol (ARP)
  • Broadcast search for IP address
  • Destination responds (only to requester using unicast) with appropriate 48-bit Ethernet address
• Caching ARP Entries
  • Efficiency concern
  • Each host maintains cache of ARP entries

Structure of the Internet

• Ad hoc interconnection of networks
  • No particular topology
  • Vastly different router and link capacities
• Send packets from source to destination by hopping through networks

Addressing in IP

• IP addresses are names of interfaces
• Domain Name System (DNS) names are names of hosts
• DNS binds host names to interfaces
• Routing binds interface names to paths
• IP Addresses
  • Fixed length: 32 bits
  • Class A (0): 128 networks, 16M hosts
  • Class B (10): 16K networks, 64K hosts
  • Class C (110): 2M networks, 256 hosts

Original IP Route Lookup

• Address would specify prefix for forwarding table
  • Simple lookup
• Forwarding table contains
  • List of class + network entries
  • A few fixed prefix lengths (8/16/24)
• End up with larage tables

Tunneling - Forcing Particular Routes

• Force a packet to go through a specific point in network
• Achieved by adding an extra IP header to packet with a new destination address
• Used increasingly to deal with special routing requirements or new features
  • Mobile IP, Multicast, IPv6, research

Extending Private Network

• Virtual Private Network (VPN)
  • Overlays private network on top of regular Internet
  • Create tunnel between client & firewall
  • Remote client appears to have direct connection to internal network

Circuit vs. Packet Switching

• Circuit Switching
  • Fast switches can be built relatively inexpensively
  • Inefficient for bursty data
  • Predictable performance (hard QoS)
  • Requires circuit establishment before communication
• Packet Switching
  • Switch design is more complex and expensive
  • Allows statistical multiplexing
  • Difficult to provide QoS guarantees
  • Deta can be sent without signaling delay and overhead
• Virtual Circuit Switching
  • Use short connection identifiers to forward packets
  • Switches know about the connections so they can more easily implement features such as quality of service
  • VC identifies long-lived stream of data that can be scheduled

Fat-Tree Topology

• K-ary fat tree: three-layer topology to inter-connect racks (of servers)
  • Edge switches: connected to k/2 servers and k/2 aggregate switches
  • Aggregation switches: connected to k/2 edge and k/2 core switches
• Fixed routing: increase rack-rack bandwidth
• Shaped traffic: maximize bandwidth for dynamic traffic patterns
• Why Fat-Tree?
  • Fat tree has identical bandwidth at any bisections
  • Each layer has the same aggregated bandwidth
  • Can be built using cheap devices with uniform capacity
• Great scalability: k-port switch supports $k^3/4$ servers
• But wiring complexity in large networks

Cloud/Data Center Networking

• Large data centers have 100s or 1000s of machines
  • 10x of VMs each
  • 1000s of network boxes
• Heavy use of Layer 2
  • Cheaper, faster than IP
• Dependence on Layer 3
  • For thigns not in same Layer 2 domain

Data Center Routing Issues

• Poor server-to-server connectivity
  • Layer 3 equipment (large buffers, marketplace factors) more expensive than Layer 2
• Static network assignment
  • Host IP addresses assigned according to physical switch connections
  • Migrating a VM changes its IP, breaks TCP and app state
• Fragmentation of resources
  • Load balancing techniques often require all IPs in same pool
  • VLANs: often policy over-loaded, hard for apps to span VLANs efficiently
• Proprietary hardware scales up, not out
  • Conventional load balancers paired up in 1+1 resiliency configs (see Fat-Tree)
  • These pairs tend to replaced togethen when load exceeds capacity

Build Data Center VLANs

• LANs and VLANs must support broadcast
  • Easy configuration because forwarding tables self discovered
  • Same IP during VM migration between (V)LAN switches
  • Avoid big (V)LANs -> big tables, lots of broadcast traffic
    • Millions of VMs in same VLAN is rarely supported (usually 100s-1000s)
• Layer 2 solves forwarding loops with spanning tree
  • One of many paths is selected for all traffic A -> B
  • Bad for bandwidth, underutilizes paths and links
• Switches must be configured with per-VLAN bandwidth
  • Why: VLANs spanning networks force too much traffic high in tree
  • This makes changes complex, often manual
• Scalability: configure all switches/routers
  • Vertically-integrated switch/router -> Admin needs to change every configuration manually
  • Software-Defined Networking (SDN)