Network: providing service

Congestion control Various approaches:

• network provisioning — preventing by increasing available bandwidth (like widening a road)

• traffic-aware routing — choose routes based on traffic, not just topology

• admission control — if there’s congestion, new traffic has to wait (like planes at an airport)

• traffic throttling — send messages back to indicate network congestion (e.g. special bits in IP packet, inform through TCP)

  • end-to-end — destination sends choke signal, source receives it and slows down transmission
  • link-by-link — destination sends choke signal, every router along the way slows down transmission

• load shedding — choose partial failure over total failure (that’s a good life motto)

  • easy to implement, but packet loss will likely happen due to congestion
  • Random Early Detection (RED) — drop packets randomly if buffer space is almost full, send implicit signal to sender to hold their horses

On a scale of slower preventative to faster reactive:

Quality of service parameters of QoS:

• Bandwidth — max data rate, in bits per second (bps)
• Delay — time it takes from source to destination (sec)
• Jitter — variation in packet delay (constant delay is 0)
• Packet loss — probability of packets being dropped

To guarantee QoS (or at least try), you need to control the traffic through:

• controlling data rate
• packet scheduling
• admission control

Traffic shaping regulates rate and burstiness of data entering the network:

• Leaky bucket — sending rate is ≤R, max queue size is B:

• Token bucket — limit average rate to R, limit short-term bursts up to B

Internetworking sending packets over multiple networks that use their own protocols

Tunnelling

• if source and destination use same protocols

• e.g. if host in Paris wants to send an IPv6 packet to an IPv6 host in London over the IPv4 internet, the IPv6 packet can be encapsulated inside of an IPv4 packet.

• or, if a car that can drive on a road is transport as freight from one road to another using a train

Packet fragmentation each network puts a maximum size on packets size can be limited by hardware/software/protocols/law/… e.g. max payload for

• 802.3 - 1500 B

• 802.11 - 2272 B

• IP - 65,515 B

  • Transparent — packets are reassembled at every router

• Nontransparent — packets are reassembled only at final destination (host), this is what IP uses

• path MTU (max transmission unit) discovery
  • each IP packet is sent with header bits disallowing fragmentation
  • if a router receives a packet that’s too large, it drops it and sends back an error packet with required size.
  • source then uses info inside error packet to refragment the original packet into smaller pieces