Back to ACN index

Goal of the lecture: what happens behind the scenes when accessing a webpage, such as https://google.com/.

• DNS resolution
  • Instead of DNS, there used to be a centrally managed /etc/hosts file
  • Of course not scalable, also not hierarchical - anyone could register any name
  • DNS is a distributed database, works by the principe of zone delegation (root -> party. -> segfault.party.)
• Once we have a destination IP:
  • Establish a connection, via the Berkeley "socket API"
  • OS wraps data in TCP segment, IP packet, Ethernet/802.11 frame
  • Link layer responsibilities:
    • Framing
      • Data-based or clock-based
    • Error detection (for performance; see also End-to-end argument#Examples). Different ways:
      • Parity checks (single bit, two-dimensional)
      • Checksumming (RFC 1071; used in TCP/UDP and IP)
      • Cyclic Redundancy Checks (used in Ethernet)
    • Medium access control
      • How to connect >2 computers?
      • Fully connected graph: obviously not scalable, (n choose 2) links
      • Shared medium/bus: problem is frame collision
        • Can be corrected (using, e.g. ECC or retransmissions)
        • We can use coordination methods that can reduce frame collisions
        • CSMA/CD, CDMA, etc.
      • Hub: broadcast frames on all links
      • Switch: send frame only on link to destination
        • Port ↔ Destination MAC address mapping (forwarding table)
        • Forwarding table can be generated automatically
        • Incoming frame on port x contains source MAC y → y is on port x
        • Entry is deleted after some time, and updated.
        • Store-and-forward vs. cut-through designs
          • Store-and-forward: buffer with frames
          • Cut-through: frames get sent right away after frame header is received
            • No error-checking possible
  • Network layer: routing
    • Routers are divided into two planes:
      • Control plane - software, manages forwarding tables
      • Data plane - hardware, uses forwarding tables to forward packets from link to link
      • How to make forwarding fast? As networks get into many-gigabit speeds, hash lookups or binary search gets too slow
        • TCAM -- Ternary Content-Addressable Memory; hardware mapping of IP address to forwarding table address
        • .
        • Ternary because the memory can contain wildcards, i.e. 192.168.xxx.xxx
      • Output port considerations:
        • Queue size
        • What happens when queues fill up?
        • Which packet should be sent first (scheduling)