A switch receives frames from 48 ports and its output queue fills. It drops the next frame. A link takes a burst of noise and a frame's CRC fails. The receiver drops it. Neither event is reported to the sender at the Ethernet layer. This is not a defect: Ethernet was designed to drop, and to let protocols above it (TCP, or the application) decide what loss means. That one decision shapes nearly all its other answers.
How does Ethernet answer each question, and what follows from choosing to be lossy?
Constraint | Ethernet |
|---|---|
C1 Distance and medium | Metres to kilometres: copper, backplanes and fibre |
C2 Bandwidth and latency | From megabits to hundreds of gigabits per second; latency varies with distance and switching |
C3 Wires | A few pairs or fibres; serial |
C4 Topology | Switched networks of any size (historically a shared medium, fact S16) |
C5 Who builds each side | Thousands of vendors; IEEE 802.3; interoperability tested independently (fact S26) |
Each answer links to the lesson that explains the idea behind it.
# | Question | Answer |
|---|---|---|
1 | When is the data valid? | Clock embedded in the data and recovered; high-speed PHYs use 64b/66b blocks and scrambling (facts S9, S13) (1.4) |
2 | Where does a message start and end? | Preamble and start-of-frame delimiter, then the frame, ended by the FCS and an interframe gap (facts S5, S10) (2.2) |
3 | Is the receiver ready? | No per-frame handshake: senders transmit, receivers accept or drop (3.1) |
4 | How much may I send before I hear back? | Lossy by default: full buffers drop frames. Optional PAUSE (802.3x) and priority flow control (802.1Qbb) make links lossless when enabled (facts S24, T12) (4.12) |
5 | What exactly is being asked? | Frames, not requests: meaning and any request/response belong to higher layers (0.7) |
6 | How many requests can be in flight? | Many frames in flight across the network; end-to-end windows (TCP) limit how many (fact T7) (4.9) |
7 | In what order do results become visible? | In order on a single link; across a network, ordering and delivery are restored end to end by higher layers (7.2) |
8 | Who gets the wire? | Switched full duplex today; classic shared Ethernet used CSMA/CD (fact S16) (8.1) |
9 | Why can it hang? | Dropping frames prevents buffer deadlock; enabling PFC brings back pause storms and deadlock risk in large fabrics (fact T12) (9.6) |
10 | What if bits go wrong? | CRC-32 detects errors and the frame is dropped; FEC corrects on high-speed lanes; retransmission is end to end (facts S5, S7, S17) (10.14) |
11 | How do the two sides agree at start-up? | Auto-negotiation selects the highest common mode; backplanes also train equalisation (fact S26) (11.5) |
12 | How do we save power without losing data? | Energy-Efficient Ethernet's Low Power Idle (fact S18) (12.2) |
13 | What if there are several copies of one value? | Not applicable |
14 | Who is allowed to ask? | MACsec encrypts and authenticates frames hop by hop (fact S20) (14.4) |
15 | How do we know it is correct? | IEEE 802.3 conformance plus independent interoperability testing (fact S26) (15.8) |
Lossy by design: drop on error, drop on congestion, and let the end points recover (the end-to-end argument, lesson 0.7).
Detection without correction at the frame level: a CRC-32 decides keep or drop; FEC appears only where the raw error rate demands it.
Losslessness as an option: PAUSE and PFC exist for traffic, such as RDMA, that cannot tolerate drops, at the cost of congestion spreading.
A data-centre network (illustrative) enabled priority flow control to carry RDMA traffic without drops. A misbehaving server stopped draining its receive queue and kept sending PAUSE frames; pauses propagated switch by switch until large parts of the fabric stopped for that priority. Watchdogs that disable PFC on a port that pauses for too long, and careful limits on which traffic is lossless, contained it.
It is designed to be lossy and relies on end-to-end recovery; occasional drops are acceptable to the layers above.
Drops it, usually counting the error; nothing is sent back.
Ethernet is lossy by design: bad or unbufferable frames are dropped, recovery is end to end.
Frames are found by preamble and SFD and checked by CRC-32; FEC on high-speed lanes.
PAUSE and PFC make links lossless when needed, at the cost of spreading congestion.