100G DAC vs AOC vs Optics: Selection & Troubleshooting - 夜莺博客

100G DAC vs AOC vs Optics: Selection & Troubleshooting

Half of the "why is this 100G link down" tickets in a data centre end with either a cable that was never going to work at that distance or a breakout that the port was never configured for. Copper and fibre assemblies look interchangeable on a bill of materials and behave nothing alike in the rack. This guide compares passive DAC, active DAC, AOC and optical modules, then covers the faults each one typically produces.

The four options compared

Attribute Passive DAC Active DAC AOC Optical modules + fibre
Media Twinax copper Twinax copper Fibre with fixed ends Fiber with pluggable ends
Power per end < 0.15 W 0.5–1 W > 1 W Depends on reach (3.5–6 W typical for 100G LR4)
Typical max reach at 100G 3 m 5–7 m 20–100 m 100 m MMF / 10 km+ SMF
Latency Lowest (pure wire) Very low Low, plus serdes/retimer hops Low, plus module DSP for coherent variants
EMI immunity Poor Poor Excellent Excellent
Failure diagnosability Low — no DOM Low Limited High — full DOM/DDM
Best fit ToR to server, adjacent racks ToR to adjacent rack, longer runs EoR/MoR, leaf-spine within a hall Cross-hall, campus, DCI

Read the reach column as a hard limit, not a suggestion. Passive copper beyond about 3 m at 100G produces exactly the kind of intermittent CRC errors that get misdiagnosed as a bad switch port for a week.

Breakout: the option people forget

A 100G QSFP28 port usually supports four 25G lanes (100G->4x25G) or four 10G lanes, which is why a single breakout assembly can replace four separate links. The catch is that the port must be configured for breakout mode before the cable is accepted, and the platform has to be asked for the right licence or feature:

# Arista EOS: split a 100G port into four 25G lanes
switch(config)# interface Ethernet49/1
switch(config-if-Et49/1)# speed forced 25gfull

# Cisco NX-OS: breakout profile on the port and reload of the module
switch(config)# interface ethernet 1/1
switch(config-if)# breakout module 1 port 1 map 4x25g-4x

# Mellanox MLNX-OS: the port must be in split mode
switch(config)# interface ethernet 1/1 speed 25G force

Mismatch here produces a link that is "up" but passing nothing, or four ports where the OS still believes it has one. Confirm with show interfaces status that the expected port count and speeds appear after the breakout before troubleshooting optics.

Coding and compatibility

  • Vendor coding. Many platforms only accept modules/assemblies coded for them. A "Cisco compatible" optic in an Arista port is not a supported combination even when the EEPROM is readable.
  • Standards vs vendor lock. Open standards (SR4, LR4, CR4) give you interoperability; vendor-coded parts give you guaranteed support. Decide which matters on that link before ordering.
  • Grade and temperature. Industrial-temperature parts cost more because they are tested beyond the commercial range; a hot ToR row is not the place for a commercial-grade module.
  • Wavelength-matched pairs. For duplex links both ends must match (LR4 with LR4); for parallel links the fibre type (OM3/OM4) and polarity both matter.

Diagnosing the faults each cable type causes

# CRC/FCS errors, clean optics, short copper run beyond reach
switch# show interfaces Ethernet1 counters errors

# Optics: DOM tells you whether the problem is local or remote
switch# show interfaces Ethernet1 transceiver
switch# show interfaces Ethernet1 transceiver dom

# Check both ends - a link works only if both sides agree
  1. Passive DAC, long run: intermittent CRC errors under load, no DOM data available to prove it. Fix by moving to active DAC or AOC.
  2. AOC, bend radius: link flaps when the rack door moves; the fibre inside the assembly has no DOM you can read.
  3. Optical module, dirty connector: receive power 3–6 dB below expected, errors climb as temperature rises. Clean, then re-measure.
  4. Polarity swap on parallel fibre: no light at all at one end — check the MPO key orientation.
  5. Breakout mismatch: port up, no traffic, or port never comes up despite good light.

Ordering rules of thumb

Within a rack or between adjacent racks: passive DAC, accepting the 3 m limit. Between racks in the same row: active DAC or AOC. End-of-row and leaf-spine within a hall: AOC, or optics if you must keep spare parts identical across the estate. Any run over 100 m, any cross-building or DCI link: optics, with DOM-based monitoring from day one. Standardising on the lowest-reach technology that actually covers the run keeps both the bill of materials and the troubleshooting surfaces small.

Related Reading on This Site

原文链接:https://equaloptics.com/dac-vs-aoc (Equal Optics - DAC vs AOC comparison)