Summary

  • RFC 7141 separates two decisions: network equipment should not scale an individual packet's marking or drop probability by size, while a transport should interpret a lost or marked packet in proportion to its octets.
  • The split avoids giving small packets a network-created discount without dictating a transport's response, promising equal bit rates or proving how widely the practice was deployed.

The same rate, twenty-five times as many packets

A packet is both a unit of work for a router and a container of bytes for a link. That difference became consequential in active queue management, where a network tries to signal congestion before a queue simply fills. In the late 1990s, Random Early Detection (RED) helped frame that debate. RFC 2309 discussed byte-mode dropping as a way to make packet loss reflect bandwidth more closely. Yet if a small packet receives a lower chance of being dropped, packet size itself becomes an advantage.

RFC 7141, published in 2014 as Best Current Practice, revisited the question. Its example compares two flows at 48 Mbps: one sends 60-byte packets, the other 1,500-byte packets. The smaller packets are 25 times as numerous. With a packet-mode drop probability of 0.1% for every packet, the small-packet flow loses more packets but, in the example, the two flows lose the same number of bits. Byte-mode drop lowers the small packet's probability 25-fold. It equalizes packet losses instead, while the small-packet flow loses far fewer bits. Those are consequences of the example's arithmetic, not a measurement of a live network.

The RFC's answer is deliberately split. In the network, the probability of dropping or ECN-marking a particular packet should not depend on that packet's size. RED's byte-mode drop variant should not be used: making small packets less likely to be marked can reward fragmentation and invite a flood of tiny packets. But the memo does not reject byte-mode queue measurement. A link limited by bits should measure its queue in bytes; a resource limited by packet-processing work may need packet counts. Measurement unit and mark-selection bias are different choices.

At the transport, RFC 7141 recommends reading a lost or marked packet's congestion indication in proportion to its octets—as if the indication applied to each byte. This lets an endpoint account for signal strength without asking shared network equipment to infer which transport's notion of fairness to enforce. The RFC calls this an interpretation goal. It does not command every transport to change its algorithm, and it expressly does not say TCP flows ought to reach equal bit rates. If equal rates are wanted, its advice is to alter TCP, not the network.

The distinction also matters when ECN-marked packets are split or merged: the recommended accounting conserves marked octets across the transformation, subject to compatibility with older protocols. RFC 7567 later replaced RFC 2309's general AQM recommendations and cited RFC 7141's split. RFC 7141's own survey of RED implementations was explicitly limited and found little, if any, installed base for byte-mode drop. The documents establish an engineering recommendation and a historical debate—not universal adoption, an observed attack or a present-day configuration.

Sources: RFC 7141 · RFC Editor record · RFC 2309 · RFC 2914 · RFC 3168 · RFC 7567 · RFC 5681 · RFC 6679 · RFC 3550 · RFC 2474