Summary

  • RFC 3194 replaced RFC 1715's hard-to-explain H ratio with Host Density: the logarithm of allocated objects divided by the logarithm of the maximum allocatable objects. The 0–1 scale made comparison easier, but its percentage was not ordinary utilization.
  • Three historical address plans reached painful change points near 87% HD even though only 3.2%, 10% and 22.8% of their literal spaces were occupied. The clustering was an empirical capacity indicator, not a universal law, a physical exhaustion threshold or proof that allocated objects were active.

A percentage arrived carrying a different denominator

Address plans invite a simple division. Count the objects assigned, divide by all possible addresses and call the result utilization. For a flat container, that might answer the question. Telephone numbers and Internet addresses were not flat containers.

They were divided into levels. A telephone prefix marked a region, another portion an exchange and the remaining digits a subscriber. A network address could allocate a block to a provider, subdivide it among organizations and divide an organization's block into subnets. Each level needed slack because growth and mobility were uneven. A region could fill while another remained sparse; a subnet could need one more host after its neighbor had reserved thousands.

Trying to reclaim every gap would demand repeated renumbering. RFC 3194 described that consequence as pain borne by operators and subscribers. Capacity was therefore not just the arithmetic number of unused strings. It was the number of objects that could be added before the hierarchy became too costly to rearrange.

This was why the familiar percentage in the memo needed careful reading. Host Density did not divide assigned objects by raw address count. It divided their logarithms:

HD = log(allocated objects) / log(maximum allocatable objects)

Any logarithm base produced the same ratio if used on both sides. HD began at zero when only one object was allocated and reached one only when every available address was assigned. Between those endpoints, however, 87% HD did not mean 87 addresses out of every hundred were occupied. It meant the allocation count had travelled 87% of the logarithmic distance between one object and the size of the whole space.

Hierarchical losses multiplied rather than added

The authors' concern came from the structure of allocation. If three levels were each 50% efficient in the ordinary sense, the whole system was not 50% full. Its direct occupancy was 0.5 × 0.5 × 0.5, or 12.5%.

That multiplication made raw occupancy a poor comparison between plans with different widths and levels. A large space could look almost empty while several local partitions were already difficult to grow. A smaller plan with a different hierarchy could tolerate a higher direct utilization.

RFC 1715 had addressed this problem during the earlier IP next-generation debate. Its H ratio used the base-10 logarithm of allocated objects divided by the address length in bits. The measure reduced dependence on the nominal size of the space, but it occupied an awkward numerical range from zero to about 0.30103. Values such as 0.25 did not carry an intuitive scale.

RFC 3194 rescaled the idea. In most cases HD was H divided by log10(2). The new output ran from zero to one and could be said as a percentage. The change did not create new capacity. It changed the representation of an earlier model so that people could compare and discuss it more easily.

That gain in communication created a new risk. A percentage normally suggests an ordinary part-to-whole fraction. The authors meant a quotient of logarithms. The clearer label could become less clear if the denominator disappeared from the explanation.

Three crises clustered only after the logarithmic transformation

The memo reused three examples from RFC 1715. France added a digit to its telephone numbers after reaching roughly ten million phones in an eight-digit plan. The United States expanded an effective 9.2-digit plan toward ten digits around one hundred million subscribers. A globally connected DECnet Phase IV network stopped visible growth around 15,000 nodes in a 16-bit space, with new nodes hidden.

Their ordinary occupancy ratios did not resemble one another: 10%, 3.2% and 22.8%. Those numbers offered no obvious shared threshold. Their HD values did: approximately 87.5%, 87.0% and 86.7%.

RFC 3194 treated the cluster as evidence that the metric captured something about practical pain. Operators delayed renumbering or redesign until the existing plan was difficult enough to justify drastic action. The ratio immediately before change might therefore estimate what a hierarchy could sustain in practice.

The inference remained limited. Three examples did not turn 87% into a natural constant. They mixed telephone and computer networks, different institutional choices and different meanings of an allocated object. The document used phrases such as “suggest”, “hypothesize” and “on the order of”. Its claim was that HD produced a useful similarity where direct occupancy did not, not that every network would fail at the same number.

Renumbering moved the indicator without filling or emptying the world

The telephone examples also showed what happened after adding capacity. The French plan moved from 87.5% HD under eight digits to 77.8% under nine. The US plan moved from roughly 87% to 80% when its effective capacity expanded.

No subscribers had to disappear for the ratio to fall. The denominator changed. The same approximate population occupied a larger address plan, so its logarithmic position moved down. The memo read the resulting 78–80% range as a more comfortable trade-off between spare structure and operational efficiency.

This was not cost-free abundance. New digits had to be communicated, equipment and directories updated, and old expectations retired. HD described the pressure that made such a change attractive. It did not measure the work of the transition or decide who paid for it.

The metric also separated one kind of capacity from others. Adding address bits did not automatically reduce routing state, registry contention, administrative fragmentation or application dependence on old identifiers. A plan could have low HD and still suffer from a poorly chosen hierarchy. It could exceed a suggested threshold because operators preferred local workarounds to global renumbering.

The IPv4 row exposed the difference between density and occupancy

RFC 3194 inverted its formula to estimate how many objects corresponded to a chosen HD:

allocatable objects = (maximum allocatable objects)^HD

For a 32-bit address space, the example table gave about 51 million objects at HD 80%, 154 million at 85%, 192 million at 86% and 240 million at 87%. It labelled the bands reasonable, painful, very painful and practical maximum.

The last label is the most tempting to misquote. Two hundred forty million is only about 5.6% of the nominal 2^32 space. The plan could therefore be “87%” along the Host-Density scale while more than 94% of its literal address values remained unassigned in this simplified calculation.

There was no contradiction. Exponentiation translated the logarithmic indicator back into an object count. The large apparent gap represented the structural margin that hierarchical allocation might consume at multiple levels. It was not a statement that the missing addresses had physically vanished.

The text also refused to turn 240 million into a hard ceiling. Practical maximum depended on the pain users and providers would tolerate. The Internet might exceed 154 million allocated IPv4 addresses if it accepted greater pain. A threshold described a negotiated operating condition, not a protocol field that stopped allocation.

The ratio measured one representation of pressure

By 2001, the 128-bit size of IPv6 had already been chosen. RFC 3194 did not reopen that decision. It offered a way to discuss the practical capacity of an address plan and to compare differently sized systems on a common scale.

Later documents on CIDR, registry practice and IPv6 site assignment dealt with controls the ratio did not contain. They decided how blocks were aggregated, who received space, how much hierarchy was desirable and what operational considerations applied. HD could inform a capacity conversation, but it could not exercise those authorities.

Nor did an “allocated object” necessarily mean an active host. It might be a subscriber, a visible node, an assigned address or another unit defined by the study. The ratio did not observe whether an address was routed, responding, transferred, reserved or economically scarce. Mixing those quantities would transform a comparison tool into false evidence about the network.

RFC 3194's historical value lies in the distinction. It recognized that unused symbols are not all operationally interchangeable inside a hierarchy. It also showed that a normalized percentage can conceal its own semantics. The number became useful only when readers preserved the formula, the chosen object, the maximum space, the hierarchy and the human decision called “too painful”.