§ 1.1  Module 1 — When Voice Became Data

The Tyranny of the Circuit

Mr. Watson — come here — I want to see you.— Alexander Graham Bell, 1876, over a circuit that guaranteed his timing

Surviving hardwarephotographed 2010

1.1.1A reserved path through the world

The telephone network — the PSTN, built on circuit switching — divides every call into two phases. First comes setup: the network finds and reserves a path of wire segments and switch contacts from one handset to the other. In 1878 the switch was a person with a cord board; from 1892 it was increasingly Almon Strowger's electromechanical stepping switch (patented 1891, first installed in La Porte, Indiana, in 1892 — Strowger was a Kansas City undertaker who suspected operators of routing his customers to a competitor), later the crossbar switch (Bell System No. 1 Crossbar, 1938), and eventually computers signaling to each other over a separate control network (Signaling System No. 7, standardized in 1980). Then comes the conversation phase, during which the network does nothing at all. The path is nailed up; electricity flows through it as if the two telephones were joined by a single wire.

It is worth being precise about what this buys, because the rest of this course is an attempt to buy it back. On an established circuit, end-to-end delay is fixed at essentially the propagation time of the medium — roughly 5 µs per kilometer in copper, about 20 ms coast to coast across the United States — plus a small, constant transit time through each switch. There is no queueing, because nothing shares the path. There is no reordering, because a single path cannot overtake itself. There is no loss short of physical failure, because no buffer ever overflows. The circuit does not merely make jitter small; it makes jitter impossible by construction.

The price is exclusivity. A circuit is consumed by a call whether or not anyone is speaking, and conversational speech is astonishingly sparse: each direction of a two-way conversation carries voice energy only about 40 percent of the time. The other 60 percent — pauses, listening, breaths — occupies reserved capacity that nobody else may touch. Worse, the reservation is all-or-nothing: a network engineered to hold thousands of simultaneous 4 kHz voice paths open for minutes at a time is structurally incapable of carrying the traffic computers generate, which arrives in bursts of microseconds separated by silences of seconds. The circuit architecture did not merely fail to anticipate the web; it was arithmetically incompatible with it.

From the archive

1.1.2Time-division multiplexing and the DS0

Reserving a physical copper pair per call was tolerable inside a town and ruinous between towns, so the Bell System multiplexed its trunks — first in frequency (analog carrier systems, beginning with the Type A carrier in 1918), and then, decisively, in time. In 1962, Illinois Bell put the first T1 carrier system into commercial service between central offices in Chicago: the first digital transmission system in the public network, designed at Bell Labs, and the ancestor of the framing inside every digital telephone trunk since.

T1 rests on two numbers derived a generation earlier. Harry Nyquist's 1928 analysis of telegraph transmission established that a band-limited signal is fully captured by sampling at twice its highest frequency; the telephone voiceband, filtered to 300–3,400 Hz, therefore needs 8,000 samples per second. Each sample is quantized to 8 bits using companded pulse-code modulation — an idea patented by Alec Reeves in 1938 and the entire subject of the next lesson. Eight bits, 8,000 times per second: 64,000 bits per second. That channel is the DS0, and it is the atom of the digital telephone network — the reason "64 kbps" recurs through forty years of telecom standards.

A T1 interleaves 24 DS0s by time. Each frame carries one 8-bit sample from each of the 24 channels plus a single framing bit: 24 × 8 + 1 = 193 bits, repeated every 125 µs — 8,000 frames per second, 1.544 Mb/s on the wire. Notice what the frame does not contain: no addresses, no headers, no lengths, no checksums. A sample's position within the frame identifies which call it belongs to, and the frame's fixed recurrence supplies the timing. Everything a packet network must spend header bytes and buffer memory to express, time-division multiplexing (TDM) expresses with position and clock — which is precisely why it can guarantee timing, and precisely why it cannot share.

From the archive1905

1.1.3Erlang's arithmetic: paying with blocking

If circuits are reserved, how many trunks must a telephone company build? The question was answered by Agner Krarup Erlang, a mathematician at the Copenhagen Telephone Company, in a line of work beginning with his 1909 paper The Theory of Probabilities and Telephone Conversations and culminating in his 1917 blocking formula. Erlang modeled call arrivals as a Poisson process and defined offered load as a = λh: the call arrival rate multiplied by the mean holding time. His "B" formula gives the probability that a call arrives to find all c trunks busy — and is simply refused, with a busy signal. The refused caller does not wait; on a circuit network there is nowhere to wait.

function erlangB(a, c) {   // a: offered load in erlangs, c: trunks
  let b = 1;               // zero trunks: every arriving call is blocked
  for (let k = 1; k <= c; k++) {
    b = (a * b) / (k + a * b);
  }
  return b;                // P(all c trunks busy) for Poisson arrivals
}

Listing 1.1 — Erlang's 1917 blocking formula, computed by its standard recurrence.

The numbers explain a century of telephone pricing. Offer 5 erlangs of load to 10 trunks and about 1.8 percent of calls are refused — the classic "P.02" grade of service that Bell engineers provisioned toward. Push blocking down to a tenth of a percent and you need 13 trunks: three additional circuits, purchased and maintained, that stand idle nearly all the time. The circuit network's quality is bought with capacity that mostly does nothing, and its failures are absolute — the unlucky 1.8 percent get no service whatsoever.

The packet alternative spends a different currency. A shared link refuses no one: every arriving packet is admitted and queued. The cost is delay, and it is paid by everyone. For a link at utilization ρ serving fixed-size packets of service time S, the M/D/1 queueing model gives a mean wait of W_q = ρS / 2(1 − ρ). At half utilization that is half a service time — negligible. At ρ = 0.96 it is twelve service times, and as ρ approaches 1 it grows without bound. Leonard Kleinrock worked out this style of analysis for communication networks in his 1961–62 MIT dissertation, published as Communication Nets in 1964 — queueing theory arriving in networking a few years before the network it would describe.

Here is the whole tradeoff in one sentence: the circuit refuses some users entirely so that admitted users experience perfection, while the packet network admits everyone and degrades continuously under load. And the economics point one way. Because each direction of a conversation is idle roughly 60 percent of the time, a statistically shared link can carry more than twice the conversations of its reserved equivalent before quality collapses. Packets won on cost. They lost — catastrophically, as the next section shows — on timing. The exercise at the end of this lesson has you build both models and watch the two currencies move.

From the archive1966

1.1.4Toll quality: the bar packet voice had to clear

"Toll quality" is not a metaphor. Toll circuits were the long-distance trunks customers paid per-minute tolls to use, and their audio quality — 64 kb/s companded PCM, the codec later standardized as G.711 (CCITT, 1972) — became the reference against which all telephone audio is judged. The measuring instrument is the mean opinion score of ITU-T P.800: panels of listeners rate speech from 1 (bad) to 5 (excellent), and toll quality means scoring about 4.0, where G.711 itself scores roughly 4.1. Delay has its own budget: ITU-T G.114 recommends at most 150 ms one-way before conversational interactivity degrades — people begin talking over each other. The analog and digital PSTN sailed under that bar without trying; a domestic long-distance circuit's delay, being propagation plus switching, ran a few tens of milliseconds.

When the first mass-market packet voice product appeared — VocalTec's Internet Phone, released by Alon Cohen and Lior Haramaty in February 1995 — it was, by these measures, not a telephone. It ran half-duplex, walkie-talkie style, over dial-up modems, with latency measured in large fractions of a second. Its existence proved demand; its quality defined the gap. The bar packet voice had to clear was never "intelligible." It was "indistinguishable from a wire that a century of engineering had made perfect," on a network that guarantees nothing.

1.1.5The thesis: timing guarantees versus statistical sharing

Packet switching was invented, deliberately, by discarding the circuit's central assumption. Paul Baran's 1964 RAND report series On Distributed Communications proposed chopping messages into independently routed blocks so a network could survive damage; Donald Davies at the UK's National Physical Laboratory arrived at the same design in 1965 and gave the blocks their lasting name, packets. On October 29, 1969, in Leonard Kleinrock's lab at UCLA, student Charley Kline sent the first ARPANET message toward SRI — the system crashed after two characters, "LO". Each packet in such a network is routed on its own and queued at every hop. For bursty computer traffic this is exactly right. For voice it resurrects, all at once, everything the circuit had abolished: variable delay, reordering, and loss.

The obvious reflex — run voice over TCP and let retransmission fix the loss — makes matters worse, which is why real-time media instead rides UDP, the transport that promises nothing and therefore never delays you.

Packet voice itself is nearly as old as the ARPANET. In 1973–74 Danny Cohen at USC's Information Sciences Institute ran the first packet voice experiments across the ARPANET under the Network Voice Protocol, later written up as RFC 741 (1977). NVP proved the physics — compressed speech could cross a packet network and be reassembled — but offered no standard framing, so every experiment was bespoke and nothing interoperated. The durable answer took until January 1996: RTP, RFC 1889, by Henning Schulzrinne, Stephen Casner, Ron Frederick, and Van Jacobson. Its two crucial header fields — a media-clock timestamp and a sequence number — are exactly the information TDM had carried implicitly in slot position and frame recurrence, now made explicit so a receiver can reorder and re-time what the network scrambled. Its companion protocol RTCP reports back what the network did to the stream. Both sit inside every WebRTC packet you will ever send.

This is the course's central tension, and you now hold both ends of it. The circuit guarantees timing by reserving physics and pays in blocking, idle capacity, and per-minute-per-mile prices; the packet network shares statistically, wins on cost by a wide margin, and surrenders every timing guarantee. Everything that follows in this book — RTP timestamps, jitter buffers, congestion control, retransmission that races the playback clock — is one project: rebuilding the circuit's guarantee in software, statistically, at the edges of a network that refuses to make it. The next lesson begins at the bottom of that rebuild, with the DS0's own question: how a pressure wave in air becomes 64,000 bits per second.

Problems for § 1.1Check your understanding

Build both economic models — circuit blocking and packet queueing — and watch each architecture pay in its own currency; then verify the circuit's guarantees against scenarios.

¶Field quirks

What the specifications say, and what the browsers actually do, are two different documents. These are the divergences that cost production teams their weekends.