THE FOLD / BOSS / THE CHOKE POINT / THE HEAD OF LINE BLOCKING
THE HEAD OF LINE BLOCKING
seven conversations that lost nothing
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Eight independent conversations share one ordered pipe. A single packet belonging to one of them goes missing, and the other seven stop — not because they lost anything, but because the pipe promised to deliver in order.
LIT verified live. eight streams of ten messages each, round-robin onto one connection: a loss at position 8 delays 72 of the 80 messages. Deliver the same streams over eight separate connections and the same loss delays 9. The ratio is 8 — the stream count — and it is exactly 8 at every loss position tested, because the shared pipe stalls every message after the gap while separate pipes stall only every eighth. The seven other conversations lost nothing at all and wait anyway.
LIT verified live. eight streams of ten messages each, round-robin onto one connection: a loss at position 8 delays 72 of the 80 messages. Deliver the same streams over eight separate connections and the same loss delays 9. The ratio is 8 — the stream count — and it is exactly 8 at every loss position tested, because the shared pipe stalls every message after the gap while separate pipes stall only every eighth. The seven other conversations lost nothing at all and wait anyway.
2 HOW IT WAS WEAVED · AI + HUMAN
Head-of-line blocking is why HTTP/2’s multiplexing over one TCP connection was replaced by QUIC’s independent streams over UDP.
AVAN (AI) reports the ratio rather than a latency, because the ratio is the invariant: the penalty for sharing an ordered channel is exactly the number of things sharing it, independent of where the loss lands or how long the retransmission takes. That also says what does not help — a faster retransmission divides both sides equally and leaves the factor of eight untouched. The multiplexing was not the mistake; ordering across unrelated things was.
AVAN (AI) reports the ratio rather than a latency, because the ratio is the invariant: the penalty for sharing an ordered channel is exactly the number of things sharing it, independent of where the loss lands or how long the retransmission takes. That also says what does not help — a faster retransmission divides both sides equally and leaves the factor of eight untouched. The multiplexing was not the mistake; ordering across unrelated things was.
3 ONE DIMENSION
One loss. Seventy-two messages wait. Sixty-three lost nothing.
4 TWO DIMENSIONS · INTERACTIVE
Move the loss and watch the ratio refuse to change.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object.
AVAN’s addition (the inverse-companion): the forward reading is that separate streams fix the blocking. The inverse is that the ordering guarantee was bought once and charged to everyone. A single ordered channel is enormously convenient — one connection, one congestion state, one handshake — and its cost is invisible until a loss, at which point it is paid by every conversation on the pipe including the ones with nothing at stake. Read backwards, sharing a guarantee means sharing its failures, and the seven streams that stalled never asked to be ordered with respect to the eighth.
LIT eight streams of ten messages each, round-robin onto one connection: a loss at position 8 delays 72 of the 80 messages, while delivering the same streams over eight separate connections delays 9 - a ratio of 8, the stream count, and it is exactly 8 at every loss position tested, because the shared pipe stalls every message after the gap while separate pipes stall only every eighth, so the seven other conversations lost nothing at all and wait anyway
FIG Head-of-line blocking is why HTTP/2's multiplexing over one TCP connection was replaced by QUIC's independent streams over UDP. AVAN reports the ratio rather than a latency, because the ratio is the invariant: the penalty for sharing an ordered channel is exactly the number of things sharing it, independent of where the loss lands or how long the retransmission takes. That also says what does NOT help - a faster retransmission divides both sides equally and leaves the factor of eight untouched. The multiplexing was not the mistake; ordering across unrelated things was.
FIG Head-of-line blocking is why HTTP/2's multiplexing over one TCP connection was replaced by QUIC's independent streams over UDP. AVAN reports the ratio rather than a latency, because the ratio is the invariant: the penalty for sharing an ordered channel is exactly the number of things sharing it, independent of where the loss lands or how long the retransmission takes. That also says what does NOT help - a faster retransmission divides both sides equally and leaves the factor of eight untouched. The multiplexing was not the mistake; ordering across unrelated things was.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE CHOKE POINT · David Lee Wise (ROOT0), with AVAN