THE FOLD / BOSS / THE WALL / THE BUFFERBLOAT
THE BUFFERBLOAT
the drop was the signal; the buffer is what silenced it
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Memory got cheap, so buffers got large, so nothing is ever dropped — and packets sit in a queue for a second instead of being discarded in a millisecond. The link is not slower. The wait in front of it is enormous.
LIT verified live. One bottleneck serving 1 packet per ms, offered 52,095 packets over 40 seconds — more than it can carry. With a 1,000-packet buffer the mean queueing delay is 957.6 ms. With a 10-packet buffer it is 10.0 ms — 95.8× less. Throughput is identical to four decimals: 1.0000 either way, a difference of 0.00%. The hundredfold delay bought nothing at all.
LIT verified live. One bottleneck serving 1 packet per ms, offered 52,095 packets over 40 seconds — more than it can carry. With a 1,000-packet buffer the mean queueing delay is 957.6 ms. With a 10-packet buffer it is 10.0 ms — 95.8× less. Throughput is identical to four decimals: 1.0000 either way, a difference of 0.00%. The hundredfold delay bought nothing at all.
2 HOW IT WAS WEAVED · AI + HUMAN
Jim Gettys named bufferbloat in 2010 after chasing terrible latency on his own home link; Kathleen Nichols and Van Jacobson’s CoDel is the standard answer.
AVAN (AI) ran both buffers over the same offered load so the throughput column would be directly comparable, which is the whole argument. My first model offered arrivals at exactly the service rate — a queue that mathematically cannot build — and reported a delay ratio of 1.0, a clean pass proving nothing. A buffer only fills when the offered load exceeds the link, so the experiment has to be run in overload or it is not the experiment.
AVAN (AI) ran both buffers over the same offered load so the throughput column would be directly comparable, which is the whole argument. My first model offered arrivals at exactly the service rate — a queue that mathematically cannot build — and reported a delay ratio of 1.0, a clean pass proving nothing. A buffer only fills when the offered load exceeds the link, so the experiment has to be run in overload or it is not the experiment.
3 ONE DIMENSION
Same link, same load, two buffer sizes.
4 TWO DIMENSIONS · INTERACTIVE
Resize the buffer. Watch delay move and throughput stay.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: a queue nobody meant to build.
AVAN’s addition (the inverse-companion): the forward reading is that oversized buffers cause latency. The inverse is that the drop was the signal, and the buffer is what silenced it. Loss is how a sender is told to slow down; a deep buffer absorbs the packets that would have carried that message, so the sender keeps accelerating into a queue it cannot see. Read backwards, this is not a memory-sizing mistake but a failure of nerve — the buffer was added to avoid discarding data, and discarding data was the only working feedback channel in the system.
LIT one bottleneck serving 1 packet per ms, offered 52,095 packets over 40 seconds: with a 1,000-packet buffer the mean queueing delay is 957.6 ms and with a 10-packet buffer it is 10.0 ms, 95.8 times less, while throughput is identical to four decimals at 1.0000 either way - a difference of 0.00%, so the hundredfold delay bought nothing
FIG Jim Gettys named bufferbloat in 2010 after chasing terrible latency on his own home link; Nichols and Jacobson's CoDel is the standard answer. AVAN ran both buffers over the same offered load so the throughput column would be directly comparable, which is the whole argument. My first model offered arrivals at exactly the service rate - a queue that mathematically cannot build - and reported a delay ratio of 1.0, a clean pass proving nothing. A buffer only fills in overload, so the experiment has to be run in overload or it is not the experiment.
FIG Jim Gettys named bufferbloat in 2010 after chasing terrible latency on his own home link; Nichols and Jacobson's CoDel is the standard answer. AVAN ran both buffers over the same offered load so the throughput column would be directly comparable, which is the whole argument. My first model offered arrivals at exactly the service rate - a queue that mathematically cannot build - and reported a delay ratio of 1.0, a clean pass proving nothing. A buffer only fills in overload, so the experiment has to be run in overload or it is not the experiment.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE WALL · David Lee Wise (ROOT0), with AVAN