| From: David Collier-Brown via talk <talk@gtalug.org> | On 22/08/16 02:05 PM, James Knott via talk wrote: | > On 08/22/2016 01:37 PM, David Collier-Brown via talk wrote: | > > Bufferbloat happens where a slower link meets a faster one, and the | > > symptoms are indeed insane latency when doing anything that has to | > > pass through the same buffer as a bulk transfer. | > TCP should detect the slow transfer and back down the transmit rate. Yeah. That should be detected by lack of timely ACKs. | > When buffers overflow, they're supposed to drop packets. | *We* know that, but people doing routers don't. They coded to never drop | packets, and added massive buffering to support that. Net result? They broke | TCP. Am I right that if the bottleneck is beyond his demark, and it the flow is predominantly inbound, code on his own router cannot affect the problem? Well, perverse code could: for example send ACKs for packets not yet seen.