Edward Ned Harvey wrote: >> The mistaken quest to never drop packets has destroyed interactivity >> under load... > > For what it's worth, I completely disagree with this...
Did you attend the talk, and if so, did your opinion change? > At the mac layer, when the receiver buffer is getting sufficiently > full, the receiver (router) sends the PAUSE frame (or newer more > powerful alternatives) to the sender, on the LAN. Jim mentioned that many enterprise routers/switches either don't send these flow control signals, or ignore them when received. In order for a flow control mechanism to work, it needs to be implemented all the way to the end points. I'm assuming the enterprise hardware vendors gave up on this option because they found the protocol was rarely implemented on the endpoints. > The so-called "congestion notification via packet drop," is a poorly > executed form of flow control. It does seem less than ideal... > It is fatal to DNS and many other stateless protocols, like XMPP, > SMS, etc. ...but if implemented correctly, it actually helps these protocols by making sure these packets don't get piled up behind big queues of TCP packets. If done right, only your "bulk" traffic will end up seeing dropped packets. Little things like DNS, and even relatively low bandwidth streams of VoIP, will pass through with minimal delay. I didn't catch whether Jim was attributing this to CoDel or another AQM algorithm (it might have been part of his description of what the best WiFi networks at technical conferences use), but he described a queue management algorithm that automatically splits up traffic into independent queues and stateful TCP connections each get their own queue, and thus a large, high-bandwidth file transfer will result in those packets piling up in a queue, while small stateless packets pass through, even when the buffers are full. -Tom -- Tom Metro Venture Logic, Newton, MA, USA "Enterprise solutions through open source." Professional Profile: http://tmetro.venturelogic.com/ _______________________________________________ bblisa mailing list [email protected] http://www.bblisa.org/mailman/listinfo/bblisa
