Messages in this thread | | | Date | Fri, 11 Sep 1998 20:36:49 -0400 (EDT) | From | Gregory Maxwell <> | Subject | Re: my broken TCP is faster on broken networks |
| |
On Sat, 12 Sep 1998, Martin Mares wrote:
> > This is not really a valid argument. Just because the people making web > > browsers are stupid doesn't mean that the technically sound solution isn't > > to use UDP. Someone could write a LGPLed libuhttp.so > > Is there any _technical_ advantage in using UDP over using persistent > connections over TCP?
Sure there is: http connections are frequently small in time, even with persistant connects (because people hop sites). Well designed sites are also small in size.
Congestion control can also persist across connections, web servers could have congestion 'buckets' that track the amount of congestion to various subnets (as calculated from RTTs and dropped packets). Web requests could be sent as small packets which routers can be programmed not to drop except as a 'last' resort (below telnet and ssh, etc)..
Images could be sent in a dynamicly lossy format: If encoded with a ranked subband encoding (like most wavlet encoders) the images is seperated into segments of importance (largest scale lumenence is most importnat) and the important parts are sent with reduncy. Less important info can be dropped, and the image quality degrades a bit. The web browser could 'right click' on the image and select full quality if they need that. Users/sites could request a minimum quality, and if that isn't met the browser could request the missing parts. Images could be predictivly shrunk (the server would keep stats on congestion) so that it wont bother trying to send the 'detal' parts to people in japan unless they ask for them.
Users could request a 'target' quality: Say the users says they want 75% quality images, that request goes to the server which then looks up congestion stats for that users network and sends the right amount to get something about 75% quality.
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/faq.html
| |