Under the thread "Size limits", Oskar wrote: > I never got a straight answer out of anybody about choking the > incoming connection. Is it a healthy thing to do? Will it overflow > your OS buffers and screw with the machine in general or will TCP > handle it graciously?
This morning, I asked a guy at the office here about it. Under the caveat that he's not a TCP guru, he said that in *theory*, the buffers get filled, and the sending machine will simply block the sending process until the buffers clear up again. No problem, unless the sending process is sensitive to long blocks. I've heard that Java's threads make Freenet resistant to such problems, so we're OK there. In practice, however, he said it "doesn't always work". That makes sense, given that full buffers and blocked processes rank high on the list of things that tend to bring out problems. No doubt there's also a lot of card drivers, OS parameters, routers, firewalls, and other junk with their finger in the pie. Each component may have differing assumptions about what a good state of affairs should look like, and when & how to give up on bad state. He said the cure would be to use the cool new Quality of Service (QoS) API's. Unfortunately, he continued, they are not done yet. Over lunch, I did a fast grep on Google for the words "TCP", "bandwidth", and "throttling". The response included several pages on the great new buzzword "QoS", and how it's supposed to fix lots and lots of problems -- which weakly implies that there are indeed lots and lots of problems to be fixed. After this decidedly weak research, it still sounds like the only way to do reciever bandwidth throttling in the short run is to hack together some crude QoS declarations and buffer management of our own. A simple example is recievers sending a "dude! slow down, I'm on a 28.8 modem here" message, and senders all being nice enough to comply. Yes, I'm cringing at the concept, too -- though I'd gladly take the assignment if it were given to me. I still want to see true bandwith throttling, for several reasons. I'm the guy who even wanted throttling to occur in response to outside messages/events. Down the road I'll probably take up the issue again, but for now, throttling recieves still seems outside the scope of 0.3. If there's a true TCP guru out there who disagrees, please speak up. --Will (never speaking for his employers, lest I be throttled by *them*) willdye at freedom.net _______________________________________________ Freenet-dev mailing list Freenet-dev at lists.sourceforge.net http://lists.sourceforge.net/mailman/listinfo/freenet-dev
