On Sat, 22 Jan 2011 07:59:32 +1300, Phillip Hutchings wrote:
> On 22/01/2011, at 7:55 AM, Dennis Nezic wrote:
> > On Fri, 21 Jan 2011 11:51:35 -0700, Ray Jones wrote:
> >> 
> >> If I understand correctly....
> >> 
> >> UDP has no such thing as flow control. So even though your machine
> >> reads only X packets per second, the sending machine is still
> >> sending and you're still receiving. If the packets build up too
> >> far your machine will drop them, but you've already used the
> >> bandwidth to get them there before they were dropped!
> > 
> > I agree it's a terrible waste -- but, I say, tough luck. Surely the
> > senders will throttle back when they start seeing some of their
> > packets not being acknowledged. (Like I said, we should avoid this
> > situation as much as possible, but ultimately the user has to be in
> > control. The network, selfish as this might sound, comes second!)
> NO! We're using UDP - UDP packets have no acknowledgment. TCP would
> behave like you describe, but UDP senders have no way of knowing that
> there's a problem.
> Freenet itself acknowledges packets,

That's what I was referring to.

> and that's part of the rate limiting, which obviously has some bugs.
> Not reading the packets will work for a while, but when packets are
> read and ack'd the senders will burst again, not knowing there's a
> limit.

Sure they will know there is a limit -- they will know how long it took
for the packets to get acked, and how many were dropped. It's not
perfect, but I'm sure they will know enough to pull back when necessary.

> The only solution is to fix the limiter bugs, which is tough
> for a project with so few developers.

Fixing that is definitely essential, although either way I think the
user's limit has to be paramount. (Perhaps there are other ways to
avoid dropping, like having the UDP packets queued into Freenet's
memory, but not used more than XKiB/second -- so the long ack-ing works
for a bit longer, rather than dropping the packets, which is worse.)
Support mailing list
Unsubscribe at http://emu.freenetproject.org/cgi-bin/mailman/listinfo/support
Or mailto:support-requ...@freenetproject.org?subject=unsubscribe

Reply via email to