On Thu, Apr 13, 2006 at 05:42:23PM +0100, Michael Rogers wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> OK, I think this is the fix: keep a separate rate for each pair of
> neighbours, and adjust it in response to both local and non-local
> RejectedOverloads.
> 
>     E   F
>     |   |
> A---B---C---D
> 
> D sends a RejectedOverload in response to traffic from A. C reduces its
> rate from B to D, and B reduces its rate from A to C. Even if A carries
> on sending at the same rate, it is throttled at B. The path EBCF is
> unaffected.

Not necessary. Load will be propagated back to A because A is spamming
requests and therefore with a good queueing algorithm ultimately gets
most of the failures.
> 
>     F       G
>     |       |
> A---B---C---D---E
> 
> E sends a RejectedOverload in response to traffic from A. D reduces its
> rate from C to E, C reduces its rate from B to D, and B reduces its rate
> from A to C. Initially the path FBCDG is affected, but it starts to
> recover as successful requests along FBCDE offset unsuccessful requests
> along ABCDE. C's rate from B to D represents a combination of the two
> paths. *However* in contrast to the previous approach, B's rate from A
> to C remains low even after C's rate from B to D recovers, because B
> reduces its rate whenever congestion occurs anywhere on the path ABCDE,
> not only when C sends a RejectedOverload after reducing its own rate.
> Thus A's traffic remains throttled at B, hopefully allowing FBCDE to
> recover...
> 
> Does this sound plausible?

Still unnecessary complexity.
> 
> Michael
-- 
Matthew J Toseland - toad at amphibian.dyndns.org
Freenet Project Official Codemonkey - http://freenetproject.org/
ICTHUS - Nothing is impossible. Our Boss says so.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: 
<https://emu.freenetproject.org/pipermail/devl/attachments/20060421/9b369b0a/attachment.pgp>

Reply via email to