On Fri, Nov 17, 2006 at 10:12:35AM +0000, Michael Rogers wrote: > A quick progress update: > > Congestion control is done but not documented.
It should be documented. But there are other important things to be done, so this is not critically urgent. It does need to be dealt with before load limiting can be implemented though IMHO, so it's fairly urgent. > > Load limiting is nearly done: the simulator now includes a simplified > version of the current backoff mechanism and an incomplete version of > the new token-passing mechanism. Both can be disabled easily, so I > should be able to make some preliminary comparisons between the two > mechanisms (and the third possibility, no flow control) in the next few > days. There are two basic problems for load balancing: - Serious overload causing timeouts etc - Serious misrouting Can you evaluate both? > > To do (short term): > > * A queue for locally generated searches Good. > * Request throttling for the current backoff mechanism - locally > generated searches should be released from the queue at a rate that > depends on the number of RejectedOverloads they cause Yes. > * Fullest-bucket, emptiest-bucket and random-bucket policies for the new > token-passing mechanism I still don't understand our token allocation scheme, but hopefully this will become clear through simulations. > * Debug both mechanisms in small-scale simulations Right. > * Badly-behaved nodes that don't throttle their searches Normally I'd say postpone attack simulations but in this instance it makes sense, as one of the main features of the new algorithm is that it should eliminate flooding. > * Wiki page describing congestion control Would be appreciated! > > To do (long term): > > * Heterogeneous nodes On bandwidth? Connections? The latter would come with churn, no? > * Swapping So in the short term you are simulating a constructed, perfect kleinberg network with its original locations. Okay, that should be useful. > * Churn We've had a suspicion for a while that node churn results in strange clustering behaviour; provincial nodes (nodes with few connections) get provincial locations (locations far from those taken by the main nodes), and then when they disconnect, those locations are lost. And thus we get selective pressure to keep the main part of the network in a relatively narrow area of keyspace. It would be interesting to see if this happens in simulation; it's just a theory at the moment, location probing doesn't yet work well enough to be sure about it. > * Other social network models Good idea. > > Cheers, > Michael -------------- 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/tech/attachments/20061117/ffb6d279/attachment.pgp>
