On Fri, Dec 15, 2006 at 01:17:45PM +0100, Jano wrote: > Matthew Toseland wrote: > > > Which is precisely why LIFO sucks... FIFO has a more or less fixed > > latency for a given load, it forces requests to wait in line. Whereas > > LIFO encourages anti-social behaviour such as cancelling and trying > > again every 10 seconds. So it's not practical - but I'm very surprised > > that we are getting apparently good results for it. > > But the fixed latency of FIFO can be *ages* (look at ed2k). This reminds me > of a discussion in the gnutella ML some years ago, where some lead > developer (of bearshare IIRC) and other people advocated that queues should > be completely dropped (admittedly they had not much support, but one of > them was the big theorist in there, kinda like oskar here?).
So you limit the length of the queues, and push everything back to the original requestor. Which is what we've always planned to do. Unlimited, but prioritised queues on the requesters, short, limited queues on the nodes just for matching requests to nodes more effectively and smoothing out load, and some sort of load propagation scheme be that client-side AIMD or token passing. > > *digging* > http://groups.yahoo.com/group/the_gdf/message/16137 > > As I see it, if LIFO indeed had better throughput (not proved in any case), > we could still use it for a low priority global queue (i.e. > inserts/downloads of big files), where the user has no practical say about > cancel/request. > > I'm also not sure of the degree of unfairness if you maintain queues per > peer instead of a single global queue, and drop duplicated queries. > > I reckon that LIFO can be seen as my current pet theory, and that's the > principal reason I want to explore it: to be convinced that we're not > overlooking a simple mechanism that can perform well. What attracts me of > it is that, in principle, is very simple: LIFO+dropping requires no tuning. > As always that some idea gets urgent in my mind, I'm already devoting to it > too much time. I'm the first one who wants to see LIFO as useless and move > on. It's not a practical solution because it encourages bad behaviour. However it would be useful to understand why it seems to work better than other things; my view is probably because of a problem in the simulation. -------------- 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/20061215/f702519f/attachment.pgp>
