On Monday 29 Aug 2011 02:48:59 Evan Daniel wrote: > On Fri, Aug 26, 2011 at 11:50 AM, Matthew Toseland > <toad at amphibian.dyndns.org> wrote: > > Freenet 0.7.5 build 1401 is now available. Please upgrade, it will be > > mandatory at midnight. This build turns off New Load Management, for the > > time being. If performance continues to be poor we will know the problem is > > elsewhere (it is possible that it is a problem with the asyncGet changes, > > although I don't see how). There are also fixes related to dropping peers > > due to the one IP per connection setting. You should not normally enable > > this setting on darknet (core settings); it can cause your friends to be > > lost. > > > > Thanks, and sorry for all the problems lately. > > Was it really necessary to have the update be mandatory on such short > notice? I'm not trying to be sarcastic, this is a serious question and > I'm curious about your opinion.
There were very serious problems. 1397 enabled NLM across the network, 1398 turned off AIMDs. While this increased success rates substantially, it also dramatically reduced throughput, especially for SSKs. It was necessary to turn AIMDs back on (which I now believe is appropriate anyway), and after that didn't appear to improve matters significantly, I turned NLM off as well. Unfortunately, inserts were so slow that 1401 didn't actually start to get deployed until *after it was already mandatory*. Which likely caused additional problems, scrambling the opennet topology for a while. > > It seems to me that you get most of the impact thanks to the > auto-updates, which most of the network uses. Some of us, however, do > not. I can't use the auto-update without it regularly interfering with > the network size graphs I produce. (Yes, the scripts that run it are > brittle and sensitive to things like that. Yes, I'd like to throw them > out and write something better, but motivation to actually do that has > yet to strike.) So I have to manually perform updates at times when > they aren't running, and I missed the window on that one. The result > is wonky network size info that I'm pretty sure is entirely an > artifact of that, and has no relationship to what the recent > performance issues have done to network size, which would have been an > interesting question. > > Anyway, a little more warning would have been nice, but obviously the > health of the network comes first. > > Is there a policy on what the requirements are for a build to be made > mandatory at all? And what the warning period should be? If not, it > seems like something we should have. Basically this was firefighting. I believed I had seriously screwed everything up and wasted 6 months' work at the same time; this also explains the discussions about load management. Generally, IMHO, builds should be mandatory in a week or so if they do anything that affects network behaviour. This can be accelerated if we need to try stuff out quickly for various reasons. Very short mandatories should only happen when something is seriously broken. -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 198 bytes Desc: This is a digitally signed message part. URL: <https://emu.freenetproject.org/pipermail/devl/attachments/20110831/72b40a42/attachment.pgp>
