On Sun, 2010-03-14 at 20:57 -0400, Bill Davidsen wrote: > Patrick O'Callaghan wrote: > > On Sat, 2010-03-13 at 16:52 -0500, Bill Davidsen wrote: > >> Patrick O'Callaghan wrote: > >>> The yum fastestmirror plugin (yum-plugin-fastestmirror) claims to > >>> evaluate the speed of a bunch of repo mirrors and use the fastest one > >>> relative to the user's location. > >>> > >>> However AFAIK what it *actually* does is make a test connection to the > >>> to the candidate mirrors and order them according to response time, > >>> which in many cases is dominated by network latency, which can distort > >>> the results. For well-connected user machines in first-world countries > >>> it probably doesn't matter much, and may have the beneficial effect of > >>> spreading the load over a wider range of mirrors, but for those of us in > >>> a less privileged position it can matter a lot. Ironically, these are > >>> the cases where such an optimization could do the most good. > >>> > >> And there you have the heart of the problem, the evaluation is not remotely > >> correct for most cases. It would be worth adding code to download some small RPM > >> from a number of sites and measure b/w for something real. However, disabling > >> the feature works, too. > > > > Sadly, downloading a "small" RPM is unlikely to give very reliable > > results either. Due to TCP slow-start, a stable effective b/w may only > > be reached after some 10's of kb have been downloaded. > > > Wow, I wasn't talking dialup. Tens of kb is a few ms on anything useful for > major D/L, the slowest DSL I've ever seen sold was 768kb, or max of ~80kB/s. > Figure 100kB per site and stop after the ten fastest are within 5-10% of each > other (that's probably max line speed). Base the speed calc on the last 10k of > 100kB to get the steady state value. You don't really care if the process is > slow, once the table is built you use it. Actually, rechecking the five fastest > is probably practical. That's kind of what I was saying actually. Maybe we have different ideas of what "small" means. My DSL line is 1Mbps and of course it's not just used for updating Fedora :-) Anyway, the general idea jibes with what I'm thinking. I'll be sure to let people know if I actually do it, which is vanishingly unlikely but not strictly impossible. > > This is not an easy problem to solve. > > > Fortunately, you don't have to find the best, virtually all of the top sites > will be close unless they're overloaded. The perfect is the enemy of the good > enough. That's certainly true (except when dealing with security, but that's a horse of a different colour). poc -- users mailing list users@xxxxxxxxxxxxxxxxxxxxxxx To unsubscribe or change subscription options: https://admin.fedoraproject.org/mailman/listinfo/users Guidelines: http://fedoraproject.org/wiki/Mailing_list_guidelines
