you may take a look at http://en.opensuse.org/Build_Service/Redirector
On Thu, Jul 17, 2008 at 8:22 PM, Shawn Walker <[EMAIL PROTECTED]> wrote: > Stephen Hahn wrote: >> * Shawn Walker <[EMAIL PROTECTED]> [2008-07-16 22:28]: >>> 3) Automatic "preferred" mirror selection >>> >>> There are a number of techniques that I've seen used over the years to >>> determine the best mirror to connect to. I've seen package managers use >>> average ping time, average transfer speed, round-robin, and even random >>> selection to determine which mirror to use from an available list. >>> >>> One thought that occurred to me today was to base selection on closest >>> timezone (a mirror in the same timezone as you is likely to provide a >>> better experience than one in another timezone). >>> >>> In fact, a combination of techniques could be used to isolate the best >>> one to mark as the preferred mirror. For example, we could get the >>> mirrors in the same timezone first (then move up to the next greater >>> criteria), then check response time, then check speed, etc. until we >>> narrow it down to a specific one. >> >> My guess is that your timezone proposal is to manage congestion. I'm > > No, actually. It was my cheap hack to get around the fact that you > can't reliably detect the client's geographical location relative to the > server. Strictly looking at the US for a moment, it is very likely that > a server in the US/Central timezone (as an example) would give me a > better network response time if I'm actually in the US/Central timezone > as well (physically) than a server in the US/Pacific timezone. > > The time of day is an interesting configuration aspect I hadn't > considered, though as you point out, actual measurements are likely to > be more reliable. > > I had strictly mentioned the timezone as a beginning part of the > heuristic. In other words, if you have a hundred servers to try out, it > makes sense to start with the ones relatively closest to the client. > > I realise timezone is strictly reliable either, but I think it's > probably a good guess most of the time. > >> not sure that this heuristic would be better than one based on actual >> measurements (since those would reflect the operation times, rather >> than the coincidence of an average day). In any case, you might enjoy >> >> Sivasubramanian et al., Replication for Web Hosting Systems, ACM >> Computing Surveys, September 2004. >> http://doi.acm.org/10.1145/1035570.1035573 >> >> which discusses choices for replication, and introduces a number of >> conventions for measuring aspects of a replicated system. We're >> trying to do something a bit outside the paper (since we want to have >> a collection of independent mirrors, rather than a singly >> administrated set of replicated sites), but it's a useful framework. > > I will definitely read that. Thanks. > > -- > Shawn Walker > _______________________________________________ > pkg-discuss mailing list > [email protected] > http://mail.opensolaris.org/mailman/listinfo/pkg-discuss > -- Andy http://blog.sartek.net _______________________________________________ pkg-discuss mailing list [email protected] http://mail.opensolaris.org/mailman/listinfo/pkg-discuss
