Mirror bandwidth and user redirection

Patrick O'Callaghan pocallaghan at gmail.com
Mon Jun 30 16:53:08 UTC 2008


On Mon, 2008-06-30 at 10:47 -0500, Matt Domsch wrote:
> On Mon, Jun 30, 2008 at 11:12:45PM +0800, Ed Greshko wrote:
> > Patrick O'Callaghan wrote:
> > >On Sun, 2008-06-29 at 23:46 -0500, Matt Domsch wrote:
> > >>The basic selection algorithm for choosing
> > >>the order in which to return mirrors to clients remains the same:
> > >>prefer same netblocks, internet2 in same country if on internet2, same
> > >>country, same continent, then global, in that order.
> > >
> > >That's totally logical, but it's wrong for some cases. Here in Venezuela
> > >there is much better bandwidth to the US than to anywhere else in South
> > >America, so the "same continent" rule is not going to work for us. I
> > >suspect the same is true for some other SA countries.
> 
> Understood.  But I don't have a way to know that.

Of course. What's needed is a way to tune these things manually.

> > The same is also true for Asia.  I would hope that the "same continent" 
> > rule has a tad bit more smarts in it.
> 
> Note, "same continent" is the 4th major sorting rule; if there are
> mirrors in the same country, they will be preferred.  We've got about
> 150 public mirrors right now, and are always looking for more.

I don't know of any in Venezuela at the moment. Furthermore, it's not
even clear that a local mirror would be faster for everyone in the
country.

> I don't have a way to know the whole global routing table to know
> which mirrors might be closer to individuals than others.  So I'm
> making what I think are pretty good guesses, and I don't get _too
> many_ complaints about Fedora mirror performance.  Except this one -
> that slow mirrors were being overloaded, which last night's change
> addresses.
> 
> I'm open to better solutions, preferably in patch form. :-)
> http://git.fedorahosted.org/git/mirrormanager has the open source code
> for the whole system.

My only suggestion for now is that the weighting of the various classes
be changeable via a config file. I don't know if that is easy or hard to
do given the existing code.

> > >Also, for the relatively few people on Internet2 it's always better than
> > >Internet1, at least here. I mean Internet2 to anywhere is better than
> > >Internet1 to the same city.
> 
> That all depends on the interconnects between the nodes on Internet2
> and the commerical internet.  As those links cost real money for our
> volunteer mirror admins, by request of some of the I2 mirrors in our
> system, I've tried to avoid sending non-Internet2 users to Internet2
> servers.

That's fine. I'm talking about I2<->I2 connections, which if available
should outweigh non I2<->I2 connections even if the latter are more
local.

poc




More information about the users mailing list