El Miércoles, 13 de Septiembre de 2006 12:02, Benoit Myard escribió: > On Wed, Sep 13, 2006 at 11:21:26AM +0200, Jose M. Prieto wrote: > > El Martes, 12 de Septiembre de 2006 20:30, Attila escribió: > > > On Dienstag, 12. September 2006 11:49 Jose M. Prieto wrote: > > > > Once the problem arises, it cannot be "reapaired". The problem with > > > > window scaling is that routers that can't handle it drop part of the > > > > packets; the greater the scale factor beeing used, the more packets > > > > are dropped. The only way to recover dropped packets is retranmitting > > > > them, which wastes bandwith. > > On the other hand, sending lots of very small packets (which is what > happens with Window Scaling disabled; remember 64Kb) does not theorically > mean speed increase as it's been written previously.
I don't think disabling scaling means lots of small packets, as the maximum size of an IP datagram is 64KB in either case. Sending lots of small packets doesn't increase speed, preventing retransmits does. > > I have tested here with scaling enabled. > > Wait. Window Scaling has been created back in 1992 and it's been enabled in > the kernel for quite a while ! > > The problem does not *magically* arise. > > It only breaks in 2.6.17 because the Scale Factor calculation scheme has > been rewritten. There were some 2.6.7 testing kernel which used to break as > well because the default scale factor had been increased from 0 to 7! > > Anyway, the *real* source is broken routers/firewalls which rewrite some > packets: those containing the Scale Factor at the very begining of a > connection. Completely agree. There is nothing wrong with the kernel. However, instead of replacing my router I'm just disabling scaling, because I don't have enough bandwith to take advantage of the feature. _______________________________________________ arch mailing list [email protected] http://www.archlinux.org/mailman/listinfo/arch
