| |||||||||||||
Van Jacobson's network channelsVan Jacobson's network channelsPosted Jan 31, 2006 22:49 UTC (Tue) by csamuel (✭ supporter ✭, #2624)Parent article: Van Jacobson's network channels
A couple of other quick comments from my notes of the two instances of this talk done at LCA 2006:
0) VJ said that the talk was *not* about fixing the Linux TCP stack as "the Linux TCP stack isn't broken" - but just because something has always been done this way (SKB's) doesn't mean it is necessarily the best way.
1) VJ was asked about this code going into the kernel - his reply was that he would slap the GPL onto his new code for the drivers and the socket layer on the plane home but that the user layer TCP side may be a bit more difficult as he needs to get agreement from others.
2) The entire user level TCP stack was done as a LD_PRELOAD'ed library and hence no actual changes to applications are necessary, so people can experiment to their hearts content with tuning TCP application by application. Fancy an Apache with a different congestion control method to your OpenSSH clients and server ? Go for it..
3) This reduced the amount of code in the interrupt handler of the e1000 considerably (from ~700 lines down to ~300) and removes all SKB code hence simplifing drivers, which can only be a good thing.
4) The new channelised napi_poll() routine is generic, rather than the current device-dependant implementations.
All in all an excellent talk
(Log in to post comments)
Van Jacobson's network channels Posted Feb 1, 2006 0:57 UTC (Wed) by nix (subscriber, #2304) [Link] Obviously in a full-userspace-TCP/IP implementation, the default implementation would go into something on a par with glibc. Giving everything that wants to do networking an LD_PRELOAD would hammer performance and give Ulrich Drepper an aneurysm ;)
Van Jacobson's network channels Posted Feb 7, 2006 18:15 UTC (Tue) by arafel (subscriber, #18557) [Link] Well, y'know, I'm sure Ulrich's not got enough to do these days... ')
|
|||||||||||||