On Fri, Jan 09, 2009 at 02:47:42PM -0800, Serge E. Hallyn wrote: > Quoting Guenter Roeck (groeck@xxxxxxxxxxx): > > On Fri, Jan 09, 2009 at 10:12:24AM -0800, Dan Smith wrote: > > > GR> I have tried something similar, only with > > > GR> CLONE_FILES|CLONE_FS|CLONE_VM|CLONE_NEWNET, and actually creating > > > GR> a virtual interface and controlling socket or thread in each new > > > GR> network namespace. > > > > > > My initial test was to create a veth pair and move one end into the > > > namespace during create. That failed in the same way, so I took the > > > veth's out of the equation with the posted test. > > > > > > GR> This scales to a couple of thousand interfaces, though interface > > > GR> creation takes a long time if more than 1,000 interfaces or so are > > > GR> created. > > > > > This is at least to some degree due to the problems I mentioned earlier. > > Enhancing the kernel name hash and the sysfs implementation improves > > performance a lot. > > Is this something you've had a chance to start addressing? (Just wondering) > Yes - I have code for the name hash change (just one line, really), and two variants of code for sysfs - one that uses a hash result as 1st step of comparison before doing a strcmp, and another which uses a hash table per directory in sysfs. The latter is of course more efficient, but also more expensive in terms of memory usage. If there is interest, I can submit a patchset once I find out how exactly to do it ;-). Guenter _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers