-- Erik B. Andersen Web: http://www.et.byu.edu/~andersee/ 2485 South State St. email: andersee@et.byu.edu Springville, Ut 84663 phone: (801) 489-1231 --This message was written using 73% post-consumer electrons--
Linus Torvalds wrote: > > On Tue, 24 Sep 1996, Hans-Georg von Zezschwitz wrote: > > > > SMC Ultra is fine running 2.1.0 > > Thanks for the report.. > > > Another report about speed: > > > > Recompiling a clean pre2.1.0-tree with my configuration took: > > > > using Linux 2.0.21: > > 554.28user 36.98system 10:18.67elapsed 95%CPU (0avgtext+0avgdata > > 0maxresident)k > > 0inputs+0outputs (0major+0minor)pagefaults 0swaps > > > > using Linux 2.1.0: > > 553.11user 36.07system 10:15.06elapsed 95%CPU (0avgtext+0avgdata > > 0maxresident)k > > 0inputs+0outputs (0major+0minor)pagefaults 0swaps > > > > Well, that's 0,6%. Of cause, I can't claim that my single pass test > > has any relevance. Who Bever thinks, that this is very little, should > > consider that the German magazine c't was already delighted when > > Maple ran 10-15% faster using Linux instead of Windows. > > Well, to be quite frank, I didn't expect all that much of a speed increase. > After all, the segment override stuff is generally just a few clock cycles, > and you end up spending more time just waiting for memory anyway (unless you > have a slow machine, and then you have a slow machine ;) > > Also, I haven't actually gotten rid of the %fs stuff - I'll be able to shave > off a few ten clock cycles off the kernel entry routines by not having to > load %fs (and making "get_fs()"/"set_fs()" work like they do on the alpha). > > But more important than the potential speed advantage (which is likely to > become more marked on newer CPU's) is the fact that I got rid of maintaining > a lot of inline assembly (look at <asm-i386/segment.h> before and after the > operation), and that the new setup works the same way on x86 (the bread and > butter of the Linux world) as on most other ports. > > I'll try to produce better numbers (your differences are definitely within > the error limits), > > Linus