On Wed, Dec 9, 2009 at 12:03 PM, <and...@msu.edu> wrote: > Quoting "Donald Allen" <donaldcal...@gmail.com>: > >> My understanding is that OpenBSD still employs the Giant Lock approach >> to SMP, serializing access to kernel services. Is this still true? If >> it is, do Theo and the other kernel developers consider it a priority >> to improve this? >> >> (I am NOT complaining. I completely understand that OpenBSD is a labor >> of love and that development resources are limited and that doing SMP >> right isn't easy. I'm simply trying to get an idea of whether this is >> likely to be addressed in the near future or not.) >> >> /Don Allen > > You have to keep in mind that thinking about race conditions and other > possible problems took (and always take) precednece over "speed". > Me, I find SMP systems to be fast--are other systems faster? Probably, > but so what? Optimizing for anything other than security, reliability and > being solid is foolish. I always tell people if OpenBSD isn't fast enough, > they can get faster hardware.
Not if they just made a major investment in today's fastest hardware. Or if the performance hit is an integral factor, as could be the case in certain system-call-intensive situations when comparing coarse- vs. fine-grained locking on a 4- or 8-processor system. > > Things get faster with time, like pf. A couple (4?) versions ago pf got > a fair amount faster. That was good, but happened only after people > saw pf working and then thought of ways to optimize it. Thats the way > it should be done -- make something *work* and then if optimizations > don't interfere with the primary goals, apply them. Running well on multi-core processors isn't an optimization, in my view. Small-to-moderate-scale SMP is where computer architecture is going, and being matched well to that architecture is more fundamental than the word "optimization" implies. Certainly I agree with you that a blazingly fast but unstable and/or insecure system isn't worth much in most, if any, settings. On the other hand, a rock-solid, secure system that simply doesn't deliver the computations at the needed rate isn't worth much either. Look, I like the OpenBSD way of doing things. It's the most stable, well-crafted, well-documented, easy-to-administer system that I know (and I've tried many of the prominent Linux distributions, plus FreeBSD). The attention to security, clean code, clear documentation, and QA set it apart from other choices. But, like it or not, people do care about performance, and frequently for more legitimate reasons than instant gratification. /Don > > --STeve Andre'
