DragonFly users List (threaded) for 2005-03
Re: Fwd: Re: Differences between constructive anddestructivecriticism
On Sat, 12 Mar 2005 18:54:12 -0500
> Processor speeds have "hit a wall", which means that they
> can't get much faster. So the best way to get more
> performance is to increase support for MP. My point was
> that current 'BSD OSes don't utilize MP very well, so
> the best price/performance is with UP.
Yes, at the present time, it probably is.
> I find the empirical
> evidence somewhat undeniable, but apparently the notion
> is irritating to people. Why do I care if other people
> waste their money? That, I can't answer.
> The "other" point, was that if I were to accept the premise
> that opterons are empirically faster than P4s, you can't
> possibly get better performance on a PCI bus than you do
> on a PCI-X bus, so quoting opteron MBs that don't have a
> PCI-X bus is not quoting a comparable system, assuming
> that networking is an important component of why you
> are buying a system in the first place. It seems that
> there are many that find this notion irritable as well,
> even though all of the empirical evidence is that running
> NICs on a faster bus substantially decreases the CPU load
> of the system.
This part (which I still have a bit of a hard time following) is
probably what confused me about when you said UP "hit the wall";
obviously if the bottleneck is not the processor, then that's not the
wall that's being hit!
> Now if someone actually had evidence to the
> contrary, it might be an interesting discussion. But what
> we have is a bunch of evidence in the form of easily
> duplicatable measurements,
I must have missed them... which message are they in?
> and a mob with no evidence at
> all, and apparently no experience even using the bus
> since they think its bleeding edge (even though its
> been the dominant but for many years, telling me I'm crazy.
> So the only possible conclusion is that there isn't any intelligent
> life amongst the mob members. Since both Matt and Joerge seem to
> consider a fast bus a "strange requirement", one has to wonder how
> dragonfly will ever be optimized for networking?
Well, hopefully someone with some sense will at some point send in some
patches that will allow DragonFly to support PCI-X :)
> Hope that clarifies things.
Yes, quite a bit. Thanks.