Re: a joint letter on low latency and Linux

[Larry McVoy]

> A generally preemptible kernel with tons of locks is IMHO clearly one of 
> these ``it is only 5-10% slower'' things that you criticized just a few
> messages ago.  

It's a tradeoff.  If you have a threaded kernel, you get SMP, which is 
generally considered to be a good thing (though I am strongly in the camp
of too much SMP is a bad thing for the health of your kernel).  

So the choices are then 

    a) non-threaded UP and threaded SMP vs
    b) threaded UP and threaded SMP.

Sun took choice (b) and SGI took choice (a).  I've worked in both places
on both kernels and I believe that the Sun approach is better.

Everything in life is some sort of tradeoff, there is no free lunch.
Part of being a designer is trying to make good tradeoffs rather than
bad ones.  In my opinion, if you are going to do SMP at all, then
threading both the UP and the SMP is a good tradeoff.  It acts as a
natural backpressure on the SMP stuff - there is an inherent cost in
threading things and if you pay that cost everywhere that will help keep
the cost from getting too high.