On Sunday 08 February 2004 04:49 am, Poul-Henning Kamp wrote: > In message <20040208094537.GA14749_at_VARK.homeunix.com>, David Schultz writes: > >> 10nsec per operation is getting you into the territory of effective > >> TSC-timecounter resolution, RAM access time, cache miss delays > >> and all sorts of other hardware effects. > > > >To avoid jitter and timestamping overhead, I read the time only at > >the start and end of the entire sequence of 10000 operations. > >I obtained the sample variance by running the entire test three > >times, i.e. > > Yes, but you have to remember that quite a lot of stuff happens > in the kernel of your iteration, and the stratification seems > to happen there. > > >Nevertheless, you're definitely right about the stratification. > > > >Yes, I realize that. I took 10 more samples of 10000 forks each > >with 5000 sleeping processes in the background and got the > >following: > > > >This data show a difference at the 95% confidence level, namely, > >that the NetBSD algorithm is about 1% faster on a system with 5000 > >processes (and only 0.1% faster if you're looking at the total > >overhead of fork() rather than vfork().) I think that pretty much > >rules out performance as the deciding factor between the two. > > Uhm, if you are using "Student's T" you have to remember that it > is only valid for gaussian noise processes. The stratification > we see is not any where near to gaussian. > > Either way: "tjr" should be our choice. Agreed. I'll defer to das_at_ on this one as he is the one who has done the actual research. Tom's version certainly seems to be the consensus. -- John Baldwin <jhb_at_FreeBSD.org> <>< http://www.FreeBSD.org/~jhb/ "Power Users Use the Power to Serve" = http://www.FreeBSD.orgReceived on Mon Feb 09 2004 - 11:46:36 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:37:42 UTC