low(er) disk performance with sched_4bsd then with sched_ule
Joseph Koshy
joseph.koshy at gmail.com
Fri Sep 16 09:21:23 PDT 2005
jk> What happens if you use a kernel with 'options
jk> NO_ADAPTIVE_MUTEXES' ?
ol> Hmm, that made it more worse
ol> http://pofo.de/tmp/gprof.4bsd.2
>From the profile it appears that with adaptive mutexes turned off the
kernel is spinning inside vm_pageout().
41.1 vm_pageout [1]
25.0 _mtx_trylock [2]
5.1 smp_tlb_shootdown [3]
Hmm ... the dd command line *did* ask for about approx 2 GB of
zero-filled memory from the kernel while the machine has
about 640M.
I'm just not able to reproduce this kind of skewed
profile on -current on a uniprocessor amd64 and
on a P4 HT machine.
Just as an off-chance: how upto-date is your kernel? In
particular, do you have the following fix:
"sys/vm/vm_pager.c:"
revision 1.105.2.1
date: 2005/08/15 14:04:47; author: kan; state: Exp; lines: +1 -0
MFC: Do not use vm_pager_init() to initialize
vnode_pbuf_freecnt variable. vm_pager_init() is run before
required nswbuf variable has been set to correct value. This
caused system to run with single pbuf available for
vnode_pager. Handle both cluster_pbuf_freecnt and
node_pbuf_freecnt variableis in the same way.
--
FreeBSD Volunteer, http://people.freebsd.org/~jkoshy
More information about the freebsd-current
mailing list