My question was:
Is it possible to tune MBUF and NBUF params in the
Digital Unix 4.x kernel? Or is it unnecessary to do
so anymore?
Knut Helleb <Knut.Hellebo_at_nho.hydro.com> responded:
These values are not necessary to fix. They are
dynamically adjusted depending on some other set of
network kernel parameters. See 'man dxkerneltuner'
and 'sysconfig'
Summary:
This is what I suspected--that you don't set those
manually anymore. But, since it appears they are
"dynamically adjusted depending on some other set of
network kernel parameters" than I still need to find
out what those other network kernel parameters are.
Man of dxkerneltuner and sysconfig doesn't tell you
what params are tunable only nor does it describe them.
They only tell you "how" to tune them once you know
what they are.
However, there is a little book called "System Tuning
and Performance Management" that provides a reference
to the tunable parameters and a brief description of them.
You can even get to it on the web. Here's the long URL:
http://www.partner.digital.com/www-swdev/pages/Home/TECH/documents/Digital_U
NIX/V4.0/Digital_UNIX_Bookshelf.html
Click on "System and Networking..." to find it.
It appears that the buffer functionality in Digital Unix
is handled either by UBC for ADVFS filesystems, which
dynamically uses and reclaims pages within VM, so other
than tuning the ubc- params there would be nothing to
do there.
Or, for traditional Unix filesystems like UFS it still uses
the BSD buffer cache. To tune that you can modify
bufcache, buffer-hash-size, name-cache-hash-size,
and name-cache-size. To determine this you use dbx
to examine the bio_stats structure. This is all explained
in the aforementioned manual, section 3.4.1.3. Give
me back the days sith sar when you could just look
at %rcache and then the ratios of bread/lread and
bwrite/lwrite to see if there is a problem.
Anyway, as you can tell I'm still reading the book,
haven't tried monitoring for this and tuning this as
suggested yet. I am also hoping that if I go with
ADVFS across the board, then I don't have to worry
about it (let UBC deal with it dynamically).
Basically, I am trying to optimize a new system
that is migrating from an older platform/older
version of unix that has an observable buffer cache
hit ratio problem (which in the old days used to
mean give more memory to MBUF/NBUF or NREMOTE);
I'm trying to make sure that I optimize for this
"hot spot" while migrating this appl. off the old
machine onto the new one.
If there is any additional conventional (or modern)
wisdom on this topic I would appreciate it. If I
get any more info. from the list I will do another
summary.
Regards/Bob
Received on Thu Aug 14 1997 - 23:50:52 NZST