[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Miscellaneous config questions


On Tue, 19 May 1998, Jonathan Rozes wrote:

> [...]

> 4a. Anybody running a news server? Anything I should be aware of? I'm
>     planning to use 2 scsi controllers with 2 disks/controller supporting
>     inn-1.7.2 with a full feed minus the binary groups and a rapidly
>     growing number of concurrent readers.

I don't know. I'm running some 1.5* locally, but with not so many groups
(367, some of them not active really), and with only me as user.
No problems at all.

> 4b: How does LFS stack up against SGI's XFS for small-block random i/o?
>     How about its performance with large-block sequential i/o? Is there
>     a direct i/o interface that can bypass the buffer cache? Can atime
>     tracking be disabled at mount-time?

LFS doesn't work at all, if nothing has substantially changed with it.
(Btw, does null/union work? There were hangs after the VFS-Lite2
integration. Didn't track it, because I don't really understand the
semantics of the Lite2 vlode locking things. Unfortunately they
aren't covered in the devil book.)

Direct I/O: You can use raw disk/partition devices (/dev/rsd0e, for
example, for raw access to the 'e' partition of the first SCSI disk).
If you want to do raw I/O to files, this *could* work:
vnconfig vnd0 /some/file
(vnd w/o s before uses VOP_BMAP to get the block numbers for the
file and does I/O to those blocks directly. It's NOT coherent
with the filesystem cache, so you must avoid accesses to /some/file
via the filesystem. If you need coherence, use svnd and lose some speed
because then, the file is accessed via the normal filesystem operations
of the filesystem /some/file resides on.)
disklabel vnd0 and then use /dev/rvnd0a.

Note that raw I/O has some constraints: The buffer for read(2)/write(2)
must meet some alignment restrictions (page alignment should work)
and the transfer sizes (i.e. the nbytes argument) must be a multiple
of the block size (usually 512 bytes).

Atime can be disabled with the noatime mount option (see mount(8)).

> [...]

>    me to? Is OpenBSD pretty good about keeping stuff that needs to be
>    writable in one place?

It does a good job, with exceptions.

/var is written, of course. If you use named, in the default configuration
/etc/named is written (for secondaries and temporary files). (You can
change named.boot to direct named to use /var/named, instead.)

/etc/master.passwd, passwd, {s,}pwd.db must be writeable if users
are to be able to change passwords, of course.

If you're using ssh, there's also /etc/ssh_random_seed.

And of course /tmp.

In /usr, xdm writes /usr/X11R6/lib/X11/xdm/xdm-pid (perhaps that has
changed, since I'm still running some oldish X - xdm-pid should belong to
/var/run). Emacs has /usr/local/com/emacs/lock (don't know what the current
emacs-20.2 port does). That's it, on my system, based on a
find, looking for ( -mtime -2 -o -ctime -2 ).

And of course, (/usr)/home is written by users.

Regards, Felix.