[vox-tech] Need Partitioning Advice

[Karsten M. Self]

Turns out I'm still subscribed - nomail - to vox-tech ;-)

on Thu, Jun 18, 2009 at 11:39:58AM -0700, Gabriel G. Rosa (grosa at ucdavis.edu) wrote:
> On Thu, Jun 18, 2009 at 11:27:06AM -0700, Rick Moen wrote:
 
> I find the argument of journal overhead to be about as relevant in a
> modern machine as the argument of software RAID overhead. That is to say,
> not at all.

Journals:

  - Take up space on disk.  Not much but below about 1GB it becomes
    large relative to the storage area.

  - More significantly:  are updated on every file write operation.
    Toss in some performance bugs and you're increasing the number of
    writes (and seeks) necessary for basic activity.  On high-activity
    filesystems (/tmp, /var/spool, /scratch) this can be more than a
    passing concern.  As with noatime and nosync, selecting a
    non-journaled filesystem may offer performance advantages.

Rule of thumb:  if you don't care whether or not your data's there at
next boot, don't use a journaled filesystem.  /tmp tends to meet that
criterion.

Doubly so as FHS specifically defines /tmp as not guaranteeing
persistance across boots.  Handily enforced by using tmpfs.
 
> > Multiple swap partitions per _spindle_, as mentioned in part of
> > Karsten's page, is indeed old hat.  On the other hand, having multiple
> > swap partitions of the one-per-spindle variety is just common sense, as
> > it improves performance considerably.
> 
> I think Bill's point is that swap spindle optimization is become largely
> irrelevant with cheap and abundant RAM. You can argue it's not a lot
> of extra work to set up, but it's also not a lot of gains to be had
> over time.

I'll point one last time at Martin Pool's essay.

Swap partitions on multiple spindles are AFAIK automatically striped by
the kernel for optimal performance.

If you've spent any time watching system performance, you'll find that
waiting for heads to seek and settle chews up an impressive amount of
time.  Even on "simple" desktop systems (perhaps moreso as task load is
highly variable and perceived subjective responsiveness important).

Swap doesn't make your box slow, swapping does.  Which indicates either
insufficient RAM, poorly written code, or both.
 
> > Anyhow, I'd feel a prize chump if I had my server set up as
> > single-filesystem plus swap on quite a few grounds, including
> > performance:  Being able to put the swap in the middle of the
> > spindle, and the most-visited portions of the file tree on either
> > side, is a huge win for keeping average seek time low.  I'd be
> > bloody incompetent if I _didn't_ do that.
> 
> Until your storage is all solid state and seek times become
> meaningless. Some of us (although not me yet) are already there.

I don't see that as being cost-effective in all cases for some time,
though it's likely true for portable devices within the next five years,
as well as most lower-end consumer desktops.  Server storage
requirements will still require spinning spindles and seeking heads for
a while to come if only for cost and space reasons.  My own speculation
is that SSD front-ending will be used to greatly accelerate caching on
most devices, which should improve performance markedly.  As Andrew
Morton says:  "solid-state disks are going to put a lot of code out of a
job."

    http://lwn.net/Articles/275087/


Peace.

-- 
Karsten M. Self <karsten at linuxmafia.com>        http://linuxmafia.com/~karsten
    Ceterum censeo, Caldera delenda est.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
Url : http://lists.lugod.org/pipermail/vox-tech/attachments/20090619/b926f617/attachment.pgp