David Boyes wrote:
>> I just want to insert a comment about a Clustered F/S (better than a C/F,
>> right?) in an application environment where there are a sh!tload of links and
>> unlinks within a single directory...
>> Don't do it.
>> Just...  don't.
>
> Newer cluster file systems like PVFS2 and Ceph have mechanisms to limit the 
> impact of this kind of activity.

The problem arises when your application-- being multi-threaded and
moving up to 2.2 million files in a day-- *depends* upon linking and
unlinking being *fast* in order to "move" the files being manipulated.
 All running on HF (Huge) Solaris boxes and the F/S's in question
residing in a SAN.  While not normally a problem, when there is a
single directory that gets all of the files being processed for
auditing you can end up with one *hell* of a nightmare that cannot be
solved without a major re-design of the application itself.

Mind you, the locking mechanism around directory manipulation is
hidden and won't show up in the stats... and it was, in hindsight, a
miracle that I even remembered, during a meeting on this, that
directory manipulation is single threaded.  (We discovered, in some
follow-ups, that we were near the limit even on a single box with the
non-clustered architecture;  We have ways to avoid this, now.)

I hate IE right now since WXP keeps dropping characters as I type them
into this browser window.

-soup

-- 
John R. Campbell         Speaker to Machines          souperb at gmail dot com
MacOS X proved it was easier to make Unix user-friendly than to fix Windows

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390
----------------------------------------------------------------------
For more information on Linux on System z, visit
http://wiki.linuxvm.org/

Reply via email to