On 14:04 20 Dec 2004, Kevin Wang <rightsock@xxxxxxxxx> wrote:
| first of all, the definition of 'load average' is 'number of processes
| waiting for something'. that can include processes that are hung on a
| particular I/O request.
Um, not really. It's the average length of the run queue - programs
_ready_to_run_, not waiting for something to happen.
For purposes of hinting at other load, processes in short term waits (such
as waiting for reads or write to/from _disc_ (== files)) are also counted,
since such requests should be satified "quickly". Conversely, programs
waiting for data from other sources (eg pipes or network connections)
don't get counted towards load - they may well wait forever and are
therefore considered idle.
As a footnote to the about, reads from NFS shares count as disc reads,
not network reads, and so contribute to load. So you'll find a down NFS
server will cause programs trying to access it to count towards the load,
although they consume no CPU.
| However, 15 of them with no actual cpu used and no other system
| problems (failing hardware is a common source of problems) does seem
| quite unusual.
| Probably best to ps auxwww, and check out column 8. it will tell you
| what state the process is in. as per the man page, you're probably
| interested in Z and D processes:
|
| PROCESS STATE CODES
| D uninterruptible sleep (usually IO)
This state "D" is the one that consumes no CPU but does count towards
load. The others either consume CPU ("R") or don't and also don't count
("S", "T", "Z").
Cheers,
--
Cameron Simpson <cs@xxxxxxxxxx> DoD#743
http://www.cskk.ezoshosting.com/cs/
Evolution has developed two basic strategies for staying alive and
well: armour, such as the triceratops or the armadillo; or intelligence
and agility, such as the primates. On the roads we see the two extremes
of these strategies in the Volvo driver and the cyclist.
- Chris Malcolm (Edinburgh AI)
