Meaning of 'load ' on Linux 2005-06-16 - By Ken Snider
Back
Mimmus wrote:
> Hi,
> recently I had a discussion with an Oracle guy apparently involved in kernel
> tuning.
> I had some issues of "high" load on a RH3 server due to the infamous problem
> with U3 (if I remember...).
> He said me that "load" on Linux (i.e. output of 'uptime' or 'top') has a
> different meaning than in other Unixes. Under HP-UX, I was comfortable with
> values under 2-2.5; he said that under Linux 10-12 is normal and that he saw
> working servers with load even of 1000 (!) using Oracle 'custom' kernels.
Load average is a useless indicator, alone, of an issue.
Let me explain why.
On our production systems, we have perhaps three custom daemons purring
away. If one of them shot the cpu usage up to 100% and hammered the IO,
while the other two were mostly idle, the load would be somewhere slightly
above 1, assuming that the other daemons running weren't blocking on IO or
resources.
Alternately, if on another server, these three daemons were all slightly
busy, lets say, using 10% CPU each, and little IO, but were constantly
"Active", the load would be around 3.
So, if I assumed that a load of 3 was the more heavily loaded server, I'd be
wrong. I'd have known this if, instead of monitoring "load average", I was
monitoring either average CPU usage, or average IO read/writes per second,
which is exactly what we do.
Another example: I happen to host http://www.boingboing.net, which is hosted
on a dual-CPU server. Because of the use of PHP, I run the apache daemon in
"prefork" mode, and have about 800 children listening for connections.
In this environment, the load is almost *always* at about 50 (more for a
slashdotting :) ), because on a server that literally gets millions of hits
a day, there's a good chunk of requests/sec coming in. But both the CPU and
IO load of the server are very low, the server moving only a MB/sec of data
via IO, and the CPU's at about 25% utilization. And this is moving about
12Mbit/sec of HTTP data around, no less.
This box is positively *bored* compared to our application servers which
have only a few daemons on them, and move 10-20MB/sec of IO, have two
processors running at about 75%, and are running with loads of about 1-2 on
average.
About the only benefit I've ever seen for load average? It's a good
*baseline*. Since I know the boingboing server runs with an average load of
about 50, if I suddenly see it with a load of 100, well, I know something's
up. by the same token, our servers with an average load of 1 worry me when
they approach 2.5+.
--
Ken Snider
--
Taroon-list mailing list
Taroon-list@(protected)
http://www.redhat.com/mailman/listinfo/taroon-list
Earn $52 per hosting referral at Lunarpages.
|
|
