New(er) kernels load averages

[lude]

I have three identical machines in a web cluster. The kernel we run in production is ancient, and I'm making some attempts to bring some of the newer features in the new kernels into production, I updated from our stable 2.6.24.7 to 2.6.32.

The machines all have the same workload, fed by a RR LVS system, each running apache 2.2.14, prefork MPM, and PHP, with the content mounted over NFS.

The 2.6.24 machines:
top - 10:30:26 up 1 day, 23:03, 3 users, load average: 0.28, 0.32, 0.25
top - 10:30:38 up 1 day, 21:15, 1 user, load average: 0.32, 0.28, 0.27

The 2.6.32 machine:
top - 10:30:43 up 19 min, 2 users, load average: 2.50, 3.28, 2.68

Now they all seem to have just as much idle CPU, and the content load seems the same, if not slightly better on the newer kernel. Can anyone explain (or at least attempt to) the difference in load average?

[sera]

Maybe it's because you switched to a tickless kernel which influences the calculation of the load average? If for a reason the the quelength is 3 every 5 seconds and 0 inbetween such numbers may occur.

If the system is responsive, I'd go for something as the above.

[durian]

The machine with the higher load has only been up for 19 minutes. At least on my machine a lot goes on after booting, and it seems to slow down a bit after a while. You should compare again after the same uptime maybe (but maybe you have done that already, not clear from the post),

-peter