glance and sar showing very different cpu idle times

[Doug Kratky]

We have a fairly busy Oracle server that we're trying to understand the performance of. One of the problems in analyzing is that sar and glance show very different statistics for cpu usage. sar shows the cpu as 70% "wio" and around 10% idle . glance shows the cpu as around 75% idle.

Can anyone explain why these are so different?


Thanks,
Doug

[harry d brown jr]

glance has hooks built into the kernel. sar does not.

live free or die
harry d brown jr

[Doug Kratky]

Okay, but is the implication that sar is just wrong? Why is sar fooled? One reason it's an issue is that the DBA tools are based on sar.

[Trever Furnish]

Wouldn't you consider "wio" time to be conceptually similar to idle time? If the cpu is Waiting on IO, then it's available for other work servicing processes that aren't waiting for IO. I'm certainly no expert, but I wouldn't really think of wio as a state for a CPU, but rather as a state for a process.

Maybe the reason your two sets of numbers are so close (80 and 75) is because they're actually showing the same thing, just at slightly different sampling times.

[Doug Kratky]

I wouldn't consider idle and wio to be the same. In one case, the CPU has nothing to schedule or to do. In the other case, the CPU is waiting for I/O to be satisfied before it can continue servicing a process.

[Trever Furnish]

But cpus don't wait for I/O, do they? Processes wait for I/O. CPUs go on about their business until I/O is ready to funnel to a process. That's why I said I thought 'wio' was a state for a process, not a cpu.

This is really an off-the-top-of-the-head response from me though, like I said I'm certainly no expert - but maybe the activity on the thread will draw some renewed attention from the real ITRC gods. :-)