Accuracy/reliability of "top" CPU usage numbers

[Grant Edwards]

On 2014-11-14, Cathey, Jim <jcathey at ciena.com> wrote:

> If you can't modify the kernel to collect more accurate data (like we
> did with our top-oid in OSE),

That would be an interesting approach, but the downside to forking the
Linux kernel can be pretty significant.

> you have to get a lot smarter. Set up some 'offline' processing of
> bulk data, NON-clock synchronized, and characterize that. (i.e. I can
> process 1000 records/second in a non-clock-synced loop, therefore
> each record takes 1msec.  Normal clock-synced behavior is 1 record
> per clock invocation.)

That's an interesting approach, but it's probably not going to happen
either.  The processing is mostly handling various communications
protocols with way too many nested, overly complicated state machines
(inside protocol stacks we acquired from elsewhere). Setting up
offline processing to measure CPU usage for them is going to be
difficult.  The access times for the hardware involved might be a
significant portion of the "CPU time" used by some of these threads as
well: one of the peripherals has very slow bus timings compared to
everything else.

-- 
Grant Edwards               grant.b.edwards        Yow! If I am elected no one
                                  at               will ever have to do their
                              gmail.com            laundry again!