Re: [RESEND PATCH V5 0/8] remove cpu_load idx

[Preeti Murthy <preeti.lkml@gmail.com>]

Hi Morten, Peter, Alex,

In a similar context, I noticed that /proc/loadavg makes use of
avenrun[] array which keeps track of the history of the global
load average. This however makes use of the sum of
nr_running + nr_uninterruptible per cpu. Why are we not
using the cpu_load[] array here which also keeps track
of the history of per-cpu load and then return a sum of it?
Of course with this patchset this might not be possible, but
I have elaborated my point  below.

Using nr_running to show the global load average would
be misleading when entire load balancing is being done on the
basis of the history of cfs_rq->runnable_load_avg/cpu_load[]
right? IOW, to the best of my understanding we do not use
nr_running anywhere to directly determine cpu load in the kernel.

My idea was that the global/per_cpu load that we reflect via
proc/sys interfaces must be consistent. I haven't really
looked at what /proc/schedstat, /proc/stat, top are all reading
from. But /proc/loadavg is reading out global nr_running +
waiting tasks when this will not give us the accurate picture
of the system load especially when there are many short running
tasks.

I observed this when looking at tuned. Tuned sets the cpu_dma_latency
depending on what it reads from /proc/loadavg. This would mean
for a small number of short running tasks also this metric could
reflect a number which makes it look like the system is loaded
reasonably. It then disables deep idle states by setting a high
pm_qos latency requirement for system. This is bad because
it disables power savings even on a lightly loaded system. This
is just an example of how users of /proc/loadavg could make
the wrong decisions based on an inaccurate measure of system
load.

Do you think we must take a look again at the avenrun[] array
and update it to reflect the right cpu load average?

Regards
Preeti U Murthy