Re: [accounting regression since rc1] scheduler updates

[Martin Schwidefsky]

On Mon, 2007-08-20 at 17:45 +0200, Ingo Molnar wrote: 
> * Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:
> 
> > 1. Jan could finish his sched_clock implementation for s390 and we 
> > would get close to the precise numbers. This would also let CFS make 
> > better decisions. [...]
> 
> i think this is the best option and it should give us the same /proc 
> accuracy on s390 as before, plus improved scheduler precision. (and 
> improved tracing accuracy, etc. etc.) Note that for architectures that 
> already have sched_clock() at least as precise as the stime/utime stats 
> there's no problem - and that seems to include all architectures except 
> s390.

For far we have used the TOD clock for sched_clock. This clocks measures
real time with an accuracy of 1usec or better. The [us]time accounting
with CONFIG_VIRT_CPU_ACCOUNTING=y is done using the CPU timer. This
timer measures virtual time with an accuracy of 1usec of better. Without
CONFIG_VIRT_CPU_ACCOUNTING the [us]time accounting is done with HZ
ticks. Which means that sched_clock() is at least as precise as [us]time
on s390 as well, only that we distinguish between real time / virtual
time if the improved accounting is used.

> could you send that precise sched_clock() patch? It should be an order 
> of magnitude simpler than the high-precision stime/utime tracking you 
> already do, and it's needed for quality scheduling anyway.

Sure if you can explain what it should do. This is still unclear to me,
for a non-idle CPU the virtual cpu time should be used but for an idle
CPU the real time should be used ? That seems rather ill-defined to me.
On s390 we have three times to consider, real time, virtual cpu time and
steal time. For a given period we have real = virtual + steal. And if a
cpu is idle we have real = steal, virtual = 0. My best interpretation of
what you want is that sched_clock should progress with virtual cpu time
if the current process is not idle and with the real time if it is. No ?

> > [...] Downside: its not as precise as before as we do some math on the 
> > numbers and it will burn cycles to compute numbers we already have 
> > (utime=sum*utime/stime).
> 
> i can see no real downside to it: if all of stime, utime and 
> sum_exec_clock are precise, then the numbers we present via /proc are 
> precise too:
> 
>    sum_exec * utime / stime;
> 
> there should be no loss of precision on s390 because the 
> multiplication/division rounding is not accumulating - we keep the 
> precise sum_exec, utime and stime values untouched.

But then sched_clock() has to return the virtual cpu time only,
otherwise it will be hard to make sum_exec exact, wouldn't it?
And why should we jump through all these loops to come up with values
that are only as good as the values we already have?

> on x86 we dont really want to slow down every irq and syscall event with 
> precise stime/utime stats for 'top' to display. On s390 the 
> multiplication and division is indeed superfluous but it keeps the code 
> generic for arches where utime/stime is less precise and irq-sampled - 
> while the sum is always precise. It also animates architectures that 
> have an imprecise sched_clock() implementation to improve its accuracy. 
> Accessing the /proc files alone is many orders of magnitude more 
> expensive than this simple multiplication and division.

Yes, I can understand why you don't want to have the exact cpu
accounting scheme on x86 since it will slow down every context switch
quite a bit (that includes user <-> kernel, softirq <-> hardirq <->
process context, ..). On s390 the cost is acceptable, for an empty
system call it is about 40 additional cycles for the precise accounting.

-- 
blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/