lkml.org 
[lkml]   [2014]   [Mar]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [BUG] Paravirtual time accounting / IRQ time accounting

Quoting Rik van Riel <riel@redhat.com>:

> On 03/20/2014 11:01 AM, Glauber Costa wrote:
>> On Wed, Mar 19, 2014 at 6:42 AM, <lwcheng@cs.hku.hk> wrote:
>
>>> ------------
>>> [src/kernel/sched/core.c]
>>> static void update_rq_clock_task(struct rq *rq, s64 delta)
>>> {
>>> ... ...
>>> #ifdef CONFIG_IRQ_TIME_ACCOUNTING
>>> irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
>>> ... ...
>>> rq->prev_irq_time += irq_delta;
>>> delta -= irq_delta;
>>> #endif
>>>
>>> #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
>>> if (static_key_false((&paravirt_steal_rq_enabled))) {
>>> steal = paravirt_steal_clock(cpu_of(rq));
>>> steal -= rq->prev_steal_time_rq;
>>> ... ...
>>> rq->prev_steal_time_rq += steal;
>>> delta -= steal;
>>> }
>>> #endif
>>>
>>> rq->clock_task += delta;
>>> ... ...
>>> }
>>> --
>>> "delta" -> the intended increment to rq->clock_task
>>> "irq_delta" -> the time spent on serving IRQ (hard + soft)
>>> "steal" -> the time stolen by the underlying hypervisor
>>> --
>>> "irq_delta" is calculated based on sched_clock_cpu(), which is vulnerable
>>> to VM scheduling delays.
>>
>> This looks like a real problem indeed. The main problem in searching
>> for a solution, is that of course not all of the irq time is steal
>> time and vice versa. In this case, we could subtract irq_time from
>> steal, and add only the steal part time that is in excess. I don't
>> think this is 100 % guaranteed, but maybe it is a good approximation.
>>
>> Rik, do you have an opinion on this ?
>
> The other way around may be better, since steal time (when it
> happens) is likely to be of "time slice" magnitude, while irq
> time will happen more frequently, and in dozens-of-microseconds
> intervals.
>
> Furthermore, we have no way to measure what the irq time is,
> except by looking at how much real time elapsed. For steal time,
> however, the hypervisor tells us exactly how much time was stolen.
>
> That means when steal time and irq time happen simultaneously,
> the amount of steal time should always be smaller than the
> calculated irq time for that period.
>
> actual irq_time = calculated irq time - reported steal time;
>
> --
> All rights reversed
>

I observe that sometimes irq_time only includes "part" of steal_time.

Like you said, irq_time is in dozens-of-microseconds. In VMs, as all
devices seen are virtual ones, irq_time seems to be not as desired
as it is in physical hosts.

A quick (but not radical) solution may be:
disable CONFIG_IRQ_TIME_ACCOUNTING if CONFIG_PARAVIRT is selected.
Just adopt tick-based accounting: CONFIG_TICK_CPU_ACCOUNTING

I am thinking what irq_time really *means* in VMs.

-Luwei




\
 
 \ /
  Last update: 2014-03-22 09:01    [W:2.160 / U:0.092 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site