[lkml]   [2005]   [Aug]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] i386 No-Idle-Hz aka Dynamic-Ticks 3
* Srivatsa Vaddagiri <> [050805 05:37]:
> On Wed, Aug 03, 2005 at 06:05:28AM +0000, Con Kolivas wrote:
> > This is the dynamic ticks patch for i386 as written by Tony Lindgen
> > <> and Tuukka Tikkanen <>.
> > Patch for 2.6.13-rc5
> >
> > There were a couple of things that I wanted to change so here is an updated
> > version. This code should have stabilised enough for general testing now.
> Con,
> I have been looking at some of the requirement of tickless idle CPUs in
> core kernel areas like scheduler and RCU. Basically, both power management and
> virtualization benefit if idle CPUs can cut off useless timer ticks. Especially
> from a virtualization standpoint, I think it makes sense that we enable this
> feature on a per-CPU basis i.e let individual CPUs cut off their ticks as and
> when they become idle. The benefit of this is more visible in platforms that
> host lot of (SMP) VMs on the same machine. Most of the time, these VMs may be
> partially idle (some CPUs in it are idle, some not) and it is good that we
> quiesce the timer ticks on the partial set of idle CPUs. Both S390 and Xen ports
> of Linux kernel have this ability today (S390 has it in mainline already and
> Xen has it out of tree).

Good point, and it would be nice to have it resolved for systems that support
idling individual CPUs. The current setup was done because when I was tinkering
with the amd76x_pm patch a while a back, I noticed that idling the cpu
disconnects all cpus from the bus. (As far as I remember)

So this may need to be configured depending on the system.

> From this viewpoint, I think the current implementation of dynamic tick
> falls short of this requirement. It cuts of the timer ticks only when
> all CPUs go idle.
> Apart from this observation, I have some others about the current dynamic tick
> patch:
> - All CPUs seem to cut off the same number of ticks (dyn_tick->skip). Isn't
> this wrong, considering that the timer list is per-CPU? This will cause
> some timers to be serviced much later than usual.

Yes if it's done on per-CPU basis. In the current setup the first interrupt
will kick the system off the dyn-tick state and the timers get checked again.

> - The fact that dyn_tick_state is global and accessed from all CPUs
> is probably a scalability concern, especially if we allow the ticks
> to be cut off on per-CPU basis.

From idling devices point of view, we still need some global variable I
believe. How else would you be able to tell all devices that the whole
system does not have any timers for next 2 seconds?

> - Again, when we allow this on a per-CPU basis, subsystems like
> RCU need to know the partial set of idle CPUs. RCU already does
> that thr' nohz_cpu_mask (which will need to replace dyn_cpu_map).

Sounds like that could work for dyn-tick too.

> - Looking at dyn_tick_timer_interrupt, would it be nice if we avoid calling
> do_timer_interrupt so many times and instead update jiffies to
> (skipped_ticks - 1) and then call do_timer_interrupt once? I think
> VST does it that way.

In the long run we would do the calculations in usecs and just emulate
jiffies from the hw timer. But yes, optimizing updating the time would be

> - dyn_tick->max_skip = 0xffffff / apic_timer_val;
> From my reading of Intel docs, APIC_TMICT is 32-bit. So why does the
> above calculation take only 24-bits into account? What am I missing here?

Hmm, could be a bug here, needs to be checked. Maybe 32-bit APIC timer is
optional support, or maybe I accidentally pulled the optional 24-bit support
from the ACPI PM timer.

But in any case on P4 systems the APIC timer is not the bottleneck as
stopping or reprogramming PIT also kills APIC. (This does not happen on P3
systems). So the bottleneck most likely is the length of PIT.

> I can take a shot at addressing these concerns in dynamic_tick patch, but it
> seems to me that VST has already addressed all these to a big extent. Had you
> considered VST before? The biggest bottleneck I see in VST going mainline is
> its dependency on HRT patch but IMO it should be possible to write a small patch
> to support VST w/o HRT.
> George, what do you think?

HRT + VST depend on APIC only, and does not use next_timer_interrupt().
CONFIG_NO_IDLE_HZ is already integrated in the mainline kernel for s390
and ARM.

You may also want to check out the ARM implementation as it does not have
the issues listed above, which are mostly x86 specific issues.



To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-08-08 09:29    [W:0.108 / U:14.764 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site