lkml.org 
[lkml]   [2009]   [Mar]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/5] [GIT PULL] updates for tip/tracing/ftrace

On Sun, 22 Mar 2009, Ingo Molnar wrote:

>
> ok, with Frederic we figured out the problem.
>
> What helped things most was this trace-dump output:
>
> 0) + 15.281 us | }
> 0) | handle_irq() {
> 1) + 35.871 us | }
> 1) | timespec_to_ktime() {
> 0) 4.608 us | }
> 0) | generic_handle_irq_desc() {
> 1) 4.097 us | }
> 1) + 14.171 us | }
> 0) 4.450 us | _spin_lock();
> 1) + 60.127 us | }
> 1) | ktime_get() {
> 0) | ack_apic_edge() {
> 1) | getnstimeofday() {
> 0) 6.486 us | }
> 0) 5.619 us | irq_complete_move();
> 1) 5.158 us | jiffies_read();
> 0) | move_native_irq() {
> 1) + 15.495 us | }
> 1) + 26.161 us | }
> 0) 5.631 us | }
> 1) 5.549 us | set_normalized_timespec();
> 0) + 16.304 us | }
> 0) | ack_APIC_irq() {
> 1) + 48.377 us | }
> 1) | timespec_to_ktime() {
> 0) 5.762 us | native_apic_mem_write();
> 1) 5.751 us | }
> 0) + 16.162 us | }
> 1) + 16.413 us | }
> 0) + 27.185 us | }
> 1) + 81.519 us | }
> 0) + 80.245 us | }
> 1) ! 154.606 us | }
> 0) | _spin_unlock() {
> 1) 5.743 us | tick_nohz_update_jiffies();
> 0) 5.781 us | }
> 1) ! 183.912 us | }
> 0) 5.327 us | preempt_schedule();
> 1) ! 202.575 us | }
> 0) + 25.827 us | }
>
> [...]
> 1) ! 2623.297 us | }
>
> i.e. all CPUs spend 2-3 milliseconds to handle a single tick. This
> is on a Core2 Extreme Edition 2.93 GHz CPU, so this kind of cost was
> unexpected.
>
> Until i saw this:
>
> CONFIG_TRACE_BRANCH_PROFILING=y
> CONFIG_PROFILE_ALL_BRANCHES=y
>
> that explains it all. The above sequence is two CPUs 'lock stepped'
> in a very high overhead series of cacheline ping-pongs. The
> ping-pongs happen due to every branch in the kernel doing:
>
> ______f.miss_hit[______r]++;
>
> where the branch info metadata is defined as global variables:
>
> static struct ftrace_branch_data \
> __attribute__((__aligned__(4))) \
> __attribute__((section("_ftrace_branch"))) \
>
> not only is it global, it's also false cacheline-shared due to a 4
> byte alignment.
>
> The proper solution would be to use percpu data and percpu_add()
> primitives for this.

Ug, that would increase the size tremendously. Remember, we have a data
structure for ever if statement in the kernel. I can't recall how much
memory it takes up now, but it was quite a bit. I can't imagine what it
would be like to multiply that by NR_CPUS.

Perhaps we could just check a single CPU?

if (!smp_processor_id()) {
[...]
}
Have CPU 0 be profiled only? Or at least make it an option?

-- Steve

>
> Anyway ... i turned off the branch tracer for my tests.



\
 
 \ /
  Last update: 2009-03-22 19:35    [from the cache]
©2003-2011 Jasper Spaans