> Date: Thu, 19 Aug 2010 14:20:08 +0800
> From: LionSky <[email protected]>
> Subject: High TLB Shootdowns interrupt
> To: [email protected]
> Message-ID:
> <[email protected]>
> Content-Type: text/plain; charset=ISO-8859-1
>
> Hi all,
> Recently, I used PowerTop to tune an workload power behavior, but
> found a new interrupt(TLB shootdowns) from its wakeups list as
> follows:
>
> Top causes for wakeups:
> 66.2% (1806.2) [ 0] [kernel scheduler] Load balancing tick
> 16.5% (451.1) [ ] [TLB shootdowns] <kernel IPI>
> 9.8% (267.7)D [ ] chrome
> 4.5% (121.5) [ ] [Rescheduling interrupts] <kernel IPI>
> 1.7% ( 46.9) [ ] [hda_intel] <interrupt>
> 0.8% ( 23.2) [ ] [i915] <interrupt>
> 0.0% ( 0.1)D [ 78] btrfs-transacti
> 0.2% ( 4.9) [ 737] syndaemon
> 0.1% ( 1.9) [ ] [ahci] <interrupt>
>
> There is about 450 TLB shootdowns interrupt per second, so high. ..
> I searched google for help, but only found what is TLB shootdowns:
> "A TLB Shootdown occurs when a process restricts access to a page in
> shared memory and must interrupt processes using that memory space on
> other processors so they flush their TLB tables"
>
> I have no idea whether "TLB shootdowns" interrupt has any impact on
> power, or even workload's performance.
> Do anyone have some idea about that?
>
>
> Thanks a lot~~
>
>
>
> ------------------------------
>
> Message: 2
> Date: Thu, 19 Aug 2010 09:38:21 -0700
> From: Auke Kok <[email protected]>
> Subject: Re: High TLB Shootdowns interrupt
> To: LionSky <[email protected]>
> Cc: "[email protected]" <[email protected]>
> Message-ID: <[email protected]>
> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
>
> On 08/18/2010 11:20 PM, LionSky wrote:
>> Hi all,
>> Recently, I used PowerTop to tune an workload power behavior, but
>> found a new interrupt(TLB shootdowns) from its wakeups list as
>> follows:
>>
>> Top causes for wakeups:
>> 66.2% (1806.2) [ 0] [kernel scheduler] Load balancing tick
>> 16.5% (451.1) [ ] [TLB shootdowns]<kernel IPI>
>> 9.8% (267.7)D [ ] chrome
>> 4.5% (121.5) [ ] [Rescheduling interrupts]<kernel IPI>
>> 1.7% ( 46.9) [ ] [hda_intel]<interrupt>
>> 0.8% ( 23.2) [ ] [i915]<interrupt>
>> 0.0% ( 0.1)D [ 78] btrfs-transacti
>> 0.2% ( 4.9) [ 737] syndaemon
>> 0.1% ( 1.9) [ ] [ahci]<interrupt>
>>
>> There is about 450 TLB shootdowns interrupt per second, so high. ..
>> I searched google for help, but only found what is TLB shootdowns:
>> "A TLB Shootdown occurs when a process restricts access to a page in
>> shared memory and must interrupt processes using that memory space on
>> other processors so they flush their TLB tables"
>>
>> I have no idea whether "TLB shootdowns" interrupt has any impact on
>> power, or even workload's performance.
>> Do anyone have some idea about that?
>
> TLB shootdowns are caused by an application in essence, and will indeed
> cause an impact on power. The most likely culprit is some memory-hungry
> application(chrome?) forcing pages to be swapped out (or something
> similar) that were shared between cores, and so each core now needs to
> update, keeping them out of a deeper C-state.
>
> Auke
>
Thanks~
I checked memory footprint and found that memory is not bottleneck and
there is not any swap operation during workload running.
But I found some things thru workload profiling. A library
implemented memory management itself and identified two api:
MemoryAllocate() and MemoryFree().
The following is some call graphs which have something to do with TLB.
--- flush_tlb_others_ipi
flush_tlb_mm
|
|--88.89%-- unmap_vmas
| unmap_region
| do_munmap
| sys_munmap
| sysenter_do_call
| 0xb77c4424
| MemoryServices::MemoryFree(void*)
| | --50.00%-- flush_tlb_mm
| | zap_page_range
| | sys_madvise
| | sysenter_do_call
| | 0xb77c4424
| |
MemoryServices::MemoryFree(void*)
--20.00%-- flush_tlb_mm
mprotect_fixup
sys_mprotect
sysenter_do_call
0xb77c4424
0xb762910a
__libc_malloc
malloc
MemoryServices::MemoryAllocate(unsigned int)
Do these operations lead to high "TLB shootdowns" interrupt?
_______________________________________________
Power mailing list
[email protected]
http://www.bughost.org/mailman/listinfo/power
