Re: [Lse-tech] AIM7 lockmeter for 29p Itanium ccNUMA

[Bill H. <ha...@au...>]

John Hawkes wrote:
> 
> All (29) CPUs
> 
> Start time: Fri Sep 14 07:03:06 2001
> End   time: Fri Sep 14 07:26:13 2001
> Delta Time: 1386.79 sec.
> Hash table slots in use:      408.
> Global read lock slots in use: 999.
> 

> 
>  34.6% 42.0%   37us(  15ms)  510us(2153ms)( 6.8%)  12836038 58.0% 42.0%    0%  kernel_flag
>   6.9% 41.3%   33us(8077us)  513us(2153ms)( 1.5%)   2896012 58.7% 41.3%    0%    ext2_get_block+0xd0
>  17.0% 64.7%  289us(  12ms)  659us(1018ms)(0.86%)    815966 35.3% 64.7%    0%    ext2_write_inode+0x50
>   1.1% 47.9%   20us(5817us)  514us( 830ms)(0.46%)    749250 52.1% 47.9%    0%    notify_change+0xd0
>   1.3% 64.9%   22us(6858us)  649us(1961ms)(0.88%)    842833 35.1% 64.9%    0%    schedule+0xd60
>   1.9% 34.5%  4.9us(3042us)  435us(2081ms)( 2.0%)   5448856 65.5% 34.5%    0%    sys_lseek+0x100
> 
> ==> kernel_flag is the dominant bottleneck for this workload on this
> ==> platform.  Kernprof profiling shows the top three CPU consumers
> ==> are sys_lseek(), ext2_get_block(), and the idle loop.
> 
I thought there was some discussion a while back about sys_lseek and the
kernel lock.  I have not followed the code path.  Does anyone know why the
kernel lock is required in this code path ?

Bill