Re: [patch V6 00/16] Improve /proc/interrupts further

[Shrikanth Hegde <sshegde@linux.ibm.com>]

Hi Thomas.

On 5/20/26 8:57 PM, Thomas Gleixner wrote:
> Shrikanth!
> 
> On Wed, May 20 2026 at 02:48, Shrikanth Hegde wrote:
>> On 5/18/26 1:31 AM, Thomas Gleixner wrote:
>>
>> Ran perf stat -r 1000 cat /proc/interrupts > tmp.txt
>> and Observed minimal improvement with series.
> 
> Can you redirect it to /dev/null instead to take the file operations out
> of the picture?
> 

Yes. Did "perf stat -r 1000 cat /proc/interrupts > /dev/null".
It shows better improvement with the series compared to file write.


>> Base:
>>            1,313,263      cycles:HG                        #    4.063 GHz                         ( +-  0.17% )
>>            2,172,511      instructions:HG                  #    1.65  insn per cycle              ( +-  0.05% )
> 
>> v6 series:
>>            1,224,666      cycles:HG                        #    4.058 GHz                         ( +-  0.25% )
>>            1,667,435      instructions:HG                  #    1.36  insn per cycle              ( +-  0.08% )
> 
> Interesting. The number of instructions goes down by 20+%, but at the
> same time IPC drops too.
> 

base:

  Performance counter stats for 'cat /proc/interrupts' (1000 runs):

               0.32 msec task-clock:HG                    #    0.615 CPUs utilized               ( +-  0.17% )
                  0      context-switches:HG              #    0.000 /sec
                  0      cpu-migrations:HG                #    0.000 /sec
                 44      page-faults:HG                   #  136.347 K/sec                       ( +-  0.03% )
          1,310,621      cycles:HG                        #    4.061 GHz                         ( +-  0.17% )
          2,182,042      instructions:HG                  #    1.66  insn per cycle              ( +-  0.04% )
            372,189      branches:HG                      #    1.153 G/sec                       ( +-  0.04% )
              4,710      branch-misses:HG                 #    1.27% of all branches             ( +-  0.33% )

        0.000525061 +- 0.000000889 seconds time elapsed  ( +-  0.17% )

v6:
  Performance counter stats for 'cat /proc/interrupts' (1000 runs):

               0.28 msec task-clock:HG                    #    0.577 CPUs utilized               ( +-  0.25% )
                  0      context-switches:HG              #    0.000 /sec
                  0      cpu-migrations:HG                #    0.000 /sec
                 44      page-faults:HG                   #  155.906 K/sec                       ( +-  0.03% )
          1,144,964      cycles:HG                        #    4.057 GHz                         ( +-  0.24% )
          1,628,375      instructions:HG                  #    1.42  insn per cycle              ( +-  0.07% )
            271,934      branches:HG                      #  963.546 M/sec                       ( +-  0.07% )
              4,683      branch-misses:HG                 #    1.72% of all branches             ( +-  0.49% )
                  
         0.00048895 +- 0.00000114 seconds time elapsed  ( +-  0.23% )   << 7-8% improvement.


v6+ ppc_hack

  Performance counter stats for 'cat /proc/interrupts' (1000 runs):
          
               0.27 msec task-clock:HG                    #    0.582 CPUs utilized               ( +-  0.15% )
                  0      context-switches:HG              #    0.000 /sec
                  0      cpu-migrations:HG                #    0.000 /sec
                 44      page-faults:HG                   #  160.232 K/sec                       ( +-  0.03% )
          1,113,983      cycles:HG                        #    4.057 GHz                         ( +-  0.15% )
          1,501,868      instructions:HG                  #    1.35  insn per cycle              ( +-  0.07% )
            251,432      branches:HG                      #  915.627 M/sec                       ( +-  0.06% )
              4,528      branch-misses:HG                 #    1.80% of all branches             ( +-  0.18% )
  
        0.000472057 +- 0.000000668 seconds time elapsed  ( +-  0.14% )  << only slightly better.


>> Looking at powerpc arch_show_interrupts,
>> It could use the similar set of optimizations.
>> - move to array based
>> - use irq_proc_emit_counts
>> - some interrupts such as machine check, is hardly set. set skip_vector.
>>
>>
>> Copilot suggested below diff to quickly try irq_proc_emit_counts integration.
>> It showed little gains compared to v6. So it maybe worth fixing that in the
>> right way. (similar to x86 stuff you have done)
>>
>>    Performance counter stats for 'cat /proc/interrupts' (1000 runs):
>>
>>                 0.29 msec task-clock:HG                    #    0.586 CPUs utilized               ( +-  0.22% )
>>                    0      context-switches:HG              #    0.000 /sec
>>                    0      cpu-migrations:HG                #    0.000 /sec
>>                   44      page-faults:HG                   #  153.067 K/sec                       ( +-  0.03% )
>>            1,166,567      cycles:HG                        #    4.058 GHz                         ( +-  0.22% )
>>            1,475,365      instructions:HG                  #    1.26  insn per cycle              ( +-  0.09% )
>>              249,051      branches:HG                      #  866.397 M/sec                       ( +-  0.10% )
>>                5,104      branch-misses:HG                 #    2.05% of all branches             ( +-  0.33% )
>>
>>          0.000490211 +- 0.000000992 seconds time elapsed  ( +-  0.20% )   <<< 3-4% improvements.
> 
> Again IPC drops ....

Yes. IPC dropping is consistent. I see the same trend in (PATCH 1/16) in the series.
Copying that snippet below.

Before:
  8,932,242      instructions      #    1.66  insn per cycle  ( +-  0.34% )
After:
  7,020,982      instructions      #    1.30  insn per cycle  ( +-  0.52% )


So it might be common pattern across archs. Maybe perf stat subsystem is slow
enough it doesn't shows the aboslute benefit.

In addition, I ran "perf stat -a -r 1000 cat /proc/interrupts > /dev/null"
It is now 10x slower. IPC is same with series And improvement vanishes.
So heavier the infra testing it, gains are getting minimal i guess.
But i don't see any regression.

As you said in the cover-letter, the micro loops you ran maybe the best way to evaluate it.
If you have the code in shareable form, I can give it a try.

Other than that, code improvement looks good to me.