Re: [PATCH 02/14] perf/bench: Default to all routines in 'perf bench mem'

[Ingo Molnar <mingo@kernel.org>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Mon, Oct 19, 2015 at 1:04 AM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> >         triton:~> perf bench mem all
> >         # Running mem/memcpy benchmark...
> >         Routine default (Default memcpy() provided by glibc)
> >                4.957170 GB/Sec (with prefault)
> >         Routine x86-64-unrolled (unrolled memcpy() in arch/x86/lib/memcpy_64.S)
> >                4.379204 GB/Sec (with prefault)
> >         Routine x86-64-movsq (movsq-based memcpy() in arch/x86/lib/memcpy_64.S)
> >                4.264465 GB/Sec (with prefault)
> >         Routine x86-64-movsb (movsb-based memcpy() in arch/x86/lib/memcpy_64.S)
> >                6.554111 GB/Sec (with prefault)
> 
> Is this skylake? And why are the numbers so low? Even on my laptop
> (Haswell), I get ~21GB/s (when setting cpufreq to performance).

No, this was on my desktop, which is a water cooled IvyBridge running at 3.6GHz:

 processor       : 11
 vendor_id       : GenuineIntel
 cpu family      : 6
 model           : 62
 model name      : Intel(R) Core(TM) i7-4960X CPU @ 3.60GHz
 stepping        : 4
 microcode       : 0x416
 cpu MHz         : 1303.031
 cache size      : 15360 KB

and I didn't really think about the validity of the numbers when I made the 
changelog, as I rarely benchmark on this box, due to it having various desktop 
loads running all the time.

AAs you noticed the results are highly variable with default settings:

  triton:~/tip> taskset 1 perf stat --null --repeat 10 perf bench mem memcpy -f x86-64-movsb 2>&1 | grep GB
       5.580357 GB/sec
       5.580357 GB/sec
      16.551907 GB/sec
      16.551907 GB/sec
      15.258789 GB/sec
      16.837284 GB/sec
      16.837284 GB/sec
      16.837284 GB/sec
      16.551907 GB/sec
      16.837284 GB/sec

They get more reliable with '-l 10000' (10,000 loops instead of the default 1):

  triton:~/tip> taskset 1 perf stat --null --repeat 10 perf bench mem memcpy -f x86-64-movsb -l 10000 2>&1 | grep GB
      15.483591 GB/sec
      16.975429 GB/sec
      17.088396 GB/sec
      20.920407 GB/sec
      21.346655 GB/sec
      21.322372 GB/sec
      21.338306 GB/sec
      21.342130 GB/sec
      21.339984 GB/sec
      21.373145 GB/sec

that's purely cached. Also note how after a few seconds it gets faster, due to 
cpufreq as you suspected.

So once I fix the frequency of all cores to the max, I get much more reliable 
results:

  triton:~/tip> taskset 1 perf stat --null --repeat 10 perf bench mem memcpy -f x86-64-movsb -l 10000 2>&1 | grep -E 'GB|elaps'
      21.356879 GB/sec
      21.378526 GB/sec
      21.351976 GB/sec
      21.375203 GB/sec
      21.369824 GB/sec
      21.353236 GB/sec
      21.283708 GB/sec
      21.380679 GB/sec
      21.347915 GB/sec
      21.378572 GB/sec
       0.459286278 seconds time elapsed                                          ( +-  0.04% )

I'll add a debug check to 'perf bench' to warn about systems that have variable 
cpufreq running - this is too easy a mistake to make :-/

So with the benchmark stabilized, I get the following results:

  triton:~/tip> taskset 1 perf bench mem memcpy -f all -l 10000
  # Running 'mem/memcpy' benchmark:
  # function 'default' (Default memcpy() provided by glibc)
  # Copying 1MB bytes ...
 
        18.356783 GB/sec
  # function 'x86-64-unrolled' (unrolled memcpy() in arch/x86/lib/memcpy_64.S)
  # Copying 1MB bytes ...
 
        16.294889 GB/sec
  # function 'x86-64-movsq' (movsq-based memcpy() in arch/x86/lib/memcpy_64.S)
  # Copying 1MB bytes ...
 
        15.760032 GB/sec
  # function 'x86-64-movsb' (movsb-based memcpy() in arch/x86/lib/memcpy_64.S)
  # Copying 1MB bytes ...

        21.145818 GB/sec

which matches your observations:

> It's interesting that 'movsb' for you is so much better. It's been
> promising before, and it *should* be able to do better than manual
> copying, but it's not been that noticeable on the machines I've
> tested. But I haven't ued Skylake or Broadwell yet.
> 
> cpufreq might be making a difference too. Maybe it's just ramping up
> the CPU? Or is that really repeatable?

So modulo the cpufreq multiplier it seems repeatable on this IB system - will try 
it on SkyLake as well.

Before relying on it I also wanted to implement the following 'perf bench' 
improvements:

 - make it more representative of kernel usage by benchmarking a list of
   characteristic lengths, not just the single stupid 1MB buffer. At smaller 
   buffer sizes I'd expect MOVSB to have even more of a fundamental advantage (due 
   to having all the differentiation in hardware) - but we don't know the
   latencies of those cases, some of which are in microcode I suspect.

 - measure aligned/unaligned buffer address and length effects as well

 - measure cache-cold numbers as well. This is pretty hard but not impossible.

With that we could start validating our fundamental memory op routines in 
user-space.

Thanks,

	Ingo