[dpdk-dev] [PATCH] mbuf: replace c memcpy() code semantics with optimized rte_memcpy()

[Olivier Matz]

Hi Sarosh,

On Tue, Jul 28, 2020 at 06:30:46PM +0500, Sarosh Arif wrote:
> Hello,
> The following things made me think that rte_memcpy() is more optimized
> than memcpy():
> 1. dpdk documentation recommends to use rte_memcpy() instead of memcpy():
>     https://doc.dpdk.org/guides/prog_guide/writing_efficient_code.html
> 2. Here some benchmarks are available:
>     https://software.intel.com/content/www/us/en/develop/articles/performance-optimization-of-memcpy-in-dpdk.html
> 3. rte_memcpy() has __attribute__((always_inline)) associated with it,
> so compiler also tries to inline it.
> 
> Using rte_memcpy() everywhere ensures consistency in code-base.
> Here are the results of the performance number measurement using "perf":
> 
> rte_memcpy()
> 
>  Performance counter stats
>           1.573864      task-clock (msec)         #    0.898 CPUs
> utilized
>                  0      context-switches          #    0.000 K/sec
>                  0      cpu-migrations            #    0.000 K/sec
>                342      page-faults               #    0.217 M/sec
>          5,483,016      cycles                    #    3.484 GHz
>          5,554,017      instructions              #    1.01  insn per
> cycle
>          1,114,593      branches                  #  708.189 M/sec
>             33,796      branch-misses             #    3.03% of all
> branches
>          1,369,247      L1-dcache-loads           #  869.991 M/sec
>      <not counted>      L1-dcache-load-misses
>                (0.00%)
>      <not counted>      LLC-loads
>                (0.00%)
>      <not counted>      LLC-load-misses
>                (0.00%)
> 
>        0.001753373 seconds time elapsed
> 
> 
> 
> memcpy()
> 
>  Performance counter stats
>           1.631135      task-clock (msec)         #    0.902 CPUs
> utilized
>                  0      context-switches          #    0.000 K/sec
>                  0      cpu-migrations            #    0.000 K/sec
>                342      page-faults               #    0.210 M/sec
>          5,676,549      cycles                    #    3.480 GHz
>                (73.99%)
>          5,739,593      instructions              #    1.01  insn per
> cycle
>          1,141,121      branches                  #  699.587 M/sec
>             34,553      branch-misses             #    3.03% of all
> branches
>          1,417,494      L1-dcache-loads           #  869.023 M/sec
>             67,312      L1-dcache-load-misses     #    4.75% of all
> L1-dcache hits    (26.01%)
>      <not counted>      LLC-loads
>                (0.00%)
>      <not counted>      LLC-load-misses
>                (0.00%)
> 
>       0.001808500 seconds time elapsed
> 

Can you give more details about your use-case? I mean what code
are you running for this benchmark.

I'll tend to agree with Stephen: memcpy() with a constant (small) size
should directly be replaced by the optimal code for this architecture.

rte_memcpy() uses vector instructions, and is probably better than
libc's memcpy for larger copies.

Thanks,
Olivier


> 
> 
> On Thu, Jul 23, 2020 at 8:47 PM Stephen Hemminger
> <stephen at networkplumber.org> wrote:
> >
> > On Thu, 23 Jul 2020 12:02:40 +0500
> > Sarosh Arif <sarosh.arif at emumba.com> wrote:
> >
> > > Since rte_memcpy is more optimized it should be used instead of memcpy
> > >
> > > Signed-off-by: Sarosh Arif <sarosh.arif at emumba.com>
> >
> > Really did you measure this.
> > For fixed size structures, compiler can inline memcpy small set of instructions.