Adhemerval Zanella 80f389ed7e resolv: Optimize inet_ntop ...

The benchtests/inet_ntop_ipv4 and benchtests/inet_ntop_ipv6 profile
shows that most of time is spent in costly sprint operations:

$ perf record ./benchtests/bench-inet_ntop_ipv4 && perf report --stdio
[...]
    38.53%  bench-inet_ntop  libc.so               [.] __printf_buffer
    18.69%  bench-inet_ntop  libc.so               [.] __printf_buffer_write
    11.01%  bench-inet_ntop  libc.so               [.] _itoa_word
     8.02%  bench-inet_ntop  bench-inet_ntop_ipv4  [.] bench_start
     6.99%  bench-inet_ntop  libc.so               [.] __memmove_avx_unaligned_erms
     3.86%  bench-inet_ntop  libc.so               [.] __strchrnul_avx2
     2.82%  bench-inet_ntop  libc.so               [.] __strcpy_avx2
     1.90%  bench-inet_ntop  libc.so               [.] inet_ntop4
     1.78%  bench-inet_ntop  libc.so               [.] __vsprintf_internal
     1.55%  bench-inet_ntop  libc.so               [.] __sprintf_chk
     1.18%  bench-inet_ntop  libc.so               [.] __GI___inet_ntop

$ perf record ./benchtests/bench-inet_ntop_ipv6 && perf report --stdio
    35.44%  bench-inet_ntop  libc.so               [.] __printf_buffer
    14.35%  bench-inet_ntop  libc.so               [.] __printf_buffer_write
    10.27%  bench-inet_ntop  libc.so               [.] __GI___inet_ntop
     7.93%  bench-inet_ntop  libc.so               [.] _itoa_word
     7.00%  bench-inet_ntop  libc.so               [.] __sprintf_chk
     6.20%  bench-inet_ntop  libc.so               [.] __vsprintf_internal
     5.26%  bench-inet_ntop  libc.so               [.] __strchrnul_avx2
     5.05%  bench-inet_ntop  bench-inet_ntop_ipv6  [.] bench_start
     3.70%  bench-inet_ntop  libc.so               [.] __memmove_avx_unaligned_erms
     2.11%  bench-inet_ntop  libc.so               [.] __printf_buffer_done

A new implementation is used instead:

  * The printf usage is replaced with an expanded function that prints
    either an IPv4 octet or an IPv6 quartet;

  * The strcpy is replaced with a memcpy (since ABIs usually tends to
    optimize the latter);

  * For IPv6, the '::' shorthanding is done in-place instead of using
    a temporary buffer.

  * An temporary buffer is used iff the size if larger than
    INET_ADDRSTRLEN/INET6_ADDRSTRLEN.

  * Inline is used for both inet_ntop4 and inet_ntop6,

The code is significand rewrote, so I take this requires a new license.

The performance results on aarch64 Neoverse1 with gcc 14.2.1:

* master

aarch64-linux-gnu-master$ ./benchtests/bench-inet_ntop_ipv4
  "inet_ntop_ipv4": {
   "workload-ipv4-random": {
    "duration": 1.43067e+09,
    "iterations": 8e+06,
    "reciprocal-throughput": 178.572,
    "latency": 179.096,
    "max-throughput": 5.59997e+06,
    "min-throughput": 5.58359e+06
   }
aarch64-linux-gnu-master$ ./benchtests/bench-inet_ntop_ipv6
  "inet_ntop_ipv6": {
   "workload-ipv6-random": {
    "duration": 1.68539e+09,
    "iterations": 4e+06,
    "reciprocal-throughput": 421.307,
    "latency": 421.388,
    "max-throughput": 2.37357e+06,
    "min-throughput": 2.37311e+06
   }
  }

* patched

aarch64-linux-gnu$ ./benchtests/bench-inet_ntop_ipv4
  "inet_ntop_ipv4": {
   "workload-ipv4-random": {
    "duration": 1.06133e+09,
    "iterations": 5.6e+07,
    "reciprocal-throughput": 18.8482,
    "latency": 19.0565,
    "max-throughput": 5.30555e+07,
    "min-throughput": 5.24755e+07
   }
  }
aarch64-linux-gnu$ ./benchtests/bench-inet_ntop_ipv6
  "inet_ntop_ipv6": {
   "workload-ipv6-random": {
    "duration": 1.01246e+09,
    "iterations": 2.4e+07,
    "reciprocal-throughput": 42.5576,
    "latency": 41.8139,
    "max-throughput": 2.34976e+07,
    "min-throughput": 2.39155e+07
   }
  }

Checked on aarch64-linux-gnu and x86_64-linux-gnu.
Reviewed-by: DJ Delorie <dj@redhat.com>

2025-06-24 13:34:20 -03:00

4.9 KiB

Raw Blame History

View Raw