zstd-dsb-repro.sh

#!/usr/bin/env sh
set -e

# Use this script to check if your platform has instability.
# Set ENWIK8 to the data you want to test. It doesn't necessarily
# have to be enwik8, any data should have the same results, but that
# is what I used. If you don't have it on your machine, just use another
# largeish file. Set COMPILERS to the compilers you want to test.

# It is good if the DSB cycles are always high and the MITE cycles are
# always low. You have performance instability if the cycles in DSB drop
# and the MITE cycles spike.

# The NICKSERV - BAD example is a case where the performance tanks.
# The NICKSERV - GOOD example is a case where the performance recovered.
# The NICKSERV - BEST example is when it is aligned at a 32+16 byte boundary.
# The DEVBIG - GOOD example is what every commit on the devbig looks like.
# You can see that the MITE cycles are 6x lower than even the best case on
# NICKSERV.

# NICKSERV - BAD
# Performance counter stats for './zstd -tq enwik8.zst' (20 runs):
#
#       461,827,836      cycles:u                                                      ( +-  0.03% )
#     1,301,501,797      instructions:u            #    2.82  insn per cycle           ( +-  0.00% )
#        67,331,424      idq.all_dsb_cycles_any_uops:u                                     ( +-  0.01% )
#       317,781,401      idq.all_mite_cycles_any_uops:u                                     ( +-  0.01% )
#                                                                                                                                                                                                                                                       # 0.1343907 +- 0.0000485 seconds time elapsed  ( +-  0.04% )

# NICKSERV - GOOD
# Performance counter stats for './zstd -tq enwik8.zst' (20 runs):
#
#       424,731,705      cycles:u                                                      ( +-  0.03% )
#     1,301,501,794      instructions:u            #    3.06  insn per cycle           ( +-  0.00% )
#       141,993,967      idq.all_dsb_cycles_any_uops:u                                     ( +-  0.00% )
#       192,747,913      idq.all_mite_cycles_any_uops:u                                     ( +-  0.00% )
#
#         0.1240921 +- 0.0000395 seconds time elapsed  ( +-  0.03% )

# NICKSERV - BEST
# Performance counter stats for './zstd -tq enwik8.zst' (20 runs):
#
#       423,652,867      cycles:u                                                      ( +-  0.06% )
#     1,301,504,853      instructions:u            #    3.07  insn per cycle           ( +-  0.00% )
#       226,354,088      idq.all_dsb_cycles_any_uops:u                                     ( +-  0.01% )
#        65,011,239      idq.all_mite_cycles_any_uops:u                                     ( +-  0.05% )
#
#          0.123488 +- 0.000150 seconds time elapsed  ( +-  0.12% )

# DEVBIG - GOOD
# Performance counter stats for './zstd -t enwik8.zst -q' (20 runs):
#
#       514,909,270      cycles                                                        ( +-  0.23% )  (24.79%)
#     1,311,267,666      instructions              #    2.55  insn per cycle           ( +-  0.20% )  (24.61%)
#       332,762,384      idq.all_dsb_cycles_any_uops                                     ( +-  0.19% )  (25.07%)
#        10,045,638      idq.all_mite_cycles_any_uops                                     ( +-  1.28% )  (24.96%)
#       253,960,156      idq.all_dsb_cycles_4_uops                                     ( +-  0.17% )  (24.78%)
#
#           0.17291 +- 0.00102 seconds time elapsed  ( +-  0.59% )

ENWIK8=/path/to/test/file
COMPILERS=(gcc)

zstd $ENWIK8 --no-compress-literals -qfo test.zst

for compiler in gcc; do
  for commit in 10f915f0 bc0e10ae c9ce6db b7b9d3; do
    git checkout $commit 2> /dev/null > /dev/null
    CC="$compiler" make clean zstd ZSTD_LEGACY_SUPPORT=0 2> /dev/null > /dev/null
    echo -e "\n\n$compiler - $commit\n"
    $compiler --version
    perf stat -r20 -e cycles,instructions,idq.all_dsb_cycles_any_uops,idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst 2>&1
  done
done

# Print the output of lscpu
lscpu