slow linking when used at WSL2

[CvH]

While trying to move our distribution (https://github.com/LibreELEC/LibreELEC.tv/pull/6875) to mold some speed regressions turned up. We investigated where they relate to and did some basic test. I tested it at WSL2 and compared it to native Linux at the same machine (8/16 core Ryzen 7) . Turned out that linking at WSL2 is a lot slower compared to native Linux while Gold and bfd is in similar ballpark across the OS.

I use WSL2 to compile the whole distribution for a while now and never saw any significant speed regressions (run at native ext4 not NTFS).

cat do-test

#!/bin/sh

if [ "$#" -lt 1 ]; then
  echo "usage: $0 count"
  exit 1
fi

MAX=$1
shift

IN="${XDG_RUNTIME_DIR}/t.c"
OBJ="${XDG_RUNTIME_DIR}/t.o"
OUT="${XDG_RUNTIME_DIR}/t"
cat << EOF > "${IN}"
#include <stdio.h>

int main(int argc, char** argv)
{
        printf("hello, world!\n");
        return 0;
}
EOF
gcc -c -o "${OBJ}" "$IN"

# args: message, args...
do_link() {
  count=0
  start="$(date +%s.%3N)"
  while [ $count -lt ${MAX} ]; do
    gcc "$@" -o "${OUT}" "${OBJ}"
    count=$((count+1))
  done
  end="$(date +%s.%3N)"
  elapsed="$(echo "${end}-${start}" | bc)"
  printf '%10.3f   %s\n' "${elapsed}" "$*"
}

COMMON_ARGS="-Wl,--as-needed"
MOLD_ARGS="-B/path/to/mold-1.4.2-x86_64-linux/libexec/mold"

do_link -fuse-ld=bfd ${COMMON_ARGS}
do_link -fuse-ld=gold ${COMMON_ARGS}
do_link ${MOLD_ARGS} ${COMMON_ARGS}
do_link ${MOLD_ARGS} ${COMMON_ARGS} -Wl,--no-threads
do_link ${MOLD_ARGS} ${COMMON_ARGS} -Wl,--thread-count=2
do_link ${MOLD_ARGS} ${COMMON_ARGS} -Wl,--thread-count=4
do_link ${MOLD_ARGS} ${COMMON_ARGS} -Wl,--thread-count=8
do_link ${MOLD_ARGS} ${COMMON_ARGS} -Wl,--thread-count=16

At Windows 11 WSL2 Ubuntu 22.04

time ./do-test 1000 -fuse-ld=bfd
real    0m19.087s
user    0m16.510s
sys     0m2.869s

time ./do-test 1000 -fuse-ld=gold
real    0m14.950s
user    0m12.782s
sys     0m2.451s

time ./do-test 1000 -B/path/to/mold-1.4.2-x86_64-linux/libexec/mold/
real    0m17.845s
user    0m11.314s
sys     0m1.682s

Linux Ubuntu 22.04 (Sorry nuked the exact numbers for Linux) At native Linux I have the expected gain compared to Gold and just need ~50% of the time to finish compared to Gold. So I think this is the real performance to expect.

[rui314]

My first guess is that multi-threading performance on WSL2 is not as good as the native Linux. Do you mind if I ask you to redo the test with -Wl,-no-threads to disable multi-threading? It's usually slower than without that flag because multi-threading usually boost its performance, but that may not be the case on WSL2.

[CvH]

I updated the script,

results at WSL

./do-test 1000
    10.339   -fuse-ld=bfd -Wl,--as-needed
     5.795   -fuse-ld=gold -Wl,--as-needed
     9.288   -B/home/le/mold-1.4.2-x86_64-linux/libexec/mold -Wl,--as-needed
     5.969   -B/home/le/mold-1.4.2-x86_64-linux/libexec/mold -Wl,--as-needed -Wl,--no-threads
     5.933   -B/home/le/mold-1.4.2-x86_64-linux/libexec/mold -Wl,--as-needed -Wl,--thread-count=2
     6.393   -B/home/le/mold-1.4.2-x86_64-linux/libexec/mold -Wl,--as-needed -Wl,--thread-count=4
     7.320   -B/home/le/mold-1.4.2-x86_64-linux/libexec/mold -Wl,--as-needed -Wl,--thread-count=8
     9.279   -B/home/le/mold-1.4.2-x86_64-linux/libexec/mold -Wl,--as-needed -Wl,--thread-count=16

[rui314]

OK, so it looks like there's something with WSL2. It doesn't scale at all but becomes slower when using more than one thread. You may want to report it to Microsoft.