mold is slower than gold linking mongodb

[igorsol]

Thanks for fixing #196. Now I can test mold with linking MongoDB binaries.

Environment: Ubuntu 20.04, 8 core intel processor, gcc 9.3.0, mold built from commit 04ad22d326d.

I failed to figure out how to use mold with scons build tool used in MongoDB builds. So I tested it this way:

1. run MongoDB build with "scons VERBOSE=1 core | tee verbose.txt". VERBOSE option forces scons to print full command lines of build tools.
2. From verbose.txt I extracted g++ command line used to link 'mongod' binary
3. Prepended extracted command line with 'mold -run'

Now testing results:

• without mold (gold is used): 2 min 23 sec
• with mold: from 3m51s to 4m11s

[igorsol]

Later I found that there is -Wl,--no-threads parameter on the generated command line. So I tried to replace it with -Wl,--threads. Result linking time is even longer: 4m 33s. I also tried to add -Wl,--thread-count=8 to the g++ command line but linking time is still 4m 40s.

[rui314]

Thank you for you report! I have trouble with building mongodb on my machine, so I couldn't reproduce the issue, but I'll try again.

[gruenich]

@igorsol I think you have to specify LINK or SHLINK to set the linker for Scons, see https://scons.org/doc/production/HTML/scons-user/apa.html

[rui314]

mold seems to be faster than gold on my machine. This is what I tried:

1. Build mongodb with python3 buildscripts/scons.py VERBOSE=1 install-mongod | tee log to get the command line argument to build build/opt/mongo/db/mongod
2. Run the same command with different linkers to rebuild mongod

Here is the result:

ld.lld 4.546s
ldmold 3.328s
ld.gold 59.935s

I run the linker commands with taskset 0xffff to simulate a 8-core 16-thread machine.

Are you sure you are using mold? mold leaves its identification string in a .comment section like this:

$ readelf -p .comment build/opt/mongo/db/mongod

String dump of section '.comment':
  [     0]  GCC: (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0
  [    2a]  GCC: (Ubuntu 10.3.0-1ubuntu1~20.04) 10.3.0
  [    56]  mold 1.0.1 (c3466b90212e7aae10fa0480d90d1ca01875cf88; compatible with GNU ld)

Please double check you are using mold with the above command.

[igorsol]

Yes, I'm sure I use mold. I checked this with readelf command. From your example I see you built mold from commit c3466b902 - I will recheck with this commit. Did you modify linker command to remove -Wl,--no-threads parameter?

[rui314]

My command line didn't have -Wl,-no-threads.

I wonder what exactly is your command. My link time of mongod is about 2 seconds, but you said it took 4 minutes and 33 seconds. It's very likely that your number includes something that is not linker.

[ishitatsuyuki]

While revisiting this issue https://github.com/rui314/mold/issues/730 comes to mind as a case of threading giving no advantage or even regressions.

Were you running Linux directly on your machine or relied on some kind of virtualization, including but not limited to WSL2?

[igorsol]

In my case that was linux guest in VitualBox on the Windows host. I will try to update my tests with the latest mold version.

[rui314]

I'm pretty sure that that's a performance issue of VirtualBox and/or Windows. mold is also slow on WSL2. My guess is that the virtual machines on Windows does not scale well for highly multi-threaded applications.