[PATCH] D147493: [ELF] Cap parallel::strategy to 8 cores when --threads= is unspecified

Wed Apr 5 06:49:51 PDT 2023

peter.smith added a comment.

In D147493#4245761 <https://reviews.llvm.org/D147493#4245761>, @mjguzik wrote:

> In D147493#4245744 <https://reviews.llvm.org/D147493#4245744>, @peter.smith wrote:
>
>> I have run a limited experiment linking clang on a 64 core AArch64 machine. Results seem to be similar to x86_64. No benefit seen linking clang past 8 threads. On the machine I had access to I was very limited in disk space so I couldn't check the debug build. From prior experience debug builds seems to benefit more from more threads as there is quite a lot of data to write out in the most parallel part of LLD.
>
> bare metal aarch64 with sufficient(tm) disk space is not hard to come by.
>
> if you think the debug thingy is worth checking out and you can't get sufficient hw as is, you can contact me in private and I'll arrange temporary access to something.
>
> perhaps while the machine is disk starved, it has a lot of ram? as in perhaps you could use a tmpfs mount for the output?

Thanks for the offer and for the suggestion of tmpfs. I'm more personal time limited than anything else. I really just wanted to check if there was any significant difference between AArch64 and X86_64 on higher core counts and so far there doesn't look to be anything significant. It looks like Windows (or Unreal) may favour a higher count though.

FWIW to link clang without debug.

| CPUS | user | system | elapsed |
| 1    | 0.57 | 0.34   | 1.06    |
| 8    | 0.66 | 0.65   | .055    |
| 16   | 0.65 | 1.08   | 0.56    |
| 32   | 0.65 | 2.02   | 0.56    |
| 64   | 0.73 | 1.94   | 0.58    |
|

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147493/new/

https://reviews.llvm.org/D147493