ci: Run benchmarks in the CI

[mattiapitossi]

Motivation

Starting from issue #6539, I would like to introduce some consistent benchmarks using Gungraun to be able to run benchmarks on the CI. For the moment we don't run any of the benchmarks due to the fact that the Criterion benchmarks use a wall-time approach and cannot considered reliable for a CI.

Solution

The iai-calgrind library is now called gungraun and it's already used by some services (https://github.com/gungraun/gungraun/network/dependents) including Rust's compiler-builtins. I wanted to start with spawn first to understand if we want to proceed in this direction.

Other Considerations

• Gungraun uses Valgrind which is only supported by Unix, therefore benchmarks on other system can only be run through Docker
• We are still discussing about whether we want to use Bencher or not

Results

Multiple runs on M1 (Docker)

ct_spawn::group::spawn_rt basic:single_rt()
  Instructions:                        8133|8133                 (No change)
  L1 Hits:                            11844|11844                (No change)
  LL Hits:                               61|61                   (No change)
  RAM Hits:                             433|433                  (No change)
  Total read+write:                   12338|12338                (No change)
  Estimated Cycles:                   27304|27304                (No change)
ct_spawn::group::spawn_rt threaded:multi_rt()
  Instructions:                        7009|7009                 (No change)
  L1 Hits:                             9858|9858                 (No change)
  LL Hits:                              104|104                  (No change)
  RAM Hits:                             273|273                  (No change)
  Total read+write:                   10235|10235                (No change)
  Estimated Cycles:                   19933|19933                (No change)
ct_spawn::group::spawn_rt_100 basic:single_rt()
  Instructions:                      267414|267414               (No change)
  L1 Hits:                           407758|407758               (No change)
  LL Hits:                              340|340                  (No change)
  RAM Hits:                             914|914                  (No change)
  Total read+write:                  409012|409012               (No change)
  Estimated Cycles:                  441448|441448               (No change)
ct_spawn::group::spawn_rt_100 threaded:multi_rt()
  Instructions:                      156427|156427               (No change)
  L1 Hits:                           231658|231659               (-0.00043%) [-1.00000x]
  LL Hits:                              486|485                  (+0.20619%) [+1.00206x]
  RAM Hits:                             699|699                  (No change)
  Total read+write:                  232843|232843               (No change)
  Estimated Cycles:                  258553|258549               (+0.00155%) [+1.00002x]

Gungraun result: Ok. 4 without regressions; 0 regressed; 0 filtered; 4 benchmarks finished in 1.12675s

[mattiapitossi]

TODO: There's an open discussion on Discord whether we want to use Bencher or not

[Darksonn]

We can try using bencher. I understand that we need to do something special to set it up. Could you share more details on what is required?

[mattiapitossi]

Sure! Technically, it should be enough to create an account on https://bencher.dev/auth/signup and generate an API token. Then, a new secret called BENCHER_API_TOKEN with the API key should be created under repo settings as described here