feat: github proving benchmark

[atanmarko]

Resolves #651

Here we download proving_benchmark_results.txt artifact form previous succesfull run.
We run the proving job with $BENCHMARK_WITNESS file.
We append the resulting time to the simple textual result file and upload it to artifacts repo.

Job is executed every night at 4AM and manually.

[sai-deng]

I think you can still use perf stat to capture CPU times (user and system time) in addition to the total wall clock time. This would give more granular insight into CPU performance.

[atanmarko]

Making it draft until we find the appropriate testing data.

[0xaatif]

info: is there somewhere I can read about the different CI runners we have, and how they're spec'd out/intended to be used?

[atanmarko]

This one is created by Ben for this particular benchmark. Best available.

[0xaatif]

I think I'd prefer storing the data differently - how do you feel about saving off wide information as follows?

Get full information from time

/usr/bin/time \
  --format='{"command":"%C","avg_unshared_data_kb":%D,"elapsed_real_hms":"%E","pg_fault_io_major_ct":%F,"fs_inputs_ct":%I,"avg_mem_kb":%K,"max_rss_kb":%M,"fs_outputs_ct":%O,"cpu_pc":"%P","pg_fault_minor_ct":%R,"cpu_kernel_secs":%S,"cpu_user_secs":%U,"swap_ct":%W,"avg_shared_text_kb":%X,"system_pg_size_bytes":%Z,"ctx_switch_forced_ct":%c,"elapsed_real_secs":%e,"signals_recv_ct":%k,"avg_unshared_stack_kb":%p,"socket_recv_ct":%r,"socket_sent_ct":%s,"avg_rss_kb":%t,"ctx_switch_yield_ct":%w,"exit_status":%x}' \
  --output time.json \
  "$COMMAND" \
  "$ARGS"

Combine information as follows

{
  "commit": $hash,
  "date": $date,
  "runner": {
    "name": $runner,
    "cores": ...,
    "arch": ...,
  },
  "time": {
    // from time.json
  },
}

I would strongly prefer moving away from bash scripts too - can we do the above in xtask?

[atanmarko]

@0xaatif I am fine to switch to xtask and switch from bash to Rust. Could we perform the switch as a separate task I'll open the ticker for?

[0xaatif]

Ok let's just get something running then

[Nashtare]

I'd also ideally have this process automated, flagging regressions directly but good as a first shot

[atanmarko]

I'd also ideally have this process automated, flagging regressions directly but good as a first shot

Yes, switching to Rust processing of the results, I think we could raise some flags and alarms in case of regressions.