chore(gpu): add benchmarking for 8xsxm5 with two processes

[andrei-stoian-zama]

Runs TPS benchmarks in parallel in two processes, starting at the same time with synchronization. Aggregates the bnech results and sends the aggregated TPS count to slab

---

This change is 

[guillermo-oyarzun]

This file shouldn't be included

[guillermo-oyarzun]

I understand this is the logic that makes each process select only a sub group of gpus, but in practice all the gpus are visible for each process, i think it would be easier if when we launch the benchmark we filter the gpus with CUDA_VISIBLE_DEVICES, then all this extra logic is not necessary, and it simulates in a better way what would happen in reality

[soonum]

Divide and conquer ERC7984 transfer ⚔️ !

@soonum reviewed 10 files and all commit messages, and made 15 comments.
Reviewable status: all files reviewed, 14 unresolved discussions (waiting on andrei-stoian-zama, IceTDrinker, soonum, and SouchonTheo).

---

Makefile line 1958 at r1 (raw file):

	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off --

.PHONY: bench_hlapi_erc7984_multi_group_gpu # Runs ERC7984 bench in two processes (half of gpus for each) and aggregates results

This recipe doesn't aggregate the results.

Suggestion:

.PHONY: bench_hlapi_erc7984_multi_group_gpu # Runs ERC7984 throughput benchmarks in two processes (half of gpus for each)

---

.github/workflows/benchmark_gpu.yml line 79 at r1 (raw file):

          - multi_bit_documentation
          - classical_documentation + multi_bit_documentation

This newline is not needed.

---

.github/workflows/benchmark_gpu_common.yml line 263 at r1 (raw file):

        if: ${{ inputs.command == 'hlapi_erc7984_multi_group' }}
        run: |
          pwd

I guess this is a debug command.

---

.github/workflows/benchmark_summary.yml line 117 at r1 (raw file):

      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}

  run-benchmarks-gpu-erc7984-multi-group:

In which category this benchmark falls ? A new line in ERC7984 dashboard ?

---

backends/tfhe-cuda-backend/CLAUDE.md line 1 at r1 (raw file):

# CLAUDE.md

I'd rather add this file in another PR.
This is not related to the benchmark changes.

---

ci/merge_multi_group_results.py line 2 at r1 (raw file):

#!/usr/bin/env python3
# TODO: ADD COMMENT

TODO to be done 😏 (same everywhere else)

---

ci/merge_multi_group_results.py line 8 at r1 (raw file):

# TODO: ADD COMMENT
def merge_multi_group_results(file1, file2, output_file):

Can't we make it generic for N files ? This should be doable as long as we use *files and keep output_file as the last argument in the command-line.

---

ci/merge_multi_group_results.py line 37 at r1 (raw file):

    with open(output_file, "w") as f:
        json.dump(result, f, indent=2)

I think Slab doesn't support pretty-printed JSON. To be tested.

---

tfhe-benchmark/benches/high_level_api/erc7984.rs line 552 at r1 (raw file):

    let params_name = params.name();

    // 300 * num_gpus seems to be enough for maximum throughput on 8xH100 SXM5

Comment to update.

---

tfhe-benchmark/src/utilities.rs line 483 at r1 (raw file):

}

pub fn get_bench_instances() -> Option<usize> {

Maybe renaming it toget_bench_gpu_instances() to avoid confusion.

---

tfhe-benchmark/src/utilities.rs line 491 at r1 (raw file):

}

/// TODO: ADD COMMENT

I'm adding a comment 😂 .
More seriously, it's more than needed since it brings libc objects handling in native Rust file.

---

tfhe-benchmark/src/utilities.rs line 493 at r1 (raw file):

/// TODO: ADD COMMENT
#[cfg(target_os = "linux")]
pub fn bench_sync_barrier(num_instances: usize) {

Maybe we can have a cleaning function that's is called before running the benchmark in the caller? That would remove the need to manually remove the files in the Makefile.
By using a struct that embed all three sem_*, pass it to this cleaning function and simply does

    unsafe {
        libc::sem_close(mutex);
        libc::sem_close(arrive);
        libc::sem_close(gate);
        libc::sem_unlink(sem_mutex.as_ptr());
        libc::sem_unlink(sem_arrive.as_ptr());
        libc::sem_unlink(sem_gate.as_ptr());
    }

I'd prefer all filesystem ops to be handled at crate-level rather a mix of crate + Makfeile.

---

tfhe-benchmark/src/utilities.rs line 500 at r1 (raw file):

    const MUTEX_NAME_PREFIX: &str = "tfhe_bench";

    let sem_mutex = CString::new(format!("/{MUTEX_NAME_PREFIX}_mutex")).unwrap();

What does sem_ stands for here? semaphore_?

---

tfhe-benchmark/src/utilities.rs line 504 at r1 (raw file):

    let sem_gate = CString::new(format!("/{MUTEX_NAME_PREFIX}_gate")).unwrap();

    let now = SystemTime::now().duration_since(UNIX_EPOCH).unwrap();

Do we want to panic if this call fails, despite being very unlikely, since it would abort the whole benchmark?

[github-actions]

✅ Backward-compat snapshot: everything looks good! No backward-compatibility issues detected.