benchmark_sage – SAGE System-Level Benchmarks and ICML Artifacts

benchmark_sage is a home for system-level benchmarks and artifacts that focus on SAGE as a complete ML systems platform.

Key points:

• SAGE is more than an LLM control plane. The LLM/embedding control plane is one subsystem. SAGE also includes components such as sage.db, sage.flow, sage.tsdb, and others, all orchestrated via a common declarative dataflow model.
• packages/sage-benchmark already contains multiple benchmark suites (agents, control-plane scheduling, DB, retrieval, memory, schedulers, refiner, libamm, etc.). benchmark_sage can aggregate cross-cutting experiments that involve several SAGE subsystems together.
• This folder may also store ICML writing prompts and experiment templates for the SAGE system track papers, under docs/.

Suggested uses:

• End-to-end experiments that span sage.flow pipelines, sage.db storage, sage.tsdb time-series monitoring, and the LLM/embedding control plane.
• Configs (config/*.yaml) for system-track experiments described in an ICML paper.
• Notebook or script entry points that reproduce figures/tables.

Q-style Workload Catalog (TPC-H/TPC-C inspired)

benchmark_sage adopts a fixed Q1..Q8 catalog where each Q denotes a workload family rather than a one-off script. This keeps paper claims, configs, and run outputs aligned.

Examples:

# Run a single workload against the default SAGE backend
python -m sage.benchmark.benchmark_sage --experiment Q1

# Run all workloads
python -m sage.benchmark.benchmark_sage --all

# Quick smoke-test
python -m sage.benchmark.benchmark_sage --experiment Q3 --quick
python -m sage.benchmark.benchmark_sage --experiment Q7 --quick

# Backend comparison: same workload, two backends, for fair comparison
python -m sage.benchmark.benchmark_sage --experiment Q1 --backend sage --repeat 3 --seed 42
python -m sage.benchmark.benchmark_sage --experiment Q1 --backend ray  --repeat 3 --seed 42

# Distributed run: 4 nodes, 8-way operator parallelism
python -m sage.benchmark.benchmark_sage --experiment Q4 \
    --backend sage --nodes 4 --parallelism 8 --output-dir results/q4_scale

# Validate config without running
python -m sage.benchmark.benchmark_sage --experiment Q2 --dry-run

Paired backend automation (Issue #7)

Use one command to launch paired sage + ray runs, archive artifacts with run_id and config_hash, and generate unified comparison outputs:

python experiments/analysis/run_paired_backends.py \
  --scheduler fifo \
  --items 10 \
  --parallelism 2 \
  --nodes 1 \
  --seed 42

Artifacts are written under:

artifacts/paired_backend_runs/run_id=<...>/config_hash=<...>/

With this structure:

• backends/sage/ and backends/ray/: raw unified metrics (unified_results.jsonl/csv)
• comparison/: summary report and merged comparison CSV
• logs/: per-step actionable logs (sage.log, ray.log, compare.log)
• manifest.json: run metadata for reproducibility

Manual GitHub Actions trigger is available in:

• .github/workflows/paired-backend-run.yml

Direct comparison report (Issue #6)

If you already have mixed backend outputs, generate a unified report directly:

python experiments/analysis/compare_backends.py \
  /path/to/sage/results \
  /path/to/ray/results \
  --output-dir artifacts/backend_comparison

See detailed usage in docs/compare_backends.md.

Installation Profiles

# Default usage (no Ray dependency)
python -m pip install -e .

# Optional: enable Ray baseline backend
python -m pip install -e .[ray-baseline]

• Default installs do not require Ray and are unaffected.
• Use --backend ray only after installing the ray-baseline extra.

Standardised CLI flags (Issue #2)

All workload entry points share the same flag contract so backend comparison runs always produce comparable run_config records.

Individual workloads may add extra flags on top of the shared contract.

Reproducibility Checklist (Issue #4)

For fair cross-backend comparisons (SAGE vs Ray), keep these controls fixed:

• Use the same --seed for both backends.
• Use the same warmup policy (--warmup-items) before timed runs.
• Use deterministic input parity split (deterministic_shuffle_v1) and fixed batch size.
• Persist and compare config_hash in output artifacts.
• Keep workload shape (--nodes, --parallelism, --repeat) identical.

Workload4 now writes repro_manifest.json with seed, parity batches, warmup split, and configuration fingerprint.

Example:

python experiments/distributed_workloads/run_workload4.py \
  --backend sage --seed 42 --warmup-items 5 --parity-batch-size 16

python experiments/distributed_workloads/run_workload4.py \
  --backend ray --seed 42 --warmup-items 5 --parity-batch-size 16

At the repo root, docs/icml-prompts/ contains reusable writing prompts. You can either reference them directly or copy customized versions into this folder when preparing a specific ICML submission.

Documentation