scripts/plot_coordination.py

#!/usr/bin/env python3
"""Regenerate the C1 coordination figures/tables from raw experiment CSVs.

Consumes the output of ../grite/scripts/run_experiments.sh (coordination.csv) and writes
PDFs into ../figures plus a LaTeX-ready summary table. Pure stdlib + matplotlib; no seaborn.

Usage:
  python plot_coordination.py --raw _raw --out ../figures
"""
import argparse
import csv
import statistics
from collections import defaultdict
from pathlib import Path

ARMS = ["no-coord", "locks-only", "locks-plus-state"]
ARM_LABEL = {
    "no-coord": "No coordination",
    "locks-only": "Locks only",
    "locks-plus-state": "Locks + shared state",
}


def load(raw_dir: Path):
    rows = []
    with open(raw_dir / "coordination.csv", newline="") as f:
        for r in csv.DictReader(f):
            rows.append(r)
    return rows


def aggregate(rows):
    """-> agg[arm][n] = {metric: (mean, ci95)} over seeds."""
    buckets = defaultdict(lambda: defaultdict(lambda: defaultdict(list)))
    for r in rows:
        arm, n = r["arm"], int(r["n_agents"])
        buckets[arm][n]["dup"].append(float(r["duplicate_work_rate"]))
        buckets[arm][n]["conf"].append(float(r["conflicting_edits"]))
        buckets[arm][n]["good"].append(float(r["goodput"]))
        buckets[arm][n]["deny"].append(float(r["lock_denials"]))

    def stat(xs):
        m = statistics.fmean(xs)
        if len(xs) > 1:
            sd = statistics.pstdev(xs)
            ci = 1.96 * sd / (len(xs) ** 0.5)
        else:
            ci = 0.0
        return m, ci

    agg = defaultdict(lambda: defaultdict(dict))
    for arm, ns in buckets.items():
        for n, metrics in ns.items():
            for k, xs in metrics.items():
                agg[arm][n][k] = stat(xs)
    return agg


def plot(agg, out_dir: Path):
    import matplotlib

    matplotlib.use("Agg")
    import matplotlib.pyplot as plt

    out_dir.mkdir(parents=True, exist_ok=True)
    ns_all = sorted({n for arm in agg for n in agg[arm]})

    # Figure 1: duplicate-work rate vs N, per arm.
    fig, ax = plt.subplots(figsize=(4.2, 3.0))
    for arm in ARMS:
        if arm not in agg:
            continue
        ns = sorted(agg[arm])
        ys = [agg[arm][n]["dup"][0] for n in ns]
        es = [agg[arm][n]["dup"][1] for n in ns]
        ax.errorbar(ns, ys, yerr=es, marker="o", capsize=3, label=ARM_LABEL[arm])
    ax.set_xscale("log", base=2)
    ax.set_xticks(ns_all)
    ax.set_xticklabels(ns_all)
    ax.set_xlabel("Concurrent agents $N$")
    ax.set_ylabel("Duplicate-work rate")
    ax.set_ylim(-0.02, 1.0)
    ax.legend(fontsize=7)
    ax.grid(True, alpha=0.3)
    fig.tight_layout()
    fig.savefig(out_dir / "duplicate_work.pdf")
    plt.close(fig)

    # Figure 2 (money figure): coordination overhead (lock denials, proxy) vs duplicate-work
    # AVOIDED relative to no-coord. One point per (arm, N): the Pareto trade-off.
    fig, ax = plt.subplots(figsize=(4.2, 3.0))
    for arm in ARMS:
        if arm not in agg or arm == "no-coord":
            continue
        xs, ys = [], []
        for n in sorted(agg[arm]):
            base = agg["no-coord"][n]["dup"][0]
            avoided = base - agg[arm][n]["dup"][0]
            overhead = agg[arm][n]["deny"][0]  # lock-wait/denial events as overhead proxy
            xs.append(overhead)
            ys.append(avoided)
        ax.plot(xs, ys, marker="s", label=ARM_LABEL[arm])
    ax.set_xlabel("Coordination overhead (lock denials)")
    ax.set_ylabel("Duplicate work avoided vs. no-coord")
    ax.legend(fontsize=7)
    ax.grid(True, alpha=0.3)
    fig.tight_layout()
    fig.savefig(out_dir / "pareto.pdf")
    plt.close(fig)

    print(f"[plot_coordination] wrote {out_dir}/duplicate_work.pdf, {out_dir}/pareto.pdf")


def write_table(agg, out_dir: Path):
    """Emit a LaTeX booktabs summary at the largest N."""
    n = max({n for arm in agg for n in agg[arm]})
    lines = [
        "% auto-generated by plot_coordination.py -- do not edit",
        "\\begin{tabular}{lrrr}",
        "\\toprule",
        f"Arm ($N={n}$) & Dup-work rate & Conflicting edits & Goodput \\\\",
        "\\midrule",
    ]
    for arm in ARMS:
        if arm not in agg or n not in agg[arm]:
            continue
        d = agg[arm][n]
        lines.append(
            f"{ARM_LABEL[arm]} & {d['dup'][0]:.2f} & {d['conf'][0]:.0f} & {d['good'][0]:.2f} \\\\"
        )
    lines += ["\\bottomrule", "\\end{tabular}"]
    (out_dir / "coordination_table.tex").write_text("\n".join(lines) + "\n")
    print(f"[plot_coordination] wrote {out_dir}/coordination_table.tex")


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--raw", default="_raw", type=Path)
    ap.add_argument("--out", default=Path("../figures"), type=Path)
    args = ap.parse_args()
    rows = load(args.raw)
    agg = aggregate(rows)
    plot(agg, args.out)
    write_table(agg, args.out)


if __name__ == "__main__":
    main()