instrumented_case/analysis/window_telemetry.py

#!/usr/bin/env python3
"""
Per-window telemetry aggregation for the instrumented multi-node case.

Reduces the raw collector NDJSON of one labelled run to the per-window,
per-node and participant-averaged aggregates. Standard library only.

    python3 window_telemetry.py --run-dir <run> --out-dir <derived>

<run> holds one directory per node, each carrying what `collect.py` and the
orchestrator wrote for that node:

    <run>/<node>/manifest.json
    <run>/<node>/windows_rank<R>.ndjson      window bounds and return codes
    <run>/<node>/gpu.<UTCDATE>.ndjson        DCGM samples
    <run>/<node>/fabric.<UTCDATE>.ndjson     interface counters
    <run>/<node>/disk.<UTCDATE>.ndjson       block-device counters
    <run>/<node>/ground_truth/*.json         per-window workload ground truth

Conventions:

  * The rank of a node is taken from the name of its windows file. A node's
    samples are attributed to a window using that node's own window bounds,
    never another node's.
  * The participant set of a window is the ranks whose window record carries a
    non-empty `return_codes` list. Non-participants idle through the window and
    are excluded from the participant average.
  * Activity and power are arithmetic means of the DCGM `sm_active`,
    `tensor_active` and `power_w` fields over the samples inside the window.
    Missing fields are skipped, never read as zero, and the number of samples
    behind every mean is reported alongside it.
  * Fabric volume is the sum of the non-negative successive differences of
    `tx_bytes + rx_bytes` on one named Ethernet interface (default `bond0`),
    over the samples inside the window. The rate divides that volume by the
    span between the first and last sample used, not by the nominal window
    duration, so a collector gap lowers the volume instead of raising the
    rate.
  * Disk write volume is computed per block device from `sectors_written`
    (512 bytes per sector) by the same successive-difference rule, and the
    window keeps the single busiest device rather than summing devices,
    because partitions and device-mapper targets double-count the same writes.
  * Counters are monotonic, so a negative successive difference means a
    counter reset or a device reappearing; such a step contributes zero rather
    than a negative volume, and the number of resets seen is reported.
"""

from __future__ import annotations

import argparse
import csv
import json
import re
import sys
from pathlib import Path
from typing import Any, Iterable

BYTES_PER_SECTOR = 512
MB = 1.0e6

GPU_MEAN_FIELDS = (
    "sm_active",
    "tensor_active",
    "gr_engine_active",
    "dram_active",
    "power_w",
    "gpu_util",
    "fb_used_mib",
    "temp_c",
)

WINDOWS_RE = re.compile(r"^windows_rank(\d+)\.ndjson$")


# --------------------------------------------------------------------- input


def read_ndjson(path: Path) -> Iterable[dict[str, Any]]:
    with path.open(encoding="utf-8") as handle:
        for line in handle:
            line = line.strip()
            if not line:
                continue
            try:
                yield json.loads(line)
            except json.JSONDecodeError:
                continue


def discover_nodes(run_dir: Path) -> list[dict[str, Any]]:
    """One entry per node directory, with its rank and its window records."""

    nodes = []
    for node_dir in sorted(p for p in run_dir.iterdir() if p.is_dir()):
        windows_files = [p for p in node_dir.iterdir() if WINDOWS_RE.match(p.name)]
        if not windows_files:
            continue
        windows_path = sorted(windows_files)[0]
        rank = int(WINDOWS_RE.match(windows_path.name).group(1))
        windows = []
        for record in read_ndjson(windows_path):
            windows.append(
                {
                    "id": record["id"],
                    "t_start": float(record["t_start"]),
                    "t_end": float(record["t_end"]),
                    "duration_s": float(record.get("duration_s", 0.0)),
                    "participant": bool(record.get("return_codes")),
                    "return_codes": record.get("return_codes", []),
                    "error": record.get("error"),
                }
            )
        manifest = {}
        manifest_path = node_dir / "manifest.json"
        if manifest_path.exists():
            manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
        nodes.append(
            {
                "node": node_dir.name,
                "rank": rank,
                "dir": node_dir,
                "windows": windows,
                "hardware": manifest.get("driver_and_gpus"),
                "sampling": manifest.get("sampling"),
            }
        )
    return sorted(nodes, key=lambda n: n["rank"])


def window_index(windows: list[dict[str, Any]]):
    """Return a lookup from a timestamp in seconds to a window position."""

    bounds = [(w["t_start"], w["t_end"], i) for i, w in enumerate(windows)]
    bounds.sort()

    def locate(ts: float) -> int | None:
        for start, end, position in bounds:
            if start <= ts <= end:
                return position
            if ts < start:
                return None
        return None

    return locate


# ------------------------------------------------------------ per-node passes


def gpu_pass(node: dict[str, Any], locate) -> list[dict[str, Any]]:
    """Means of the DCGM fields over the samples inside each window."""

    acc = [
        {"sums": {f: 0.0 for f in GPU_MEAN_FIELDS},
         "counts": {f: 0 for f in GPU_MEAN_FIELDS},
         "samples": 0,
         "fallback_samples": 0,
         "first_ts": None,
         "last_ts": None}
        for _ in node["windows"]
    ]
    for path in sorted(node["dir"].glob("gpu.*.ndjson")):
        for record in read_ndjson(path):
            ts = record.get("ts")
            if ts is None:
                continue
            position = locate(ts / 1000.0)
            if position is None:
                continue
            bucket = acc[position]
            bucket["samples"] += 1
            if record.get("src") != "dcgm":
                bucket["fallback_samples"] += 1
            if bucket["first_ts"] is None:
                bucket["first_ts"] = ts / 1000.0
            bucket["last_ts"] = ts / 1000.0
            for field in GPU_MEAN_FIELDS:
                value = record.get(field)
                if isinstance(value, (int, float)):
                    bucket["sums"][field] += float(value)
                    bucket["counts"][field] += 1
    return acc


def counter_pass(
    node: dict[str, Any],
    locate,
    glob_pattern: str,
    key_of,
    value_of,
) -> list[dict[str, dict[str, float]]]:
    """Successive-difference volumes per counter key, per window.

    `key_of` names the series a record belongs to (an interface or a block
    device) and `value_of` extracts its monotonic counter value.
    """

    acc: list[dict[str, dict[str, float]]] = [dict() for _ in node["windows"]]
    last: dict[str, tuple[float, float]] = {}
    for path in sorted(node["dir"].glob(glob_pattern)):
        for record in read_ndjson(path):
            ts = record.get("ts")
            if ts is None:
                continue
            key = key_of(record)
            if key is None:
                continue
            value = value_of(record)
            if value is None:
                continue
            ts = ts / 1000.0
            previous = last.get(key)
            last[key] = (ts, value)
            if previous is None:
                continue
            position = locate(ts)
            if position is None:
                continue
            # Only pair up samples that both fall inside the same window.
            if not (node["windows"][position]["t_start"] <= previous[0]
                    <= node["windows"][position]["t_end"]):
                continue
            series = acc[position].setdefault(
                key, {"volume": 0.0, "resets": 0, "samples": 0,
                      "first_ts": previous[0], "last_ts": ts},
            )
            delta = value - previous[1]
            if delta < 0:
                series["resets"] += 1
            else:
                series["volume"] += delta
            series["samples"] += 1
            series["last_ts"] = ts
    return acc


def fabric_pass(node, locate, interface: str):
    return counter_pass(
        node,
        locate,
        "fabric.*.ndjson",
        key_of=lambda r: r.get("dev") if r.get("class") == "eth" else None,
        value_of=lambda r: (
            float(r["tx_bytes"]) + float(r["rx_bytes"])
            if isinstance(r.get("tx_bytes"), (int, float))
            and isinstance(r.get("rx_bytes"), (int, float))
            else None
        ),
    )


def disk_pass(node, locate):
    return counter_pass(
        node,
        locate,
        "disk.*.ndjson",
        key_of=lambda r: r.get("dev"),
        value_of=lambda r: (
            float(r["sectors_written"]) * BYTES_PER_SECTOR
            if isinstance(r.get("sectors_written"), (int, float))
            else None
        ),
    )


def clock_pass(node, locate) -> list[dict[str, Any]]:
    """Clock-channel delivery, kept as a three-state coverage record."""

    acc = [
        {"samples": 0, "offset_samples": 0, "sources": set()}
        for _ in node["windows"]
    ]
    for path in sorted(node["dir"].glob("clock.*.ndjson")):
        for record in read_ndjson(path):
            ts = record.get("ts")
            if ts is None:
                continue
            position = locate(ts / 1000.0)
            if position is None:
                continue
            acc[position]["samples"] += 1
            acc[position]["sources"].add(record.get("src"))
            if isinstance(record.get("system_time_offset_s"), (int, float)):
                acc[position]["offset_samples"] += 1
    return [
        {
            "samples": a["samples"],
            "offset_samples": a["offset_samples"],
            "sources": sorted(s for s in a["sources"] if s),
            # No offset value at all is a delivery gap, not a zero offset.
            "offset_state": "observed" if a["offset_samples"] else
                            ("synchronization_flag_only" if a["samples"] else "missing"),
        }
        for a in acc
    ]


# --------------------------------------------------------------- aggregation


def rate(volume: float, first_ts: float | None, last_ts: float | None) -> float | None:
    if first_ts is None or last_ts is None:
        return None
    span = last_ts - first_ts
    if span <= 0:
        return None
    return volume / span


def mean(total: float, count: int) -> float | None:
    return total / count if count else None


def per_node_rows(node, gpu, fabric, disk, clock, interface: str):
    rows = []
    for position, window in enumerate(node["windows"]):
        gpu_acc = gpu[position]
        fabric_series = fabric[position].get(interface, {})
        disk_series = disk[position]
        busiest = max(
            disk_series.items(), key=lambda item: item[1]["volume"], default=None
        )
        row = {
            "window": window["id"],
            "node": node["node"],
            "rank": node["rank"],
            "participant": window["participant"],
            "return_codes": window["return_codes"],
            "error": window["error"],
            "t_start": window["t_start"],
            "t_end": window["t_end"],
            "duration_s": window["duration_s"],
            "gpu_samples": gpu_acc["samples"],
            "gpu_fallback_samples": gpu_acc["fallback_samples"],
            "gpu_sample_hz": rate(
                max(gpu_acc["samples"] - 1, 0), gpu_acc["first_ts"], gpu_acc["last_ts"]
            ),
            "fabric_interface": interface,
            "fabric_bytes": fabric_series.get("volume"),
            "fabric_samples": fabric_series.get("samples"),
            "fabric_counter_resets": fabric_series.get("resets"),
            "fabric_mb_per_s": None,
            "disk_busiest_device": busiest[0] if busiest else None,
            "disk_write_bytes": busiest[1]["volume"] if busiest else None,
            "disk_write_mb_per_s": None,
            "clock": clock[position],
        }
        if fabric_series:
            row["fabric_mb_per_s"] = (
                rate(fabric_series["volume"], fabric_series["first_ts"],
                     fabric_series["last_ts"]) or 0.0
            ) / MB
        if busiest:
            row["disk_write_mb_per_s"] = (
                rate(busiest[1]["volume"], busiest[1]["first_ts"],
                     busiest[1]["last_ts"]) or 0.0
            ) / MB
        for field in GPU_MEAN_FIELDS:
            row[field] = mean(gpu_acc["sums"][field], gpu_acc["counts"][field])
            row[f"{field}_samples"] = gpu_acc["counts"][field]
        rows.append(row)
    return rows


AVERAGED_FIELDS = (
    "sm_active",
    "tensor_active",
    "gr_engine_active",
    "dram_active",
    "power_w",
    "fabric_mb_per_s",
    "disk_write_mb_per_s",
)


def participant_average(rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
    """Unweighted mean over the participant nodes of each window."""

    order: list[str] = []
    grouped: dict[str, list[dict[str, Any]]] = {}
    for row in rows:
        if row["window"] not in grouped:
            grouped[row["window"]] = []
            order.append(row["window"])
        grouped[row["window"]].append(row)

    out = []
    for window in order:
        members = grouped[window]
        participants = [r for r in members if r["participant"]] or members
        record = {
            "window": window,
            "participant_ranks": sorted(r["rank"] for r in members if r["participant"]),
            "participant_nodes": sorted(r["node"] for r in members if r["participant"]),
            "nodes_reporting": len(members),
            "duration_s": members[0]["duration_s"],
            "return_codes": sorted(
                {code for r in members for code in (r["return_codes"] or [])}
            ),
        }
        for field in AVERAGED_FIELDS:
            values = [r[field] for r in participants if r[field] is not None]
            record[field] = sum(values) / len(values) if values else None
        out.append(record)
    return out


# ------------------------------------------------------------------ ground truth


def load_ground_truth(run_dir: Path) -> dict[str, list[dict[str, Any]]]:
    truth: dict[str, list[dict[str, Any]]] = {}
    for path in sorted(run_dir.glob("*/ground_truth/*.json")):
        record = json.loads(path.read_text(encoding="utf-8"))
        truth.setdefault(record.get("run_id", path.stem), []).append(record)
    return truth


# ------------------------------------------------------------------------ main


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__,
                                     formatter_class=argparse.RawDescriptionHelpFormatter)
    parser.add_argument("--run-dir", type=Path, required=True,
                        help="directory holding one subdirectory per node")
    parser.add_argument("--out-dir", type=Path, required=True,
                        help="where window_telemetry.json and the CSVs are written")
    parser.add_argument("--interface", default="bond0",
                        help="Ethernet interface used as the fabric counter (default bond0)")
    args = parser.parse_args()

    nodes = discover_nodes(args.run_dir)
    if not nodes:
        raise SystemExit(f"no node directories with a windows_rank*.ndjson under {args.run_dir}")

    rows: list[dict[str, Any]] = []
    for node in nodes:
        locate = window_index(node["windows"])
        print(f"  node {node['node']} (rank {node['rank']}): "
              f"{len(node['windows'])} windows", file=sys.stderr)
        gpu = gpu_pass(node, locate)
        fabric = fabric_pass(node, locate, args.interface)
        disk = disk_pass(node, locate)
        clock = clock_pass(node, locate)
        rows.extend(per_node_rows(node, gpu, fabric, disk, clock, args.interface))

    averaged = participant_average(rows)
    truth = load_ground_truth(args.run_dir)

    args.out_dir.mkdir(parents=True, exist_ok=True)
    payload = {
        "schema": "trace-window-telemetry/1",
        "conventions": {
            "participant_set": "ranks whose window record has a non-empty return_codes list",
            "activity_and_power": "arithmetic mean of the DCGM field over samples inside the node's own window bounds",
            "fabric": f"sum of non-negative successive differences of tx_bytes+rx_bytes on {args.interface}, divided by the span of the samples used",
            "disk": "single busiest block device by sectors_written x 512, never a sum over devices",
            "participant_average": "unweighted mean across participant nodes",
            "missing": "a field absent from a sample is skipped, never read as zero",
        },
        "nodes": [
            {"node": n["node"], "rank": n["rank"], "hardware": n["hardware"],
             "sampling": n["sampling"]}
            for n in nodes
        ],
        "windows_participant_averaged": averaged,
        "windows_per_node": rows,
        "ground_truth_runs": sorted(truth),
    }
    (args.out_dir / "window_telemetry.json").write_text(
        json.dumps(payload, indent=2, sort_keys=False) + "\n", encoding="utf-8"
    )

    with (args.out_dir / "window_table.csv").open("w", newline="", encoding="utf-8") as fh:
        writer = csv.writer(fh)
        writer.writerow(["window", "participant_ranks", "sm_active", "tensor_active",
                         "fabric_mb_per_s", "disk_write_mb_per_s", "power_w"])
        for record in averaged:
            writer.writerow([
                record["window"],
                " ".join(str(r) for r in record["participant_ranks"]),
                _fmt(record["sm_active"], 2),
                _fmt(record["tensor_active"], 2),
                _fmt(record["fabric_mb_per_s"], 0),
                _fmt(record["disk_write_mb_per_s"], 1),
                _fmt(record["power_w"], 0),
            ])

    with (args.out_dir / "window_table_per_node.csv").open("w", newline="", encoding="utf-8") as fh:
        writer = csv.writer(fh)
        writer.writerow(["window", "node", "rank", "participant", "sm_active",
                         "tensor_active", "power_w", "fabric_mb_per_s",
                         "disk_busiest_device", "disk_write_mb_per_s",
                         "gpu_samples", "gpu_sample_hz", "clock_offset_state"])
        for row in rows:
            writer.writerow([
                row["window"], row["node"], row["rank"], int(row["participant"]),
                _fmt(row["sm_active"], 4), _fmt(row["tensor_active"], 4),
                _fmt(row["power_w"], 2), _fmt(row["fabric_mb_per_s"], 2),
                row["disk_busiest_device"], _fmt(row["disk_write_mb_per_s"], 3),
                row["gpu_samples"], _fmt(row["gpu_sample_hz"], 3),
                row["clock"]["offset_state"],
            ])

    print(f"wrote {args.out_dir/'window_telemetry.json'}", file=sys.stderr)


def _fmt(value: float | None, digits: int) -> str:
    if value is None:
        return ""
    return f"{value:.{digits}f}"


if __name__ == "__main__":
    main()