File size: 34,320 Bytes
96bb363 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 | """Plotting functions and data-loading helpers for openg2g simulation results.
Reproduces the figures from the G2G paper, reading data from the library's
``SimulationLog`` and ``LLMDatacenterState`` objects.
This module lives outside the ``openg2g`` library on purpose: the library
exports simulation state and metrics, while all matplotlib-dependent
visualization code stays here.
"""
from __future__ import annotations
import math
import re
from collections.abc import Sequence
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any, cast
import matplotlib.figure
import matplotlib.pyplot as plt
import numpy as np
from matplotlib.axes import Axes
from matplotlib.patches import Patch
from openg2g.datacenter.base import LLMDatacenterState
from openg2g.datacenter.offline import OfflineDatacenterState
from openg2g.datacenter.online import OnlineDatacenterState
from openg2g.grid.base import GridState
Figure = matplotlib.figure.Figure
# ── Bus color map (IEEE 13-bus, tab20-based) ─────────────────────────
# Deterministic colors so all voltage plots use consistent bus coloring.
BUS_COLOR_MAP: dict[str, Any] = {
"611": (0.1216, 0.4667, 0.7059, 1.0),
"632": (0.6824, 0.7804, 0.9098, 1.0),
"633": (1.0, 0.4980, 0.0549, 1.0),
"634": (1.0, 0.7333, 0.4706, 1.0),
"645": (0.1725, 0.6275, 0.1725, 1.0),
"646": (0.5961, 0.8745, 0.5412, 1.0),
"650": (0.8392, 0.1529, 0.1569, 1.0),
"652": (1.0, 0.5961, 0.5882, 1.0),
"670": (0.5804, 0.4039, 0.7412, 1.0),
"671": (0.7725, 0.6902, 0.8353, 1.0),
"675": (0.5490, 0.3373, 0.2941, 1.0),
"680": (0.7686, 0.6118, 0.5804, 1.0),
"684": (0.8902, 0.4667, 0.7608, 1.0),
"692": (0.9686, 0.7137, 0.8235, 1.0),
"rg60": "black",
"sourcebus": (0.4980, 0.4980, 0.4980, 1.0),
}
# ── Per-model time-series extraction ─────────────────────────────────
@dataclass
class PerModelTimeSeries:
"""Per-model time series extracted from ``SimulationLog.dc_states``."""
time_s: np.ndarray
power_w: dict[str, np.ndarray] = field(default_factory=dict)
itl_s: dict[str, np.ndarray] = field(default_factory=dict)
batch_size: dict[str, np.ndarray] = field(default_factory=dict)
active_replicas: dict[str, np.ndarray] = field(default_factory=dict)
def extract_per_model_timeseries(
dc_states: Sequence[LLMDatacenterState],
) -> PerModelTimeSeries:
"""Build per-model arrays from a list of ``LLMDatacenterState`` objects.
``power_w`` is populated only when states are ``OfflineDatacenterState``
(which carries ``power_by_model_w``).
"""
if not dc_states:
raise ValueError("dc_states is empty.")
time_s = np.array([s.time_s for s in dc_states])
model_labels = sorted(dc_states[0].batch_size_by_model.keys())
batch_size: dict[str, np.ndarray] = {}
active_replicas: dict[str, np.ndarray] = {}
itl_s: dict[str, np.ndarray] = {}
power_w: dict[str, np.ndarray] = {}
for label in model_labels:
batch_size[label] = np.array([s.batch_size_by_model.get(label, 0) for s in dc_states])
active_replicas[label] = np.array([s.active_replicas_by_model.get(label, 0) for s in dc_states])
itl_s[label] = np.array([s.observed_itl_s_by_model.get(label, float("nan")) for s in dc_states])
if dc_states and isinstance(dc_states[0], OfflineDatacenterState):
offline_states = cast(list[OfflineDatacenterState], dc_states)
for label in model_labels:
power_w[label] = np.array([s.power_by_model_w.get(label, 0.0) for s in offline_states])
return PerModelTimeSeries(
time_s=time_s,
power_w=power_w,
itl_s=itl_s,
batch_size=batch_size,
active_replicas=active_replicas,
)
# ── Helper: bus ordering ─────────────────────────────────────────────
def _bus_sort_key(b: str) -> tuple[int, str]:
"""Sort buses numerically when possible: '650' < '671' < '692'."""
m = re.match(r"^(\d+)", str(b).strip())
if m:
return (int(m.group(1)), str(b))
return (999999, str(b))
# ══════════════════════════════════════════════════════════════════════
# Paper Fig. 5: 2-panel, 3-phase power (MW) + per-model average ITL
# ══════════════════════════════════════════════════════════════════════
def plot_power_and_itl_2panel(
time_s: np.ndarray,
kW_A: np.ndarray,
kW_B: np.ndarray,
kW_C: np.ndarray,
avg_itl_by_model: dict[str, np.ndarray],
*,
itl_time_s: np.ndarray | None = None,
t_train_start_s: float = 1000.0,
t_train_end_s: float = 2000.0,
t_ramp_start_s: float = 2500.0,
t_ramp_end_s: float = 3600.0,
ramp_label: str = "Less active GPUs",
figsize: tuple[float, float] = (7.2, 3.6),
dpi: int = 300,
save_path: Path | str | None = None,
) -> Figure:
"""2-panel figure: (a) 3-phase power in MW, (b) per-model average ITL.
Args:
time_s: Time array (seconds) aligned with kW_A/B/C.
kW_A, kW_B, kW_C: Per-phase power in kW.
avg_itl_by_model: {model_label: itl_array_s}. Time base is
*itl_time_s* if provided, otherwise *time_s*.
itl_time_s: Separate time base for ITL data (e.g. DC step rate).
save_path: If given, save figure to this path.
Returns:
The matplotlib Figure object.
"""
t_min = np.asarray(time_s) / 60.0
t_itl_min = (np.asarray(itl_time_s) / 60.0) if itl_time_s is not None else t_min
t_train_start_min = t_train_start_s / 60.0
t_train_end_min = t_train_end_s / 60.0
t_ramp_start_min = t_ramp_start_s / 60.0
t_ramp_end_min = t_ramp_end_s / 60.0
fig, axes = plt.subplots(
2,
1,
sharex=True,
figsize=figsize,
dpi=dpi,
constrained_layout=True,
)
title_pad = 4
label_fs = 10
tick_fs = 10
legend_fs = 9
lw_main = 0.8
lw_lat = 0.6
phase_colors = {
"A": "#4E79A7",
"B": "#59A14F",
"C": "#9C755F",
}
# (a) Three-phase power
ax = axes[0]
ax.plot(t_min, np.asarray(kW_A) / 1e3, lw=lw_main, color=phase_colors["A"], label="Phase A")
ax.plot(t_min, np.asarray(kW_B) / 1e3, lw=lw_main, color=phase_colors["B"], label="Phase B")
ax.plot(t_min, np.asarray(kW_C) / 1e3, lw=lw_main, color=phase_colors["C"], label="Phase C")
ax.axvspan(t_train_start_min, t_train_end_min, color="tab:blue", alpha=0.10, zorder=0)
ax.axvline(t_train_start_min, ls="--", lw=1.0, color="tab:blue", alpha=0.9)
ax.axvline(t_train_end_min, ls="--", lw=1.0, color="tab:blue", alpha=0.9)
ax.set_ylabel("Power (MW)", fontsize=label_fs)
ax.set_title("(a) Synthetic three-phase data center power demand", pad=title_pad)
ax.grid(True, alpha=0.25)
ax.tick_params(labelsize=tick_fs, labelbottom=False, bottom=True)
handles, labels = ax.get_legend_handles_labels()
handles.append(Patch(facecolor="tab:blue", alpha=0.10, edgecolor="none"))
labels.append("Training window")
ax.legend(handles, labels, fontsize=legend_fs, ncol=4, loc="best", framealpha=0.9)
# (b) Per-model average ITL
ax = axes[1]
for model, lat in avg_itl_by_model.items():
ax.plot(t_itl_min, np.asarray(lat), lw=lw_lat, label=model)
ax.axvspan(
t_ramp_start_min,
t_ramp_end_min,
color="tab:orange",
alpha=0.12,
zorder=0,
label=ramp_label,
)
ax.axvline(t_ramp_start_min, ls="--", lw=1.0, color="tab:orange", alpha=0.9)
ax.axvline(t_ramp_end_min, ls="--", lw=1.0, color="tab:orange", alpha=0.9)
ax.set_ylim(0.0, 0.11)
ax.set_ylabel("Avg ITL (s)", fontsize=label_fs)
ax.set_title("(b) Per-model average ITL", pad=title_pad)
ax.grid(True, alpha=0.25)
ax.legend(fontsize=legend_fs, ncol=1, loc="upper left", framealpha=0.9)
ax.set_xlabel("Time (minutes)", fontsize=label_fs)
ax.tick_params(labelsize=tick_fs)
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
# ══════════════════════════════════════════════════════════════════════
# Paper Fig. 6 / Fig. 7: Per-phase all-bus voltages with bus colormap
# ══════════════════════════════════════════════════════════════════════
def plot_allbus_voltages_per_phase(
grid_states: list[GridState],
time_s: np.ndarray,
*,
save_dir: Path | str,
v_min: float = 0.95,
v_max: float = 1.05,
bus_color_map: dict[str, Any] | None = None,
figsize_main: tuple[float, float] = (7.2, 2.0),
figsize_c: tuple[float, float] = (7.2, 2.0),
dpi: int = 200,
legend_mode: str = "best",
drop_buses: tuple[str, ...] = ("sourcebus",),
reg_bus: str = "rg60",
reg_label: str = "Regulator bus",
reg_color: str = "black",
reg_lw: float = 2.4,
reg_zorder: int = 5,
title_template: str = "Voltage trajectories (Phase {label})",
shared_legend_phase: str = "B",
filename_template: str = "allbus_voltages_phase_{label}.png",
) -> None:
"""Per-phase all-bus voltage plots with bus-specific colors and shared legend.
Produces one PNG per phase (A, B, C). Legend is placed on the
``shared_legend_phase`` (default B).
Args:
grid_states: List of ``GridState`` from ``SimulationLog``.
time_s: Time array (seconds) aligned with *grid_states*.
save_dir: Directory to write PNG files into.
bus_color_map: {bus_name: color}. Defaults to ``BUS_COLOR_MAP``.
"""
if bus_color_map is None:
bus_color_map = BUS_COLOR_MAP
save_dir = Path(save_dir)
t_min = np.asarray(time_s) / 60.0
drop_set = {str(b).strip().lower() for b in drop_buses}
reg_bus_lc = str(reg_bus).strip().lower()
PHASE_MAP = {"A": 1, "B": 2, "C": 3}
shared_legend_phase_int = PHASE_MAP.get(shared_legend_phase.upper(), 2)
# Build buses-with-phase from grid_states
buses_with_phase: dict[str, set[str]] = {"A": set(), "B": set(), "C": set()}
for snap in grid_states[: min(10, len(grid_states))]:
for bus in snap.voltages.buses():
b_lc = bus.lower()
if b_lc in drop_set:
continue
v = snap.voltages[bus]
if not np.isnan(v.a):
buses_with_phase["A"].add(bus)
if not np.isnan(v.b):
buses_with_phase["B"].add(bus)
if not np.isnan(v.c):
buses_with_phase["C"].add(bus)
# Build V arrays per phase
V_by_phase: dict[str, tuple[list[str], np.ndarray]] = {}
for phase_letter in ("A", "B", "C"):
buses = sorted(buses_with_phase.get(phase_letter, set()), key=_bus_sort_key)
if not buses:
continue
V = np.full((len(buses), len(t_min)), np.nan, dtype=float)
for k, snap in enumerate(grid_states):
for i, b in enumerate(buses):
if b in snap.voltages:
v = snap.voltages[b]
V[i, k] = getattr(v, phase_letter.lower())
V_by_phase[phase_letter] = (buses, V)
y_lo = 0.93
y_hi = 1.11
AX_RECT = (0.10, 0.18, 0.88, 0.72)
shared_handles = None
shared_labels = None
for phase_letter in ("A", "B", "C"):
if phase_letter not in V_by_phase:
continue
buses, V = V_by_phase[phase_letter]
ph_int = PHASE_MAP[phase_letter]
figsize = figsize_c if phase_letter == "C" else figsize_main
fig = plt.figure(figsize=figsize, dpi=dpi)
ax = fig.add_axes(AX_RECT)
handles_by_label: dict[str, Any] = {}
# Sort buses: reg_bus last so it draws on top
ordered = [b for b in buses if b.lower() != reg_bus_lc]
ordered += [b for b in buses if b.lower() == reg_bus_lc]
for b in ordered:
b_lc = b.lower()
idx = buses.index(b)
if b_lc == reg_bus_lc:
color = reg_color
lw = reg_lw
zorder = reg_zorder
label = reg_label
else:
color = bus_color_map.get(b, bus_color_map.get(b_lc, "gray"))
lw = 1.2
zorder = 2
label = b
(line,) = ax.plot(
t_min,
V[idx],
label=label,
color=color,
linewidth=lw,
zorder=zorder,
)
if label not in handles_by_label:
handles_by_label[label] = line
ax.axhline(v_min, linestyle="--", linewidth=2.0, alpha=0.9)
ax.axhline(v_max, linestyle="--", linewidth=2.0, alpha=0.9)
ax.set_ylim(y_lo, y_hi)
ax.set_xlabel("Time (minutes)")
ax.set_ylabel("Voltage (pu)")
ax.set_title(title_template.format(label=phase_letter))
ax.grid(True, alpha=0.3)
if shared_handles is None and ph_int == shared_legend_phase_int:
ordered_labels: list[str] = []
for b in buses:
lab = reg_label if b.lower() == reg_bus_lc else b
if lab in handles_by_label and lab not in ordered_labels:
ordered_labels.append(lab)
shared_labels = ordered_labels
shared_handles = [handles_by_label[lab] for lab in shared_labels]
if ph_int == shared_legend_phase_int and shared_handles is not None and shared_labels is not None:
if legend_mode == "best":
ax.legend(
shared_handles,
shared_labels,
loc="upper center",
ncol=6,
fontsize=7,
frameon=True,
framealpha=0.95,
)
else:
fig.legend(
shared_handles,
shared_labels,
ncol=7,
fontsize=7,
frameon=True,
framealpha=0.95,
loc="lower center",
bbox_to_anchor=(0.5, 0.02),
)
fig.savefig(
save_dir / filename_template.format(label=phase_letter),
bbox_inches="tight",
)
plt.close(fig)
# ══════════════════════════════════════════════════════════════════════
# Paper Fig. 8: 4-panel, batch, power/replica, ITL, throughput (OFO)
# ══════════════════════════════════════════════════════════════════════
def plot_model_timeseries_4panel(
time_s: np.ndarray,
per_model: PerModelTimeSeries,
model_labels: list[str],
*,
regime_shading: bool = True,
t_regime_edges_s: tuple[float, ...] | None = None,
regime_colors: tuple[str, ...] | None = None,
regime_labels: tuple[str, ...] | None = None,
figsize: tuple[float, float] = (7.2, 6.2),
dpi: int = 300,
save_path: Path | str | None = None,
) -> Figure:
"""4-panel OFO time-series: batch, power/replica, ITL, throughput.
Args:
time_s: DSS/DC time base (seconds) aligned with *per_model* arrays.
per_model: Extracted per-model time series (batch sizes read from
`per_model.batch_size`).
model_labels: Ordered model labels for consistent color assignment.
regime_shading: If True, draw colored background spans.
t_regime_edges_s: Edge times (seconds) for regime spans.
Defaults to (0, 950, 2000, 2950, 3100, 3600).
regime_colors: Colors for each span (len = len(edges) - 1).
Defaults to ("tab:gray", "tab:red", "tab:gray", "tab:red", "tab:green").
regime_labels: Legend labels for unique regime colors.
Defaults to ("Throughput-driven", "Voltage-driven", "Latency-driven").
Returns:
The matplotlib Figure object.
"""
color_cycle = plt.rcParams["axes.prop_cycle"].by_key().get("color", [])
if not color_cycle:
cmap = plt.get_cmap("tab10")
color_cycle = [cmap(i) for i in range(10)]
model_to_color = {lab: color_cycle[i % len(color_cycle)] for i, lab in enumerate(model_labels)}
title_pad = 4
label_fs = 10
tick_fs = 10
legend_fs = 9
lw_main = 1.3
fig, axes = plt.subplots(
4,
1,
sharex=True,
figsize=figsize,
dpi=dpi,
constrained_layout=True,
)
# Regime shading setup
if t_regime_edges_s is None:
t_regime_edges_s = (0, 950, 2000, 2950, 3100, 3600)
if regime_colors is None:
regime_colors = ("tab:gray", "tab:red", "tab:gray", "tab:red", "tab:green")
if regime_labels is None:
regime_labels = ("Throughput-driven", "Voltage-driven", "Latency-driven")
t_edges_min = np.array(t_regime_edges_s, dtype=float) / 60.0
def _apply_overlays(ax: Axes) -> None:
if not regime_shading:
return
for i in range(len(t_edges_min) - 1):
ax.axvspan(
t_edges_min[i],
t_edges_min[i + 1],
color=regime_colors[i],
alpha=0.12,
zorder=0,
)
legend_patches: list[Patch] = []
if regime_shading:
seen_colors: set[str] = set()
label_idx = 0
for c in regime_colors:
if c not in seen_colors:
seen_colors.add(c)
lab = regime_labels[label_idx] if label_idx < len(regime_labels) else ""
label_idx += 1
legend_patches.append(
Patch(
facecolor=c,
edgecolor=c,
linewidth=1.2,
alpha=0.12,
label=lab,
)
)
# (a) log2(batch)
ax = axes[0]
t_dc_min = per_model.time_s / 60.0
for lab in model_labels:
b = per_model.batch_size.get(lab)
if b is None or b.size == 0:
continue
ax.step(
t_dc_min[: b.size],
np.log2(b),
where="post",
lw=lw_main,
color=model_to_color[lab],
label=lab,
)
_apply_overlays(ax)
ax.set_ylabel(r"$\log_2(\mathrm{batch})$", fontsize=label_fs)
ax.set_title("(a) Per-model batch size setting (log scale)", pad=title_pad)
ax.grid(True, alpha=0.25)
ax.tick_params(labelsize=tick_fs)
ymin, ymax = ax.get_ylim()
ax.set_yticks(np.arange(np.floor(ymin), np.ceil(ymax) + 1))
# (b) Power per active replica (kW)
ax = axes[1]
for lab in model_labels:
if lab not in per_model.power_w:
continue
p_tot = per_model.power_w[lab]
wrep = np.maximum(per_model.active_replicas[lab].astype(float), 1.0)
y = p_tot / wrep / 1e3 # W -> kW
ax.plot(per_model.time_s / 60.0, y, lw=1, color=model_to_color[lab], label=lab)
_apply_overlays(ax)
ax.set_ylabel("Power (kW)", fontsize=label_fs)
ax.set_title("(b) Per-replica power measurement", pad=title_pad)
ax.grid(True, alpha=0.25)
ax.tick_params(labelsize=tick_fs)
# (c) Per-model ITL (legend placed here)
ax = axes[2]
for lab in model_labels:
if lab not in per_model.itl_s:
continue
y = per_model.itl_s[lab]
ax.plot(per_model.time_s / 60.0, y, lw=0.8, color=model_to_color[lab], label=lab)
_apply_overlays(ax)
ax.set_ylabel("ITL (s)", fontsize=label_fs)
ax.set_title("(c) Per-model average ITL", pad=title_pad)
ax.grid(True, alpha=0.25)
ax.tick_params(labelsize=tick_fs)
ax.set_ylim(0, 0.2)
ax.legend(fontsize=legend_fs, ncol=3, loc="upper left", framealpha=0.9)
# (d) Total throughput per model (tokens/s, log scale)
ax = axes[3]
for lab in model_labels:
if lab not in per_model.itl_s or lab not in per_model.batch_size:
continue
itl = per_model.itl_s[lab]
bs = per_model.batch_size[lab].astype(float)
wrep = per_model.active_replicas[lab].astype(float)
with np.errstate(divide="ignore", invalid="ignore"):
throughput = np.where(itl > 0, bs * wrep / itl, np.nan)
ax.plot(per_model.time_s / 60.0, throughput, lw=lw_main, color=model_to_color[lab], label=lab)
ax.set_yscale("log")
_apply_overlays(ax)
ax.set_ylabel("Tokens/s", fontsize=label_fs)
ax.set_title("(d) Per-model total token throughput (log scale)", pad=title_pad)
ax.grid(True, alpha=0.25)
ax.tick_params(labelsize=tick_fs)
ax.set_xlabel("Time (minutes)", fontsize=label_fs)
if legend_patches:
fig.legend(
handles=legend_patches,
loc="upper center",
ncol=len(legend_patches),
fontsize=10,
frameon=False,
bbox_to_anchor=(0.5, 1.05),
handlelength=2.2,
handleheight=1.0,
columnspacing=1.6,
borderaxespad=0.2,
)
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
# ══════════════════════════════════════════════════════════════════════
# Standalone plots
# ══════════════════════════════════════════════════════════════════════
def plot_power_3ph(
time_s: np.ndarray,
kW_A: np.ndarray,
kW_B: np.ndarray,
kW_C: np.ndarray,
*,
save_path: Path | str | None = None,
title: str = "DC Power by Phase",
t_train_start_s: float = 1000.0,
t_train_end_s: float = 2000.0,
t_ramp_start_s: float = 2500.0,
t_ramp_end_s: float = 3000.0,
) -> Figure:
"""3-phase power in MW with training/ramp overlays."""
t = np.asarray(time_s) / 60.0
fig, ax = plt.subplots(figsize=(11, 3.2), dpi=160)
ax.plot(t, np.asarray(kW_A) / 1e3, lw=1.0, label="Phase A")
ax.plot(t, np.asarray(kW_B) / 1e3, lw=1.0, label="Phase B")
ax.plot(t, np.asarray(kW_C) / 1e3, lw=1.0, label="Phase C")
ax.axvspan(
t_train_start_s / 60.0,
t_train_end_s / 60.0,
alpha=0.15,
label=f"Training overlay ({t_train_start_s:.0f}\u2013{t_train_end_s:.0f}s)",
)
ax.axvspan(
t_ramp_start_s / 60.0,
t_ramp_end_s / 60.0,
alpha=0.12,
label=f"Ramp to 20% ({t_ramp_start_s:.0f}\u2013{t_ramp_end_s:.0f}s)",
)
ax.set_xlabel("Time (min)")
ax.set_ylabel("Power (MW)")
ax.set_title(title)
ax.grid(True, alpha=0.25)
ax.legend(ncol=3, fontsize=9)
fig.tight_layout()
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
def plot_voltage_dc_bus(
time_s: np.ndarray,
Va: np.ndarray,
Vb: np.ndarray,
Vc: np.ndarray,
*,
v_min: float = 0.95,
v_max: float = 1.05,
save_path: Path | str | None = None,
title: str = "Voltage at DC bus (3-phase)",
) -> Figure:
"""DC-bus 3-phase voltage with limit lines."""
t = np.asarray(time_s) / 60.0
fig, ax = plt.subplots(figsize=(11, 3.2), dpi=160)
ax.plot(t, np.asarray(Va), lw=1.0, label="Va @ DC bus")
ax.plot(t, np.asarray(Vb), lw=1.0, alpha=0.8, label="Vb @ DC bus")
ax.plot(t, np.asarray(Vc), lw=1.0, alpha=0.8, label="Vc @ DC bus")
ax.axhline(v_min, ls="--", lw=1.6, label="v_min")
ax.axhline(v_max, ls="--", lw=1.6, label="v_max")
ax.set_xlabel("Time (min)")
ax.set_ylabel("Voltage (pu)")
ax.set_title(title)
ax.grid(True, alpha=0.25)
ax.legend(ncol=5, fontsize=9)
fig.tight_layout()
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
def plot_batch_schedule(
per_model: PerModelTimeSeries,
*,
save_path: Path | str | None = None,
title: str = "Per-model batch schedule (closed-loop)",
) -> Figure:
"""log2 batch schedule, step plot with integer y-ticks."""
fig, ax = plt.subplots(figsize=(11, 2.6), dpi=160)
t_min = per_model.time_s / 60.0
for label, b in per_model.batch_size.items():
ax.step(t_min[: b.size], np.log2(b), where="post", lw=1.4, label=label)
ax.set_xlabel("Time (min)")
ax.set_ylabel(r"$\log_2(\mathrm{batch\ size})$")
ax.set_title(title)
ax.grid(True, alpha=0.25)
ymin, ymax = ax.get_ylim()
ax.set_yticks(np.arange(math.floor(ymin), math.ceil(ymax) + 1))
ax.legend(ncol=2, fontsize=8)
fig.tight_layout()
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
def plot_latency_samples(
per_model: PerModelTimeSeries,
*,
itl_deadlines: dict[str, float] | None = None,
save_path: Path | str | None = None,
title: str = "Sampled ITL measurements (applied batch)",
) -> Figure:
"""Sampled average ITL per DC timestep with optional deadline overlays."""
fig, ax = plt.subplots(figsize=(11, 2.8), dpi=160)
t_min = per_model.time_s / 60.0
for label, y in per_model.itl_s.items():
ax.plot(t_min, y, lw=1.2, marker="o", ms=2.5, label=f"ITL[{label}]")
if itl_deadlines and label in itl_deadlines:
ax.axhline(float(itl_deadlines[label]), ls="--", lw=1.0)
ax.set_xlabel("Time (min)")
ax.set_ylabel("Sampled avg ITL (s)")
ax.set_title(title)
ax.grid(True, alpha=0.25)
ax.legend(ncol=2, fontsize=8)
fig.tight_layout()
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
def plot_per_model_power(
per_model: PerModelTimeSeries,
*,
save_path: Path | str | None = None,
title: str = "Per-model measured power (whole simulation)",
) -> Figure:
"""Per-model total power in MW."""
fig, ax = plt.subplots(figsize=(11, 3.2), dpi=160)
t_min = per_model.time_s / 60.0
for label, y_w in per_model.power_w.items():
ax.plot(t_min, np.asarray(y_w) / 1e6, lw=1.1, label=label)
ax.set_xlabel("Time (min)")
ax.set_ylabel("Power (MW)")
ax.set_title(title)
ax.grid(True, alpha=0.25)
ax.legend(ncol=2, fontsize=8)
fig.tight_layout()
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
# ══════════════════════════════════════════════════════════════════════
# Online multi-panel: GPU power, batch size, ITL, KV cache
# (ported from g2g/analyze.py, adapted for SimulationLog)
# ══════════════════════════════════════════════════════════════════════
@dataclass
class OnlinePerModelTimeSeries:
"""Per-model time series extracted from online datacenter states."""
time_s: np.ndarray
measured_power_w: dict[str, np.ndarray] = field(default_factory=dict)
batch_size: dict[str, np.ndarray] = field(default_factory=dict)
itl_s: dict[str, np.ndarray] = field(default_factory=dict)
kv_cache_pct: dict[str, np.ndarray] = field(default_factory=dict)
num_requests_running: dict[str, np.ndarray] = field(default_factory=dict)
def extract_online_per_model_timeseries(
dc_states: Sequence[OnlineDatacenterState],
) -> OnlinePerModelTimeSeries:
"""Build per-model arrays from a list of `OnlineDatacenterState` objects."""
if not dc_states:
raise ValueError("dc_states is empty.")
time_s = np.array([s.time_s for s in dc_states])
model_labels = sorted(dc_states[0].batch_size_by_model.keys())
measured_power_w: dict[str, np.ndarray] = {}
batch_size: dict[str, np.ndarray] = {}
itl_s: dict[str, np.ndarray] = {}
kv_cache_pct: dict[str, np.ndarray] = {}
num_requests_running: dict[str, np.ndarray] = {}
for label in model_labels:
measured_power_w[label] = np.array([s.measured_power_w_by_model.get(label, 0.0) for s in dc_states])
batch_size[label] = np.array([s.batch_size_by_model.get(label, 0) for s in dc_states])
itl_s[label] = np.array([s.observed_itl_s_by_model.get(label, float("nan")) for s in dc_states])
kv_cache_pct[label] = np.array(
[
s.prometheus_metrics_by_model.get(label, {}).get("kv_cache_usage_perc", float("nan")) * 100.0
for s in dc_states
]
)
num_requests_running[label] = np.array(
[s.prometheus_metrics_by_model.get(label, {}).get("num_requests_running", float("nan")) for s in dc_states]
)
return OnlinePerModelTimeSeries(
time_s=time_s,
measured_power_w=measured_power_w,
batch_size=batch_size,
itl_s=itl_s,
kv_cache_pct=kv_cache_pct,
num_requests_running=num_requests_running,
)
def plot_online_timeseries(
per_model: OnlinePerModelTimeSeries,
model_labels: list[str],
*,
schedule_times_s: list[float] | None = None,
figsize: tuple[float, float] = (14, 12),
dpi: int = 200,
save_path: Path | str | None = None,
) -> Figure:
"""Multi-panel online time series: GPU power, batch size, ITL, KV cache.
Args:
per_model: Extracted per-model time series from online states.
model_labels: Ordered model labels for consistent color assignment.
schedule_times_s: Optional list of batch size change times (seconds)
to draw as vertical lines on all panels.
figsize: Figure size in inches.
dpi: Figure resolution.
save_path: If given, save figure to this path.
Returns:
The matplotlib Figure object.
"""
color_cycle = plt.rcParams["axes.prop_cycle"].by_key().get("color", [])
if not color_cycle:
cmap = plt.get_cmap("tab10")
color_cycle = [cmap(i) for i in range(10)]
model_to_color = {lab: color_cycle[i % len(color_cycle)] for i, lab in enumerate(model_labels)}
t_s = per_model.time_s
label_fs = 10
tick_fs = 9
legend_fs = 8
fig, axes = plt.subplots(4, 1, sharex=True, figsize=figsize, dpi=dpi, constrained_layout=True)
def _draw_schedule_lines(ax: Axes) -> None:
if schedule_times_s:
for t in schedule_times_s:
ax.axvline(t, color="gray", alpha=0.3, linewidth=0.7, linestyle=":")
# (a) GPU Power (measured, per model)
ax = axes[0]
for lab in model_labels:
if lab in per_model.measured_power_w:
ax.plot(t_s, per_model.measured_power_w[lab], lw=0.8, color=model_to_color[lab], label=lab)
ax.set_ylabel("Measured GPU Power (W)", fontsize=label_fs)
ax.set_title("(a) Per-model measured GPU power")
ax.set_ylim(bottom=0)
ax.grid(True, alpha=0.25)
ax.legend(fontsize=legend_fs, ncol=4, loc="upper right")
ax.tick_params(labelsize=tick_fs)
_draw_schedule_lines(ax)
# (b) Batch size (from prometheus num_requests_running or state)
ax = axes[1]
for lab in model_labels:
if lab in per_model.num_requests_running:
y = per_model.num_requests_running[lab]
if not np.all(np.isnan(y)):
ax.step(t_s, y, where="post", lw=1.2, color=model_to_color[lab], label=f"{lab} (running)")
if lab in per_model.batch_size:
ax.step(
t_s,
per_model.batch_size[lab],
where="post",
lw=0.8,
color=model_to_color[lab],
linestyle="--",
alpha=0.6,
label=f"{lab} (set)",
)
ax.set_ylabel("Requests / Batch Size", fontsize=label_fs)
ax.set_title("(b) Batch size and requests running")
ax.set_ylim(bottom=0)
ax.grid(True, alpha=0.25)
ax.legend(fontsize=legend_fs, ncol=4, loc="upper right")
ax.tick_params(labelsize=tick_fs)
_draw_schedule_lines(ax)
# (c) ITL
ax = axes[2]
for lab in model_labels:
if lab in per_model.itl_s:
ax.plot(t_s, per_model.itl_s[lab] * 1e3, lw=0.8, color=model_to_color[lab], label=lab)
ax.set_ylabel("ITL (ms)", fontsize=label_fs)
ax.set_title("(c) Per-model average ITL")
ax.set_ylim(bottom=0)
ax.grid(True, alpha=0.25)
ax.legend(fontsize=legend_fs, ncol=4, loc="upper right")
ax.tick_params(labelsize=tick_fs)
_draw_schedule_lines(ax)
# (d) KV cache usage
ax = axes[3]
for lab in model_labels:
if lab in per_model.kv_cache_pct:
y = per_model.kv_cache_pct[lab]
if not np.all(np.isnan(y)):
ax.plot(t_s, y, lw=1.0, color=model_to_color[lab], label=lab)
ax.set_ylabel("KV Cache Usage (%)", fontsize=label_fs)
ax.set_title("(d) KV cache usage")
ax.set_ylim(bottom=0)
ax.grid(True, alpha=0.25)
ax.legend(fontsize=legend_fs, ncol=4, loc="upper right")
ax.tick_params(labelsize=tick_fs)
ax.set_xlabel("Time (s)", fontsize=label_fs)
_draw_schedule_lines(ax)
if save_path is not None:
fig.savefig(save_path, bbox_inches="tight")
plt.close(fig)
return fig
|