diff --git a/pytorch/analyze.py b/pytorch/analyze.py new file mode 100755 index 0000000..bf1c73d --- /dev/null +++ b/pytorch/analyze.py @@ -0,0 +1,189 @@ +#! /bin/python3 + +from perf_stat import Stat, CPU + +import argparse +import os, glob +import re +import json + +from enum import Enum + +import math +import numpy as np +import matplotlib.pyplot as plt +import itertools + +class Plot(Enum): + BOX = 'box' + LINE = 'line' + +def accumulate(stats_list: list[dict[str, str | int | float]], category: str, value: str): + print(category) + print(value) + category_list = np.array([stats[category] for stats in stats_list if value in stats]) + value_list = np.array([stats[value] for stats in stats_list if value in stats]) + + result: dict[np.ndarray] = dict() + for category in np.sort(np.unique(category_list)): + result[category] = value_list[category_list == category] + + return result + +def box_plot(ax, stats_list: list[dict[str, str | int | float]], x: Stat, y: Stat): + data: dict[str, np.ndarray] = accumulate(stats_list, x, y) + + print("Plotted data: " + str(data)) + + ax.boxplot(data.values(), tick_labels=data.keys()) + ax.set_ylabel(y.value) + +def line_plot( + ax, stats_list: list[dict[str, str | int | float]], + x: Stat, y: Stat, color: Stat +): + x_data: dict[str, np.ndarray] = accumulate(stats_list, color, x) + y_data: dict[str, np.ndarray] = accumulate(stats_list, color, y) + + for category in x_data.keys(): + sorted_indices = np.argsort(x_data[category]) + x_data[category] = x_data[category][sorted_indices] + y_data[category] = y_data[category][sorted_indices] + ax.plot(x_data[category], y_data[category], label=category) + print("Plotted x data: " + str(x_data[category])) + print("Plotted y data: " + str(y_data[category])) + + ax.set_ylabel(y) + ax.grid(True) + +def visualize( + stats_list: list[dict[str, str | int | float]], + plot: Plot, + rows: int, + size_multiplier: int, + font_size: int, + x: Stat, + y: Stat, + color: Stat, + filter_list: list[str] = [] +): + # Remove stats entries containing undesired values (like a specific CPU). +# stats_list = [stats for stats in stats_list +# if len([stats[key] for key in stats.keys() +# if stats[key] in filter_list]) == 0] + + #x = Stat.MAXWELL_SIZE + #y = Stat.DTLB_MISS_RATE + #color = Stat.SOLVER + + if y is None: + #ys = [stat for stat in Stat if stat.value in stats_list[0].keys() + ys = [stat for stat in stats_list[0].keys() if "power" not in stat] + #and stat is not x + #and y != color + #and y != marker + #and stat.value not in filter_list] + fig, axes = plt.subplots(rows, int(math.ceil(len(ys) / rows)), + figsize = (16 * size_multiplier, 9 * size_multiplier)) + match plot: + case Plot.BOX: + for i, y in enumerate(ys): + box_plot(axes[i % rows][int(i / rows)], stats_list, x, y) + case Plot.LINE: + for i, y in enumerate(ys): + line_plot(axes[i % rows][int(i / rows)], stats_list, x, y, color) + + handles, labels = axes[i % rows][int(i / rows)].get_legend_handles_labels() + else: + fig, ax = plt.subplots() + + match plot: + case Plot.BOX: + box_plot(ax, stats_list, x, y) + case Plot.LINE: + line_plot(ax, stats_list, x, y, color) + + handles, labels = ax.get_legend_handles_labels() + + #box_plot(ax, stats, x, y) + #line_plot(ax, stats, x, y, color) + + match plot: + case Plot.BOX: + title = f"{plot.value}_plot_of_{y.value.replace(' ', '_')}_vs_{x.value.replace(' ', '_')}_excluding_{filter_list}" + case Plot.LINE: + #title = f"{plot.value}_plot_of_{y.replace(' ', '_')}_vs_{x.replace(' ', '_')}_by_{color.replace(' ', '_')}_excluding_{filter_list}" + title = "altra_spmv" + fig.suptitle(title, fontsize = font_size) + fig.legend(handles, labels, fontsize = font_size) + fig.supxlabel(x, fontsize = font_size) + + plt.savefig(title + ".png", dpi = 100) + plt.show() + +def main(): + class Command(Enum): + PARSE = 'parse' + VISUALIZE = 'visualize' + + parser = argparse.ArgumentParser() + parser.add_argument('command', choices=[x.value for x in Command]) + parser.add_argument('filepath', + help='the output for the ' + Command.PARSE.value + ' command or the input for the ' + Command.VISUALIZE.value + ' command') + parser.add_argument('-i', '--input_dir', + help='the input directory for the parse command') + parser.add_argument('-p', '--plot', + choices=[x.name.lower() for x in Plot], + help = 'the type of plot') + parser.add_argument('-r', '--rows', type=int, + help = 'the number of rows to display when -y is not specified', + default = 5) + parser.add_argument('-s', '--size', type=int, + help = 'figure size multiplier', + default = 4) + parser.add_argument('-fs', '--font_size', type=int, + help = 'font size', + default = 40) + parser.add_argument('-x', + #choices=[x.name.lower() for x in Stat], + help = 'the name of the x axis') + parser.add_argument('-y', + #choices=[x.name.lower() for x in Stat], + help = 'the name of the y axis') + parser.add_argument('-c', '--color', + #choices=[x.name.lower() for x in Stat], + help = 'the name of the color') + parser.add_argument('-f', '--filter', nargs = '+', + help = 'a comma-separated string of names and values to filter out.', + default = []) + + args = parser.parse_args() + + stats_list: list[dict] = list() + if args.command == Command.PARSE.value: + if (args.input_dir) is None: + print("An input directory is required with -i") + exit(-1) + + for filename in glob.glob(f'{args.input_dir.rstrip("/")}/*.json'): + with open(filename, 'r') as file: + stats_list.append(json.load(file)) + print(filename + " loaded.") + + with open(args.filepath, 'w') as file: + json.dump(stats_list, file, indent = 2) + + elif args.command == Command.VISUALIZE.value: + with open(args.filepath, 'r') as file: + stats_list = json.load(file) + + #x = Stat[args.x.upper()] if args.x is not None else None + x = args.x + #y = Stat[args.y.upper()] if args.y is not None else None + y = args.y + #color = Stat[args.color.upper()] if args.color is not None else None + color = args.color + visualize(stats_list, Plot[args.plot.upper()], args.rows, args.size, args.font_size, x, y, color, args.filter) + +if __name__ == '__main__': + main() diff --git a/pytorch/output.json b/pytorch/output.json new file mode 100644 index 0000000..6948648 --- /dev/null +++ b/pytorch/output.json @@ -0,0 +1,859 @@ +[ + { + "power_before": [ + 29.76, + 33.16 + ], + "shape": [ + 22687, + 22687 + ], + "nnz": 54705, + "% density": 0.00010628522108964806, + "time_s": 0.14322686195373535, + "power": [ + 22.6, + 22.6, + 26.16, + 29.2 + ], + "power_after": [ + 34.0, + 30.16 + ], + "task clock (msec)": 64.71, + "page faults": 3319, + "cycles": 57611295, + "instructions": 83148228, + "branch mispredictions": 318386, + "branches": 19233431, + "ITLB accesses": 27039805, + "ITLB misses": 6375, + "DTLB misses": 17290, + "DTLB accesses": 36688544, + "L1I cache accesses": 32508072, + "L1I cache misses": 297568, + "L1D cache misses": 477654, + "L1D cache accesses": 34044579, + "LL cache misses": 549474, + "LL cache accesses": 561939, + "L2D TLB accesses": 185622, + "L2D TLB misses": 23295, + "L2D cache misses": 305878, + "L2D cache accesses": 1763089, + "instructions per cycle": 1.4432626102225268, + "branch miss rate": 0.01655378075809771, + "ITLB miss rate": 0.00023576353453732377, + "DTLB miss rate": 0.00047126427257511227, + "L2D TLB miss rate": 0.12549697772893298, + "L1I cache miss rate": 0.009153664972810446, + "L1D cache miss rate": 0.014030251336049713, + "L2D cache miss rate": 0.17348982382625042, + "LL cache miss rate": 0.9778178770293573, + "architecture": "altra" + }, + { + "power_before": [ + 20.48, + 20.96 + ], + "shape": [ + 24115, + 24115 + ], + "nnz": 116056, + "% density": 0.0001995689928120616, + "time_s": 0.3271017074584961, + "power": [ + 25.28, + 26.08, + 31.28, + 32.96 + ], + "power_after": [ + 33.4, + 30.24 + ], + "task clock (msec)": 59.88, + "page faults": 3313, + "cycles": 58169777, + "instructions": 57993431, + "branch mispredictions": 330494, + "branches": 20578427, + "ITLB accesses": 27982097, + "ITLB misses": 6614, + "DTLB misses": 17270, + "DTLB accesses": 37728899, + "L1I cache accesses": 29754926, + "L1I cache misses": 278786, + "L1D cache misses": 454742, + "L1D cache accesses": 31173246, + "LL cache misses": 543243, + "LL cache accesses": 560716, + "L2D TLB accesses": 162281, + "L2D TLB misses": 19847, + "L2D cache misses": 300577, + "L2D cache accesses": 1696278, + "instructions per cycle": 0.9969684257170179, + "branch miss rate": 0.016060216847478187, + "ITLB miss rate": 0.0002363654160729984, + "DTLB miss rate": 0.00045773930482307474, + "L2D TLB miss rate": 0.12230020766448321, + "L1I cache miss rate": 0.009369406598423401, + "L1D cache miss rate": 0.014587572946365611, + "L2D cache miss rate": 0.1771979592967662, + "LL cache miss rate": 0.9688380570556218, + "architecture": "altra" + }, + { + "power_before": [ + 20.28, + 20.32 + ], + "shape": [ + 36692, + 36692 + ], + "nnz": 367662, + "% density": 0.0002730901120626302, + "time_s": 1.030203104019165, + "power": [ + 32.08, + 47.84, + 55.76, + 58.08, + 58.24 + ], + "power_after": [ + 48.76, + 45.16 + ], + "task clock (msec)": 60.43, + "page faults": 3319, + "cycles": 66114448, + "instructions": 90786829, + "branch mispredictions": 341625, + "branches": 20129354, + "ITLB accesses": 27441303, + "ITLB misses": 6807, + "DTLB misses": 20551, + "DTLB accesses": 36867114, + "L1I cache accesses": 31744243, + "L1I cache misses": 271027, + "L1D cache misses": 464135, + "L1D cache accesses": 33441141, + "LL cache misses": 539935, + "LL cache accesses": 552519, + "L2D TLB accesses": 188291, + "L2D TLB misses": 24177, + "L2D cache misses": 301281, + "L2D cache accesses": 1737575, + "instructions per cycle": 1.3731768432824245, + "branch miss rate": 0.016971483535934636, + "ITLB miss rate": 0.00024805673404065397, + "DTLB miss rate": 0.0005574344658494288, + "L2D TLB miss rate": 0.12840231344036623, + "L1I cache miss rate": 0.008537831568388637, + "L1D cache miss rate": 0.01387916159918108, + "L2D cache miss rate": 0.17339165215889962, + "LL cache miss rate": 0.9772243126480719, + "architecture": "altra" + }, + { + "power_before": [ + 50.88, + 50.88 + ], + "shape": [ + 11806, + 11806 + ], + "nnz": 65460, + "% density": 0.0004696458003979807, + "time_s": 0.1896660327911377, + "power": [ + 25.52, + 32.28, + 33.12, + 33.12 + ], + "power_after": [ + 32.88, + 26.52 + ], + "task clock (msec)": 42.01, + "page faults": 3263, + "cycles": 47084933, + "instructions": 77895119, + "branch mispredictions": 330923, + "branches": 19740519, + "ITLB accesses": 27761239, + "ITLB misses": 6471, + "DTLB misses": 17268, + "DTLB accesses": 36993265, + "L1I cache accesses": 31834980, + "L1I cache misses": 298333, + "L1D cache misses": 466901, + "L1D cache accesses": 33528976, + "LL cache misses": 525505, + "LL cache accesses": 546521, + "L2D TLB accesses": 184884, + "L2D TLB misses": 22933, + "L2D cache misses": 292367, + "L2D cache accesses": 1706226, + "instructions per cycle": 1.6543534000568716, + "branch miss rate": 0.016763642333821112, + "ITLB miss rate": 0.00023309478370183695, + "DTLB miss rate": 0.0004667876706746485, + "L2D TLB miss rate": 0.12403993855606758, + "L1I cache miss rate": 0.009371232524725947, + "L1D cache miss rate": 0.013925298523879763, + "L2D cache miss rate": 0.1713530329510862, + "LL cache miss rate": 0.9615458509371094, + "architecture": "altra" + }, + { + "power_before": [ + 30.44, + 35.52 + ], + "shape": [ + 25181, + 25181 + ], + "nnz": 125750, + "% density": 0.00019831796057928155, + "time_s": 0.29622840881347656, + "power": [ + 23.84, + 29.44, + 33.0, + 33.04 + ], + "power_after": [ + 36.32, + 30.0 + ], + "task clock (msec)": 60.77, + "page faults": 3361, + "cycles": 63493475, + "instructions": 91578911, + "branch mispredictions": 329084, + "branches": 20406595, + "ITLB accesses": 26859919, + "ITLB misses": 6237, + "DTLB misses": 16689, + "DTLB accesses": 36348977, + "L1I cache accesses": 30979764, + "L1I cache misses": 292038, + "L1D cache misses": 469219, + "L1D cache accesses": 32411890, + "LL cache misses": 571870, + "LL cache accesses": 598306, + "L2D TLB accesses": 205488, + "L2D TLB misses": 26392, + "L2D cache misses": 342141, + "L2D cache accesses": 1857697, + "instructions per cycle": 1.442335783322617, + "branch miss rate": 0.01612635522976763, + "ITLB miss rate": 0.00023220472109390948, + "DTLB miss rate": 0.0004591325912693499, + "L2D TLB miss rate": 0.12843572374055906, + "L1I cache miss rate": 0.009426734173959492, + "L1D cache miss rate": 0.014476755289494072, + "L2D cache miss rate": 0.1841748142996409, + "LL cache miss rate": 0.9558152517273769, + "architecture": "altra" + }, + { + "power_before": [ + 30.48, + 33.04 + ], + "shape": [ + 131828, + 131828 + ], + "nnz": 841372, + "% density": 4.841419648464106e-05, + "time_s": 2.848874092102051, + "power": [ + 65.52, + 75.88, + 71.16, + 71.16, + 59.72, + 47.92, + 48.68 + ], + "power_after": [ + 68.68, + 67.88 + ], + "task clock (msec)": 49.87, + "page faults": 3300, + "cycles": 51935476, + "instructions": 83731856, + "branch mispredictions": 326464, + "branches": 20341367, + "ITLB accesses": 27590154, + "ITLB misses": 6210, + "DTLB misses": 17536, + "DTLB accesses": 36763243, + "L1I cache accesses": 31663300, + "L1I cache misses": 289727, + "L1D cache misses": 462864, + "L1D cache accesses": 33262254, + "LL cache misses": 530272, + "LL cache accesses": 551373, + "L2D TLB accesses": 196152, + "L2D TLB misses": 23542, + "L2D cache misses": 301998, + "L2D cache accesses": 1732662, + "instructions per cycle": 1.6122285275675532, + "branch miss rate": 0.01604926551888081, + "ITLB miss rate": 0.000225080294948698, + "DTLB miss rate": 0.0004769981799483794, + "L2D TLB miss rate": 0.12001916880786329, + "L1I cache miss rate": 0.00915024649989104, + "L1D cache miss rate": 0.013915593332911234, + "L2D cache miss rate": 0.17429712200071334, + "LL cache miss rate": 0.9617300810884828, + "architecture": "altra" + }, + { + "power_before": [ + 20.44, + 20.2 + ], + "shape": [ + 24818, + 24818 + ], + "nnz": 239978, + "% density": 0.00038961697406616504, + "time_s": 0.556269645690918, + "power": [ + 25.24, + 32.16, + 33.0, + 32.52 + ], + "power_after": [ + 34.24, + 30.28 + ], + "task clock (msec)": 62.49, + "page faults": 3312, + "cycles": 76783170, + "instructions": 77095702, + "branch mispredictions": 323514, + "branches": 19769937, + "ITLB accesses": 26809325, + "ITLB misses": 6925, + "DTLB misses": 19003, + "DTLB accesses": 36516965, + "L1I cache accesses": 31104231, + "L1I cache misses": 285499, + "L1D cache misses": 468498, + "L1D cache accesses": 32677465, + "LL cache misses": 559358, + "LL cache accesses": 571935, + "L2D TLB accesses": 194840, + "L2D TLB misses": 23481, + "L2D cache misses": 313487, + "L2D cache accesses": 1779730, + "instructions per cycle": 1.004070319055595, + "branch miss rate": 0.016363936819829016, + "ITLB miss rate": 0.00025830564551699827, + "DTLB miss rate": 0.0005203882633729282, + "L2D TLB miss rate": 0.12051426811742968, + "L1I cache miss rate": 0.009178783426601994, + "L1D cache miss rate": 0.01433703624194839, + "L2D cache miss rate": 0.1761430104566423, + "LL cache miss rate": 0.9780097388689274, + "architecture": "altra" + }, + { + "power_before": [ + 16.52, + 16.64 + ], + "shape": [ + 81871, + 81871 + ], + "nnz": 545671, + "% density": 8.140867447881048e-05, + "time_s": 1.3372814655303955, + "power": [ + 23.92, + 38.6, + 46.04, + 48.2, + 48.2 + ], + "power_after": [ + 45.0, + 44.08 + ], + "task clock (msec)": 59.01, + "page faults": 3448, + "cycles": 73062796, + "instructions": 88329175, + "branch mispredictions": 331091, + "branches": 20013316, + "ITLB accesses": 26330936, + "ITLB misses": 5193, + "DTLB misses": 16837, + "DTLB accesses": 35930477, + "L1I cache accesses": 31853890, + "L1I cache misses": 306147, + "L1D cache misses": 479933, + "L1D cache accesses": 33426019, + "LL cache misses": 540302, + "LL cache accesses": 553181, + "L2D TLB accesses": 173206, + "L2D TLB misses": 21390, + "L2D cache misses": 300032, + "L2D cache accesses": 1739931, + "instructions per cycle": 1.2089487377406143, + "branch miss rate": 0.016543535314187813, + "ITLB miss rate": 0.0001972204861991993, + "DTLB miss rate": 0.000468599401004334, + "L2D TLB miss rate": 0.12349456716280037, + "L1I cache miss rate": 0.009610976869701, + "L1D cache miss rate": 0.014358066391334247, + "L2D cache miss rate": 0.17243902200719455, + "LL cache miss rate": 0.9767182893121781, + "architecture": "altra" + }, + { + "power_before": [ + 50.68, + 49.4 + ], + "shape": [ + 10879, + 10879 + ], + "nnz": 39994, + "% density": 0.0003379223282393842, + "time_s": 0.11296772956848145, + "power": [ + 26.2, + 29.76, + 33.64, + 34.44 + ], + "power_after": [ + 36.84, + 29.44 + ], + "task clock (msec)": 67.56, + "page faults": 3829, + "cycles": 47862000, + "instructions": 84392375, + "branch mispredictions": 331622, + "branches": 19800140, + "ITLB accesses": 25905045, + "ITLB misses": 6746, + "DTLB misses": 17547, + "DTLB accesses": 35220079, + "L1I cache accesses": 30359576, + "L1I cache misses": 283204, + "L1D cache misses": 465520, + "L1D cache accesses": 31843274, + "LL cache misses": 560542, + "LL cache accesses": 575610, + "L2D TLB accesses": 173643, + "L2D TLB misses": 21499, + "L2D cache misses": 313335, + "L2D cache accesses": 1741621, + "instructions per cycle": 1.7632438051063475, + "branch miss rate": 0.016748467435078743, + "ITLB miss rate": 0.0002604125953072075, + "DTLB miss rate": 0.0004982101261044871, + "L2D TLB miss rate": 0.12381149830399152, + "L1I cache miss rate": 0.009328325270418797, + "L1D cache miss rate": 0.014619099782264852, + "L2D cache miss rate": 0.17990998041479747, + "LL cache miss rate": 0.9738225534650197, + "architecture": "altra" + }, + { + "power_before": [ + 53.64, + 46.88 + ], + "shape": [ + 82144, + 82144 + ], + "nnz": 549202, + "% density": 8.13917555860553e-05, + "time_s": 1.2292509078979492, + "power": [ + 40.64, + 52.44, + 54.8, + 54.96, + 46.8 + ], + "power_after": [ + 47.88, + 47.08 + ], + "task clock (msec)": 61.26, + "page faults": 3303, + "cycles": 44515786, + "instructions": 81513738, + "branch mispredictions": 328019, + "branches": 19893662, + "ITLB accesses": 27248112, + "ITLB misses": 5792, + "DTLB misses": 16632, + "DTLB accesses": 36929042, + "L1I cache accesses": 31702830, + "L1I cache misses": 295778, + "L1D cache misses": 470423, + "L1D cache accesses": 33155119, + "LL cache misses": 545220, + "LL cache accesses": 562139, + "L2D TLB accesses": 192206, + "L2D TLB misses": 24891, + "L2D cache misses": 307033, + "L2D cache accesses": 1782260, + "instructions per cycle": 1.8311198189334452, + "branch miss rate": 0.01648861833482443, + "ITLB miss rate": 0.0002125651861677609, + "DTLB miss rate": 0.0004503772396803578, + "L2D TLB miss rate": 0.12950168048864238, + "L1I cache miss rate": 0.009329703373484323, + "L1D cache miss rate": 0.014188548079106578, + "L2D cache miss rate": 0.17227172241984895, + "LL cache miss rate": 0.9699024618466251, + "architecture": "altra" + }, + { + "power_before": [ + 34.04, + 43.96 + ], + "shape": [ + 32580, + 32580 + ], + "nnz": 155598, + "% density": 0.00014658915806621921, + "time_s": 0.4164857864379883, + "power": [ + 23.72, + 23.72, + 29.88, + 33.32 + ], + "power_after": [ + 33.36, + 32.52 + ], + "task clock (msec)": 61.63, + "page faults": 3304, + "cycles": 64734203, + "instructions": 53597991, + "branch mispredictions": 330777, + "branches": 20357034, + "ITLB accesses": 27381387, + "ITLB misses": 6248, + "DTLB misses": 17636, + "DTLB accesses": 37436110, + "L1I cache accesses": 32505993, + "L1I cache misses": 303849, + "L1D cache misses": 467426, + "L1D cache accesses": 34241110, + "LL cache misses": 550075, + "LL cache accesses": 562829, + "L2D TLB accesses": 199285, + "L2D TLB misses": 24424, + "L2D cache misses": 310155, + "L2D cache accesses": 1783824, + "instructions per cycle": 0.8279701999266138, + "branch miss rate": 0.016248781625063848, + "ITLB miss rate": 0.00022818420410916364, + "DTLB miss rate": 0.00047109595521543235, + "L2D TLB miss rate": 0.12255814536969667, + "L1I cache miss rate": 0.009347476325365603, + "L1D cache miss rate": 0.01365101773861887, + "L2D cache miss rate": 0.17387085272986572, + "LL cache miss rate": 0.9773394761108614, + "architecture": "altra" + }, + { + "power_before": [ + 16.32, + 16.2 + ], + "shape": [ + 116835, + 116835 + ], + "nnz": 766396, + "% density": 5.614451099680581e-05, + "time_s": 2.2665774822235107, + "power": [ + 35.16, + 50.8, + 53.4, + 53.4, + 46.08, + 46.88 + ], + "power_after": [ + 58.4, + 57.32 + ], + "task clock (msec)": 50.43, + "page faults": 3285, + "cycles": 54118679, + "instructions": 77692421, + "branch mispredictions": 325039, + "branches": 19383216, + "ITLB accesses": 26060519, + "ITLB misses": 4749, + "DTLB misses": 16865, + "DTLB accesses": 34819729, + "L1I cache accesses": 30777115, + "L1I cache misses": 293980, + "L1D cache misses": 461522, + "L1D cache accesses": 32216597, + "LL cache misses": 567700, + "LL cache accesses": 588689, + "L2D TLB accesses": 189417, + "L2D TLB misses": 22360, + "L2D cache misses": 328306, + "L2D cache accesses": 1908607, + "instructions per cycle": 1.4355934482436277, + "branch miss rate": 0.0167690954896236, + "ITLB miss rate": 0.00018222967854170517, + "DTLB miss rate": 0.00048435184547243316, + "L2D TLB miss rate": 0.11804642666708902, + "L1I cache miss rate": 0.009551902444397404, + "L1D cache miss rate": 0.014325597455249542, + "L2D cache miss rate": 0.172013410827897, + "LL cache miss rate": 0.9643461997761127, + "architecture": "altra" + }, + { + "power_before": [ + 20.16, + 20.08 + ], + "shape": [ + 31379, + 31379 + ], + "nnz": 106762, + "% density": 0.00010842726485909405, + "time_s": 0.336850643157959, + "power": [ + 24.28, + 30.72, + 30.72, + 34.56 + ], + "power_after": [ + 37.32, + 32.92 + ], + "task clock (msec)": 60.78, + "page faults": 3300, + "cycles": 66733059, + "instructions": 87889334, + "branch mispredictions": 326300, + "branches": 19832700, + "ITLB accesses": 27233629, + "ITLB misses": 5868, + "DTLB misses": 16893, + "DTLB accesses": 36409508, + "L1I cache accesses": 30924532, + "L1I cache misses": 288199, + "L1D cache misses": 462816, + "L1D cache accesses": 32428375, + "LL cache misses": 551997, + "LL cache accesses": 568528, + "L2D TLB accesses": 193991, + "L2D TLB misses": 24353, + "L2D cache misses": 312207, + "L2D cache accesses": 1821196, + "instructions per cycle": 1.3170284011707, + "branch miss rate": 0.016452626218316214, + "ITLB miss rate": 0.0002154688969288669, + "DTLB miss rate": 0.00046397221297250155, + "L2D TLB miss rate": 0.125536751704976, + "L1I cache miss rate": 0.009319429635992551, + "L1D cache miss rate": 0.014271945479845968, + "L2D cache miss rate": 0.17142965391973186, + "LL cache miss rate": 0.9709231559395491, + "architecture": "altra" + }, + { + "power_before": [ + 20.56, + 20.28 + ], + "shape": [ + 36682, + 36682 + ], + "nnz": 88328, + "% density": 6.564359899804003e-05, + "time_s": 0.30861377716064453, + "power": [ + 23.88, + 27.6, + 39.8, + 40.12 + ], + "power_after": [ + 39.28, + 35.2 + ], + "task clock (msec)": 65.91, + "page faults": 3247, + "cycles": 92293071, + "instructions": 76208632, + "branch mispredictions": 320083, + "branches": 19285106, + "ITLB accesses": 26853940, + "ITLB misses": 6728, + "DTLB misses": 13955, + "DTLB accesses": 37111059, + "L1I cache accesses": 32554796, + "L1I cache misses": 298729, + "L1D cache misses": 473779, + "L1D cache accesses": 34117102, + "LL cache misses": 535040, + "LL cache accesses": 547502, + "L2D TLB accesses": 179876, + "L2D TLB misses": 21809, + "L2D cache misses": 298620, + "L2D cache accesses": 1722959, + "instructions per cycle": 0.8257243059990929, + "branch miss rate": 0.016597419791210898, + "ITLB miss rate": 0.0002505405165871377, + "DTLB miss rate": 0.0003760334621547717, + "L2D TLB miss rate": 0.12124463519313304, + "L1I cache miss rate": 0.009176190199440968, + "L1D cache miss rate": 0.013886847716432655, + "L2D cache miss rate": 0.17331811145825293, + "LL cache miss rate": 0.9772384393116372, + "architecture": "altra" + }, + { + "power_before": [ + 34.6, + 37.16 + ], + "shape": [ + 115406, + 115406 + ], + "nnz": 572066, + "% density": 4.295259032005559e-05, + "time_s": 1.0817186832427979, + "power": [ + 34.32, + 50.84, + 52.12, + 52.4, + 52.76 + ], + "power_after": [ + 49.0, + 45.08 + ], + "task clock (msec)": 60.55, + "page faults": 3490, + "cycles": 49977496, + "instructions": 78622993, + "branch mispredictions": 327078, + "branches": 20135808, + "ITLB accesses": 27608093, + "ITLB misses": 6616, + "DTLB misses": 17185, + "DTLB accesses": 36866957, + "L1I cache accesses": 32639204, + "L1I cache misses": 309643, + "L1D cache misses": 478856, + "L1D cache accesses": 34280618, + "LL cache misses": 555275, + "LL cache accesses": 578455, + "L2D TLB accesses": 188723, + "L2D TLB misses": 24635, + "L2D cache misses": 319663, + "L2D cache accesses": 1799940, + "instructions per cycle": 1.573167911413569, + "branch miss rate": 0.016243599462211798, + "ITLB miss rate": 0.00023963987661154286, + "DTLB miss rate": 0.00046613556958335347, + "L2D TLB miss rate": 0.13053522888042263, + "L1I cache miss rate": 0.009486842877663316, + "L1D cache miss rate": 0.013968709665619214, + "L2D cache miss rate": 0.17759647543807017, + "LL cache miss rate": 0.9599277385449171, + "architecture": "altra" + }, + { + "power_before": [ + 16.52, + 16.24 + ], + "shape": [ + 26518, + 26518 + ], + "nnz": 65369, + "% density": 9.295875717624285e-05, + "time_s": 0.1715233325958252, + "power": [ + 18.56, + 24.92, + 27.84, + 27.84 + ], + "power_after": [ + 33.2, + 27.28 + ], + "task clock (msec)": 61.92, + "page faults": 3281, + "cycles": 66250810, + "instructions": 75178179, + "branch mispredictions": 332366, + "branches": 19076182, + "ITLB accesses": 27005133, + "ITLB misses": 4791, + "DTLB misses": 13403, + "DTLB accesses": 36457054, + "L1I cache accesses": 32367686, + "L1I cache misses": 287524, + "L1D cache misses": 467557, + "L1D cache accesses": 34022862, + "LL cache misses": 535707, + "LL cache accesses": 556316, + "L2D TLB accesses": 150149, + "L2D TLB misses": 18418, + "L2D cache misses": 297042, + "L2D cache accesses": 1687364, + "instructions per cycle": 1.1347510920998551, + "branch miss rate": 0.017423088121092577, + "ITLB miss rate": 0.00017741071669597036, + "DTLB miss rate": 0.00036763804338112453, + "L2D TLB miss rate": 0.12266481961251822, + "L1I cache miss rate": 0.008883057009388932, + "L1D cache miss rate": 0.013742435895016709, + "L2D cache miss rate": 0.1760390763344483, + "LL cache miss rate": 0.9629545078696281, + "architecture": "altra" + } +] \ No newline at end of file diff --git a/pytorch/output_HPC.json b/pytorch/output_HPC.json new file mode 100644 index 0000000..9be58d5 --- /dev/null +++ b/pytorch/output_HPC.json @@ -0,0 +1,1406 @@ +[ + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 80.64, + 75.2, + 61.4, + 49.84, + 38.12, + 38.12, + 24.16, + 22.68, + 22.36, + 22.2 + ], + "matrix": "ASIC_680k", + "shape": [ + 682862, + 682862 + ], + "nnz": 3871773, + "% density": 8.303171256088674e-06, + "time_s": 41.51614689826965, + "power": [ + 92.4, + 92.44, + 84.28, + 73.04, + 59.64, + 53.28, + 56.32, + 65.88, + 80.28, + 93.76, + 96.16, + 94.44, + 94.44, + 94.2, + 93.92, + 92.48, + 92.16, + 91.84, + 92.08, + 91.84, + 91.68, + 90.68, + 90.88, + 90.28, + 90.28, + 92.44, + 92.52, + 92.84, + 90.0, + 89.64, + 88.16, + 87.28, + 88.12, + 88.24, + 88.08, + 85.72, + 85.12, + 85.12, + 81.72, + 82.52, + 83.84, + 86.32, + 88.8, + 91.0, + 90.2 + ], + "power_after": [ + 21.92, + 21.88, + 21.92, + 21.88, + 21.88, + 21.72, + 21.72, + 21.72, + 21.72, + 21.44 + ], + "task clock (msec)": 55.74, + "page faults": 3266, + "cycles": 51085608, + "instructions": 88049969, + "branch mispredictions": 332704, + "branches": 20219525, + "ITLB accesses": 27856157, + "ITLB misses": 6496, + "DTLB misses": 17046, + "DTLB accesses": 37522360, + "L1I cache accesses": 31475230, + "L1I cache misses": 277921, + "L1D cache misses": 462005, + "L1D cache accesses": 33126938, + "LL cache misses": 558923, + "LL cache accesses": 571263, + "L2D TLB accesses": 190627, + "L2D TLB misses": 24234, + "L2D cache misses": 314815, + "L2D cache accesses": 1760110, + "instructions per cycle": 1.7235768046452535, + "branch miss rate": 0.01645459030318467, + "ITLB miss rate": 0.00023319799640704206, + "DTLB miss rate": 0.0004542891225392006, + "L2D TLB miss rate": 0.12712784652751186, + "L1I cache miss rate": 0.008829832220447635, + "L1D cache miss rate": 0.013946504805243395, + "L2D cache miss rate": 0.17886098027964162, + "LL cache miss rate": 0.978398741035215, + "size": 466300511044 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 31.36, + 30.64, + 31.12, + 24.52, + 24.16, + 23.12, + 22.08, + 21.28, + 21.16, + 20.88 + ], + "matrix": "rma10", + "shape": [ + 46835, + 46835 + ], + "nnz": 2374001, + "% density": 0.0010822805369125833, + "time_s": 68.86891412734985, + "power": [ + 81.8, + 81.32, + 75.08, + 63.48, + 51.92, + 51.96, + 51.8, + 65.0, + 65.0, + 75.12, + 82.68, + 82.32, + 82.08, + 82.76, + 82.8, + 83.6, + 83.36, + 83.08, + 82.88, + 83.0, + 83.32, + 83.32, + 83.36, + 84.64, + 84.56, + 84.24, + 83.52, + 83.4, + 83.36, + 83.36, + 83.72, + 84.16, + 83.24, + 82.76, + 82.76, + 82.96, + 82.36, + 82.24, + 81.64, + 81.6, + 81.4, + 81.6, + 81.88, + 82.32, + 83.04, + 83.48, + 83.48, + 84.32, + 84.04, + 84.32, + 83.16, + 82.44, + 81.96, + 81.4, + 81.8, + 82.08, + 81.8, + 81.84, + 82.04, + 82.04, + 82.08, + 82.44, + 82.6, + 82.84, + 83.8, + 84.24, + 84.6, + 85.4, + 85.6, + 86.0, + 85.72, + 85.36 + ], + "power_after": [ + 21.96, + 21.88, + 21.96, + 21.96, + 22.0, + 21.68, + 21.44, + 21.16, + 21.04, + 20.92 + ], + "task clock (msec)": 58.3, + "page faults": 3281, + "cycles": 81319364, + "instructions": 90830397, + "branch mispredictions": 342237, + "branches": 20641135, + "ITLB accesses": 27974213, + "ITLB misses": 6660, + "DTLB misses": 18441, + "DTLB accesses": 37780346, + "L1I cache accesses": 31166891, + "L1I cache misses": 291301, + "L1D cache misses": 477186, + "L1D cache accesses": 32682323, + "LL cache misses": 538552, + "LL cache accesses": 552543, + "L2D TLB accesses": 202351, + "L2D TLB misses": 24178, + "L2D cache misses": 298051, + "L2D cache accesses": 1775481, + "instructions per cycle": 1.1169590185186398, + "branch miss rate": 0.01658033824205888, + "ITLB miss rate": 0.00023807640272132053, + "DTLB miss rate": 0.00048811093471722044, + "L2D TLB miss rate": 0.11948544855226809, + "L1I cache miss rate": 0.00934648887500521, + "L1D cache miss rate": 0.014600736918241704, + "L2D cache miss rate": 0.1678705657790762, + "LL cache miss rate": 0.9746788937693537, + "size": 2193517225 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 16.44, + 16.44, + 16.44, + 16.84, + 16.72, + 16.6, + 16.72, + 16.84, + 16.68, + 16.84 + ], + "matrix": "p2p-Gnutella30", + "shape": [ + 36682, + 36682 + ], + "nnz": 88328, + "% density": 6.564359899804003e-05, + "time_s": 2.896674871444702, + "power": [ + 56.32, + 68.24, + 71.76, + 59.48, + 47.6, + 48.76, + 52.6 + ], + "power_after": [ + 16.92, + 17.0, + 16.96, + 16.8, + 16.48, + 16.52, + 16.52, + 16.52, + 16.24, + 16.36 + ], + "task clock (msec)": 56.47, + "page faults": 3222, + "cycles": 69105836, + "instructions": 89065155, + "branch mispredictions": 333669, + "branches": 20078755, + "ITLB accesses": 26015038, + "ITLB misses": 5212, + "DTLB misses": 17039, + "DTLB accesses": 35296010, + "L1I cache accesses": 31837486, + "L1I cache misses": 293353, + "L1D cache misses": 462358, + "L1D cache accesses": 33478540, + "LL cache misses": 546516, + "LL cache accesses": 559865, + "L2D TLB accesses": 190400, + "L2D TLB misses": 23787, + "L2D cache misses": 307032, + "L2D cache accesses": 1768186, + "instructions per cycle": 1.288822480926213, + "branch miss rate": 0.016618012421586895, + "ITLB miss rate": 0.00020034566161310238, + "DTLB miss rate": 0.00048274578344691083, + "L2D TLB miss rate": 0.12493172268907562, + "L1I cache miss rate": 0.009214075508348869, + "L1D cache miss rate": 0.013810578358554464, + "L2D cache miss rate": 0.17364236567872385, + "LL cache miss rate": 0.9761567520741607, + "size": 1345569124 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 32.08, + 31.8, + 28.68, + 27.6, + 22.96, + 22.08, + 21.0, + 20.84, + 20.68, + 20.72 + ], + "matrix": "va2010", + "shape": [ + 285762, + 285762 + ], + "nnz": 1402128, + "% density": 1.717033263003816e-05, + "time_s": 14.632386922836304, + "power": [ + 85.16, + 83.48, + 76.96, + 67.44, + 54.04, + 51.4, + 54.24, + 66.76, + 83.2, + 96.44, + 96.44, + 95.84, + 94.24, + 92.36, + 91.2, + 89.32, + 87.48, + 88.68, + 88.24 + ], + "power_after": [ + 21.12, + 21.0, + 21.16, + 21.4, + 21.32, + 21.36, + 21.36, + 21.12, + 20.76, + 20.84 + ], + "task clock (msec)": 57.32, + "page faults": 3280, + "cycles": 39497791, + "instructions": 64385555, + "branch mispredictions": 332792, + "branches": 19983954, + "ITLB accesses": 27156853, + "ITLB misses": 6466, + "DTLB misses": 18244, + "DTLB accesses": 36466301, + "L1I cache accesses": 30929971, + "L1I cache misses": 291811, + "L1D cache misses": 473063, + "L1D cache accesses": 32462905, + "LL cache misses": 544953, + "LL cache accesses": 565172, + "L2D TLB accesses": 183225, + "L2D TLB misses": 23924, + "L2D cache misses": 301362, + "L2D cache accesses": 1756590, + "instructions per cycle": 1.6301052127193645, + "branch miss rate": 0.01665296067034582, + "ITLB miss rate": 0.00023809828038616994, + "DTLB miss rate": 0.000500297521264907, + "L2D TLB miss rate": 0.13057170145995362, + "L1I cache miss rate": 0.009434570759862659, + "L1D cache miss rate": 0.014572417348354991, + "L2D cache miss rate": 0.17156080815671274, + "LL cache miss rate": 0.964225050073252, + "size": 81659920644 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 16.6, + 16.64, + 17.04, + 17.08, + 16.92, + 17.24, + 16.88, + 16.36, + 16.4, + 16.4 + ], + "matrix": "ri2010", + "shape": [ + 25181, + 25181 + ], + "nnz": 125750, + "% density": 0.00019831796057928155, + "time_s": 2.970583200454712, + "power": [ + 23.04, + 23.28, + 23.76, + 24.12, + 21.4, + 26.28, + 26.36 + ], + "power_after": [ + 16.16, + 16.16, + 16.52, + 16.48, + 16.52, + 16.44, + 16.36, + 16.48, + 16.76, + 16.6 + ], + "task clock (msec)": 52.61, + "page faults": 3292, + "cycles": 42915672, + "instructions": 71002596, + "branch mispredictions": 344300, + "branches": 20224759, + "ITLB accesses": 26039851, + "ITLB misses": 5035, + "DTLB misses": 16402, + "DTLB accesses": 34820806, + "L1I cache accesses": 31878105, + "L1I cache misses": 299057, + "L1D cache misses": 471869, + "L1D cache accesses": 33450518, + "LL cache misses": 530093, + "LL cache accesses": 551126, + "L2D TLB accesses": 188315, + "L2D TLB misses": 22856, + "L2D cache misses": 299885, + "L2D cache accesses": 1763155, + "instructions per cycle": 1.6544677664607, + "branch miss rate": 0.01702368863826758, + "ITLB miss rate": 0.00019335748119296073, + "DTLB miss rate": 0.0004710402165877493, + "L2D TLB miss rate": 0.12137110692191275, + "L1I cache miss rate": 0.009381266546427399, + "L1D cache miss rate": 0.014106478111938357, + "L2D cache miss rate": 0.1700843090936418, + "LL cache miss rate": 0.9618363132931489, + "size": 634082761 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 30.08, + 25.12, + 24.68, + 23.68, + 22.84, + 21.96, + 21.08, + 20.96, + 20.8, + 20.96 + ], + "matrix": "p2p-Gnutella04", + "shape": [ + 10879, + 10879 + ], + "nnz": 39994, + "% density": 0.0003379223282393842, + "time_s": 0.9992897510528564, + "power": [ + 29.48, + 30.52, + 31.88, + 31.24, + 34.32 + ], + "power_after": [ + 20.4, + 20.6, + 20.64, + 20.76, + 20.92, + 20.84, + 20.88, + 20.88, + 20.88, + 20.84 + ], + "task clock (msec)": 52.68, + "page faults": 3272, + "cycles": 63019732, + "instructions": 73518898, + "branch mispredictions": 333423, + "branches": 19435905, + "ITLB accesses": 27447537, + "ITLB misses": 6417, + "DTLB misses": 18300, + "DTLB accesses": 37569384, + "L1I cache accesses": 30830481, + "L1I cache misses": 290545, + "L1D cache misses": 473875, + "L1D cache accesses": 32284772, + "LL cache misses": 529403, + "LL cache accesses": 549794, + "L2D TLB accesses": 198306, + "L2D TLB misses": 24497, + "L2D cache misses": 298519, + "L2D cache accesses": 1772795, + "instructions per cycle": 1.1666012480027683, + "branch miss rate": 0.017155002558409294, + "ITLB miss rate": 0.00023379146915805232, + "DTLB miss rate": 0.000487098750408045, + "L2D TLB miss rate": 0.12353131019737174, + "L1I cache miss rate": 0.009423952873132274, + "L1D cache miss rate": 0.014677972636758903, + "L2D cache miss rate": 0.16838890001381998, + "LL cache miss rate": 0.9629115632400499, + "size": 118352641 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 22.08, + 21.88, + 21.88, + 21.88, + 21.56, + 21.64, + 21.84, + 21.88, + 21.72, + 21.92 + ], + "matrix": "de2010", + "shape": [ + 24115, + 24115 + ], + "nnz": 116056, + "% density": 0.0001995689928120616, + "time_s": 2.7533018589019775, + "power": [ + 29.48, + 30.24, + 27.96, + 28.4, + 26.84, + 30.6, + 30.92 + ], + "power_after": [ + 20.84, + 21.24, + 21.2, + 21.24, + 21.28, + 20.88, + 20.68, + 20.56, + 20.52, + 20.56 + ], + "task clock (msec)": 61.38, + "page faults": 3315, + "cycles": 65013274, + "instructions": 87442627, + "branch mispredictions": 328392, + "branches": 19496396, + "ITLB accesses": 28311619, + "ITLB misses": 6963, + "DTLB misses": 17888, + "DTLB accesses": 38223408, + "L1I cache accesses": 30063404, + "L1I cache misses": 272797, + "L1D cache misses": 468341, + "L1D cache accesses": 31519623, + "LL cache misses": 538689, + "LL cache accesses": 552789, + "L2D TLB accesses": 192995, + "L2D TLB misses": 23339, + "L2D cache misses": 300578, + "L2D cache accesses": 1764035, + "instructions per cycle": 1.344996515634638, + "branch miss rate": 0.016843728451145536, + "ITLB miss rate": 0.0002459414277933028, + "DTLB miss rate": 0.00046798548156668814, + "L2D TLB miss rate": 0.12093059405684085, + "L1I cache miss rate": 0.009074055619250568, + "L1D cache miss rate": 0.01485871198395996, + "L2D cache miss rate": 0.17039231081015965, + "LL cache miss rate": 0.9744929801425137, + "size": 581533225 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 51.04, + 38.64, + 22.84, + 22.24, + 21.88, + 21.88, + 21.6, + 21.4, + 21.24, + 21.28 + ], + "matrix": "ga2010", + "shape": [ + 291086, + 291086 + ], + "nnz": 1418056, + "% density": 1.6735964475229304e-05, + "time_s": 15.249999523162842, + "power": [ + 88.88, + 89.52, + 78.6, + 64.88, + 52.64, + 52.64, + 54.76, + 60.16, + 71.44, + 86.84, + 90.72, + 89.6, + 90.56, + 90.36, + 91.68, + 91.84, + 93.4, + 93.4, + 92.72 + ], + "power_after": [ + 21.68, + 21.4, + 21.28, + 21.04, + 21.04, + 20.96, + 20.92, + 20.76, + 20.8, + 20.96 + ], + "task clock (msec)": 72.45, + "page faults": 3289, + "cycles": 24836161, + "instructions": 74134706, + "branch mispredictions": 325643, + "branches": 19697746, + "ITLB accesses": 27767290, + "ITLB misses": 5832, + "DTLB misses": 18134, + "DTLB accesses": 37063060, + "L1I cache accesses": 32135376, + "L1I cache misses": 302429, + "L1D cache misses": 484427, + "L1D cache accesses": 33639686, + "LL cache misses": 548380, + "LL cache accesses": 561312, + "L2D TLB accesses": 186006, + "L2D TLB misses": 25022, + "L2D cache misses": 304539, + "L2D cache accesses": 1750107, + "instructions per cycle": 2.9849502908279586, + "branch miss rate": 0.01653199305138771, + "ITLB miss rate": 0.00021003129941740803, + "DTLB miss rate": 0.0004892742261432272, + "L2D TLB miss rate": 0.13452254228358226, + "L1I cache miss rate": 0.009411092622659838, + "L1D cache miss rate": 0.014400461407398393, + "L2D cache miss rate": 0.17401164614506429, + "LL cache miss rate": 0.976961119662505, + "size": 84731059396 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 22.04, + 21.32, + 21.32, + 21.32, + 21.12, + 21.12, + 21.0, + 20.68, + 20.72, + 20.56 + ], + "matrix": "mac_econ_fwd500", + "shape": [ + 206500, + 206500 + ], + "nnz": 1273389, + "% density": 2.9862143765866013e-05, + "time_s": 15.046087741851807, + "power": [ + 91.88, + 91.12, + 83.92, + 72.88, + 57.76, + 51.24, + 53.12, + 62.84, + 78.32, + 91.64, + 95.8, + 95.8, + 94.08, + 92.48, + 91.6, + 89.88, + 87.36, + 87.84, + 87.32 + ], + "power_after": [ + 20.92, + 21.04, + 21.12, + 20.92, + 20.92, + 20.88, + 20.88, + 20.92, + 21.04, + 20.96 + ], + "task clock (msec)": 62.46, + "page faults": 3243, + "cycles": 57150420, + "instructions": 94155455, + "branch mispredictions": 320781, + "branches": 19491698, + "ITLB accesses": 27433101, + "ITLB misses": 7382, + "DTLB misses": 19213, + "DTLB accesses": 37123052, + "L1I cache accesses": 32027284, + "L1I cache misses": 290368, + "L1D cache misses": 471338, + "L1D cache accesses": 33366668, + "LL cache misses": 571063, + "LL cache accesses": 583554, + "L2D TLB accesses": 196434, + "L2D TLB misses": 25171, + "L2D cache misses": 329198, + "L2D cache accesses": 1814040, + "instructions per cycle": 1.6475024155553013, + "branch miss rate": 0.016457314288370363, + "ITLB miss rate": 0.0002690909788142434, + "DTLB miss rate": 0.0005175490420345827, + "L2D TLB miss rate": 0.1281397314110592, + "L1I cache miss rate": 0.009066269871650684, + "L1D cache miss rate": 0.014126013421537926, + "L2D cache miss rate": 0.1814722938854711, + "LL cache miss rate": 0.9785949543658342, + "size": 42642250000 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 94.16, + 91.68, + 78.92, + 60.88, + 46.72, + 28.36, + 22.08, + 21.64, + 21.64, + 21.64 + ], + "matrix": "mc2depi", + "shape": [ + 525825, + 525825 + ], + "nnz": 2100225, + "% density": 7.595972132902821e-06, + "time_s": 11.03979206085205, + "power": [ + 95.44, + 94.0, + 88.76, + 72.12, + 59.48, + 51.92, + 53.88, + 68.6, + 83.2, + 97.76, + 98.4, + 97.12, + 97.12, + 95.28, + 94.12 + ], + "power_after": [ + 21.48, + 21.44, + 21.28, + 21.24, + 21.16, + 21.08, + 21.24, + 21.24, + 21.24, + 21.16 + ], + "task clock (msec)": 56.14, + "page faults": 3289, + "cycles": 47515158, + "instructions": 72388154, + "branch mispredictions": 327042, + "branches": 19309026, + "ITLB accesses": 26093030, + "ITLB misses": 6189, + "DTLB misses": 17253, + "DTLB accesses": 35168741, + "L1I cache accesses": 30539322, + "L1I cache misses": 285404, + "L1D cache misses": 465747, + "L1D cache accesses": 31932803, + "LL cache misses": 530261, + "LL cache accesses": 551030, + "L2D TLB accesses": 183570, + "L2D TLB misses": 23883, + "L2D cache misses": 297006, + "L2D cache accesses": 1721848, + "instructions per cycle": 1.5234749719236964, + "branch miss rate": 0.01693726032581861, + "ITLB miss rate": 0.0002371897782664566, + "DTLB miss rate": 0.0004905776979619486, + "L2D TLB miss rate": 0.13010295799967314, + "L1I cache miss rate": 0.009345459601231487, + "L1D cache miss rate": 0.014585221347465175, + "L2D cache miss rate": 0.1724925777420539, + "LL cache miss rate": 0.9623087672177558, + "size": 276491930625 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 30.72, + 30.6, + 28.68, + 26.48, + 22.44, + 21.4, + 21.28, + 21.08, + 21.32, + 21.6 + ], + "matrix": "p2p-Gnutella24", + "shape": [ + 26518, + 26518 + ], + "nnz": 65369, + "% density": 9.295875717624285e-05, + "time_s": 1.718301773071289, + "power": [ + 31.52, + 32.48, + 33.64, + 33.88, + 33.44, + 31.52 + ], + "power_after": [ + 20.96, + 20.84, + 20.92, + 20.8, + 20.76, + 20.76, + 20.76, + 20.68, + 20.72, + 20.92 + ], + "task clock (msec)": 67.08, + "page faults": 3303, + "cycles": 61261862, + "instructions": 83757591, + "branch mispredictions": 329248, + "branches": 19953212, + "ITLB accesses": 27084694, + "ITLB misses": 7107, + "DTLB misses": 17529, + "DTLB accesses": 36684333, + "L1I cache accesses": 32158234, + "L1I cache misses": 286484, + "L1D cache misses": 474161, + "L1D cache accesses": 33730073, + "LL cache misses": 550064, + "LL cache accesses": 565245, + "L2D TLB accesses": 191046, + "L2D TLB misses": 23775, + "L2D cache misses": 307419, + "L2D cache accesses": 1772169, + "instructions per cycle": 1.3672060930828385, + "branch miss rate": 0.016501002445120115, + "ITLB miss rate": 0.0002623991247602797, + "DTLB miss rate": 0.0004778334118818516, + "L2D TLB miss rate": 0.12444646838981188, + "L1I cache miss rate": 0.008908573773049851, + "L1D cache miss rate": 0.014057514788064645, + "L2D cache miss rate": 0.1734704760099065, + "LL cache miss rate": 0.973142619572044, + "size": 703204324 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 86.48, + 72.16, + 59.36, + 41.84, + 28.44, + 22.96, + 22.92, + 22.92, + 23.04, + 23.24 + ], + "matrix": "p2p-Gnutella25", + "shape": [ + 22687, + 22687 + ], + "nnz": 54705, + "% density": 0.00010628522108964806, + "time_s": 1.431199073791504, + "power": [ + 35.16, + 36.2, + 36.72, + 37.52, + 37.52 + ], + "power_after": [ + 21.32, + 21.2, + 21.2, + 21.28, + 21.52, + 21.44, + 21.92, + 21.68, + 21.6, + 21.36 + ], + "task clock (msec)": 59.85, + "page faults": 3318, + "cycles": 76505130, + "instructions": 72343215, + "branch mispredictions": 322338, + "branches": 19784096, + "ITLB accesses": 27270404, + "ITLB misses": 6607, + "DTLB misses": 17981, + "DTLB accesses": 36751047, + "L1I cache accesses": 30620441, + "L1I cache misses": 302139, + "L1D cache misses": 471011, + "L1D cache accesses": 32141810, + "LL cache misses": 531907, + "LL cache accesses": 545159, + "L2D TLB accesses": 188244, + "L2D TLB misses": 23034, + "L2D cache misses": 293848, + "L2D cache accesses": 1757551, + "instructions per cycle": 0.945599530384433, + "branch miss rate": 0.016292783860329025, + "ITLB miss rate": 0.00024227730546272803, + "DTLB miss rate": 0.0004892649725054092, + "L2D TLB miss rate": 0.12236246573595971, + "L1I cache miss rate": 0.009867232153841285, + "L1D cache miss rate": 0.014654152955294054, + "L2D cache miss rate": 0.1671917344077071, + "LL cache miss rate": 0.9756914955086498, + "size": 514699969 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 77.2, + 64.12, + 64.12, + 48.92, + 36.2, + 21.72, + 21.88, + 22.36, + 22.36, + 22.44 + ], + "matrix": "vt2010", + "shape": [ + 32580, + 32580 + ], + "nnz": 155598, + "% density": 0.00014658915806621921, + "time_s": 3.5892834663391113, + "power": [ + 33.44, + 30.68, + 31.08, + 26.96, + 26.88, + 32.48, + 32.24 + ], + "power_after": [ + 21.24, + 21.24, + 21.36, + 21.36, + 21.2, + 21.04, + 20.84, + 20.72, + 20.72, + 20.56 + ], + "task clock (msec)": 55.26, + "page faults": 3297, + "cycles": 49276491, + "instructions": 64763517, + "branch mispredictions": 340611, + "branches": 20355849, + "ITLB accesses": 27946393, + "ITLB misses": 6805, + "DTLB misses": 17877, + "DTLB accesses": 38226912, + "L1I cache accesses": 31946141, + "L1I cache misses": 295259, + "L1D cache misses": 468136, + "L1D cache accesses": 33395666, + "LL cache misses": 527109, + "LL cache accesses": 540409, + "L2D TLB accesses": 192519, + "L2D TLB misses": 24204, + "L2D cache misses": 290933, + "L2D cache accesses": 1743452, + "instructions per cycle": 1.3142883286880147, + "branch miss rate": 0.016732831924622747, + "ITLB miss rate": 0.00024350190738389746, + "DTLB miss rate": 0.0004676548291423592, + "L2D TLB miss rate": 0.1257226559456469, + "L1I cache miss rate": 0.009242399574959616, + "L1D cache miss rate": 0.014017866869311724, + "L2D cache miss rate": 0.16687181522634406, + "LL cache miss rate": 0.9753890109158063, + "size": 1061456400 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 93.52, + 87.76, + 71.6, + 58.32, + 39.6, + 26.24, + 26.24, + 22.16, + 22.24, + 22.24 + ], + "matrix": "ut2010", + "shape": [ + 115406, + 115406 + ], + "nnz": 572066, + "% density": 4.295259032005559e-05, + "time_s": 8.478580713272095, + "power": [ + 89.68, + 88.92, + 80.84, + 68.96, + 56.64, + 54.52, + 55.88, + 70.44, + 85.36, + 85.36, + 98.2, + 96.52 + ], + "power_after": [ + 21.24, + 21.32, + 21.16, + 21.44, + 21.68, + 21.76, + 21.72, + 22.0, + 21.72, + 21.72 + ], + "task clock (msec)": 53.84, + "page faults": 3291, + "cycles": 66389970, + "instructions": 74935543, + "branch mispredictions": 330515, + "branches": 19475058, + "ITLB accesses": 26125490, + "ITLB misses": 6431, + "DTLB misses": 13728, + "DTLB accesses": 35274185, + "L1I cache accesses": 30428652, + "L1I cache misses": 288897, + "L1D cache misses": 475615, + "L1D cache accesses": 31855716, + "LL cache misses": 553829, + "LL cache accesses": 574192, + "L2D TLB accesses": 181148, + "L2D TLB misses": 23202, + "L2D cache misses": 307806, + "L2D cache accesses": 1767037, + "instructions per cycle": 1.1287178319255153, + "branch miss rate": 0.016971194642911976, + "ITLB miss rate": 0.00024615806248992844, + "DTLB miss rate": 0.0003891797925309968, + "L2D TLB miss rate": 0.12808311435952924, + "L1I cache miss rate": 0.009494242465949527, + "L1D cache miss rate": 0.014930287550278261, + "L2D cache miss rate": 0.17419329646181717, + "LL cache miss rate": 0.9645362526820297, + "size": 13318544836 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 20.72, + 20.8, + 20.96, + 21.08, + 21.4, + 21.48, + 21.48, + 21.36, + 21.08, + 21.04 + ], + "matrix": "fl2010", + "shape": [ + 484481, + 484481 + ], + "nnz": 2346294, + "% density": 9.99606174861054e-06, + "time_s": 14.43001127243042, + "power": [ + 93.04, + 93.04, + 89.16, + 77.68, + 62.92, + 55.12, + 53.84, + 64.72, + 77.04, + 89.56, + 94.4, + 94.76, + 93.52, + 93.52, + 96.04, + 97.12, + 96.44, + 93.88, + 93.72 + ], + "power_after": [ + 21.08, + 21.28, + 21.28, + 21.36, + 21.08, + 21.24, + 21.08, + 20.8, + 21.04, + 20.88 + ], + "task clock (msec)": 61.6, + "page faults": 3276, + "cycles": 41408849, + "instructions": 49118917, + "branch mispredictions": 331330, + "branches": 19331189, + "ITLB accesses": 27367982, + "ITLB misses": 6160, + "DTLB misses": 17157, + "DTLB accesses": 36828216, + "L1I cache accesses": 30147304, + "L1I cache misses": 280082, + "L1D cache misses": 454022, + "L1D cache accesses": 31595140, + "LL cache misses": 536056, + "LL cache accesses": 550006, + "L2D TLB accesses": 185998, + "L2D TLB misses": 23735, + "L2D cache misses": 296648, + "L2D cache accesses": 1723525, + "instructions per cycle": 1.1861937287848787, + "branch miss rate": 0.017139659645353425, + "ITLB miss rate": 0.00022508053388810325, + "DTLB miss rate": 0.00046586562867992305, + "L2D TLB miss rate": 0.12760889902041958, + "L1I cache miss rate": 0.009290449321770198, + "L1D cache miss rate": 0.014369994878959232, + "L2D cache miss rate": 0.172117027603313, + "LL cache miss rate": 0.97463664032756, + "size": 234721839361 + }, + { + "architecture": "altra", + "iterations": 1000, + "baseline_time_s": 10, + "baseline_delay_s": 30, + "power_before": [ + 37.56, + 23.12, + 22.32, + 22.28, + 22.28, + 21.96, + 21.76, + 21.72, + 21.88, + 21.84 + ], + "matrix": "tn2010", + "shape": [ + 240116, + 240116 + ], + "nnz": 1193966, + "% density": 2.070855328296721e-05, + "time_s": 16.282614707946777, + "power": [ + 85.48, + 85.84, + 79.28, + 70.16, + 55.52, + 49.48, + 49.48, + 60.48, + 76.32, + 88.88, + 91.0, + 91.0, + 90.68, + 88.32, + 86.92, + 86.4, + 88.08, + 86.8, + 87.32, + 87.8 + ], + "power_after": [ + 21.68, + 21.48, + 21.44, + 21.36, + 21.52, + 21.4, + 21.4, + 21.32, + 21.2, + 21.04 + ], + "task clock (msec)": 68.11, + "page faults": 3486, + "cycles": 70427921, + "instructions": 85638293, + "branch mispredictions": 333780, + "branches": 19402540, + "ITLB accesses": 26935483, + "ITLB misses": 5639, + "DTLB misses": 16688, + "DTLB accesses": 36421540, + "L1I cache accesses": 33029213, + "L1I cache misses": 302558, + "L1D cache misses": 481598, + "L1D cache accesses": 34668833, + "LL cache misses": 551659, + "LL cache accesses": 564579, + "L2D TLB accesses": 188346, + "L2D TLB misses": 24479, + "L2D cache misses": 311796, + "L2D cache accesses": 1767924, + "instructions per cycle": 1.215970765344614, + "branch miss rate": 0.017202902300420462, + "ITLB miss rate": 0.0002093521025778524, + "DTLB miss rate": 0.00045819040051573877, + "L2D TLB miss rate": 0.12996824992301403, + "L1I cache miss rate": 0.00916031514284037, + "L1D cache miss rate": 0.013891381922200843, + "L2D cache miss rate": 0.17636278482559206, + "LL cache miss rate": 0.9771156915152707, + "size": 57655693456 + } +] \ No newline at end of file