Commit bd160bde2cb92f5a0b1ec43b698e977791bd7d21
1 parent
38e0d558
Added script to plot tuning comparison
Showing
1 changed file
with
161 additions
and
0 deletions
figures/tuning_comparison.py
0 → 100644
| 1 | +import os,sys,argparse | ||
| 2 | +import json | ||
| 3 | +import itertools | ||
| 4 | +from collections import namedtuple | ||
| 5 | +from matplotlib import pyplot as plt | ||
| 6 | +from common import * | ||
| 7 | + | ||
| 8 | +def within_threshold(value, base, threshold): | ||
| 9 | + bmax = base * ((100 + threshold)/100.0) | ||
| 10 | + if value <= bmax: | ||
| 11 | + return True | ||
| 12 | + else: | ||
| 13 | + return False | ||
| 14 | + | ||
| 15 | +def find_optimal(data_points_list, inefficiency, ineff_threshold, cluster_threshold): | ||
| 16 | + frontier = [] | ||
| 17 | + inefficiencies = [] | ||
| 18 | + energies = [] | ||
| 19 | + | ||
| 20 | + # Find emin | ||
| 21 | + emin = min(data_points_list, key=lambda x: x['energy'])['energy'] | ||
| 22 | + # Check | ||
| 23 | + for point in data_points_list: | ||
| 24 | + assert float(point['energy']) >= emin | ||
| 25 | + | ||
| 26 | + # Find all inefficiencies | ||
| 27 | + for point in data_points_list: | ||
| 28 | + inefficiencies.append(float(point['energy']) / float(emin)) | ||
| 29 | + | ||
| 30 | + # Find the frontier | ||
| 31 | + for idx, ineff in enumerate(inefficiencies): | ||
| 32 | + if within_threshold(ineff, inefficiency, ineff_threshold): | ||
| 33 | + frontier.append(data_points_list[idx]) | ||
| 34 | + | ||
| 35 | + # Find the optimal point - point with best performance | ||
| 36 | + optimal_point = min(frontier, key=lambda x: x['performance']) | ||
| 37 | + # Check | ||
| 38 | + for point in frontier: | ||
| 39 | + assert point['performance'] >= optimal_point['performance'] | ||
| 40 | + | ||
| 41 | + return optimal_point, optimal_point['performance'], optimal_point['energy'] | ||
| 42 | + | ||
| 43 | + | ||
| 44 | + | ||
| 45 | + | ||
| 46 | +def no_tuning(args, bmark, inefficiency, ineff_threshold=0.0, cluster_threshold=0.0): | ||
| 47 | + file_path = os.path.join(os.path.join(os.path.join(args.input_dir, "aggr_data"), bmark), 'frontiers.json') | ||
| 48 | + data = json.loads(open(file_path).read()) | ||
| 49 | + | ||
| 50 | + | ||
| 51 | + | ||
| 52 | + nt_point, performance, energy = find_optimal(data['data'], inefficiency, ineff_threshold, cluster_threshold) | ||
| 53 | + | ||
| 54 | + # Get the simulation corresponding to the optimal | ||
| 55 | + file_path = os.path.join(os.path.join(os.path.join(args.input_dir, "per_sample_data"), bmark), 'per_sample_frontiers.json') | ||
| 56 | + data = json.loads(open(file_path).read()) | ||
| 57 | + no_tuning_data = [] | ||
| 58 | + for point in data['data']: | ||
| 59 | + for sample in point: | ||
| 60 | + if sample['cpu_freq'] == nt_point['cpu_freq'] and sample['mem_freq'] == nt_point['mem_freq']: | ||
| 61 | + no_tuning_data.append(sample) | ||
| 62 | + assert len(no_tuning_data) == len(data['data']) | ||
| 63 | + | ||
| 64 | + return nt_point, no_tuning_data, performance, energy | ||
| 65 | + | ||
| 66 | + | ||
| 67 | + | ||
| 68 | + | ||
| 69 | +def with_tuning(args, bmark, inefficiency, ineff_threshold=0.0, cluster_threshold=0.0): | ||
| 70 | + file_path = os.path.join(os.path.join(os.path.join(args.input_dir, "per_sample_data"), bmark), 'per_sample_frontiers.json') | ||
| 71 | + data = json.loads(open(file_path).read()) | ||
| 72 | + | ||
| 73 | + optimal_points = [] | ||
| 74 | + performance = 0 | ||
| 75 | + energy = 0 | ||
| 76 | + # For each sample | ||
| 77 | + for point in data['data']: | ||
| 78 | + assert len(point) == len(data['data'][0]) # 496 | ||
| 79 | + | ||
| 80 | + optimal_point, performance, energy = find_optimal(point, inefficiency, ineff_threshold, cluster_threshold) | ||
| 81 | + optimal_points.append(optimal_point) | ||
| 82 | + | ||
| 83 | + return optimal_points, sum(x['performance'] for x in optimal_points), sum(x['energy'] for x in optimal_points) | ||
| 84 | + | ||
| 85 | + | ||
| 86 | + | ||
| 87 | + | ||
| 88 | +def plot(args, bmark, inefficiency, nt_points_list, t_points_list): | ||
| 89 | + fig, axes_array = plt.subplots(3, sharex=True) | ||
| 90 | + x_axis = range(len(t_points_list[0])) | ||
| 91 | + x_ticklabels = [str(x) for x in x_axis] | ||
| 92 | + | ||
| 93 | + ax = axes_array[0] | ||
| 94 | + ax.set_ylabel('Inefficiency') | ||
| 95 | + | ||
| 96 | + for nt_points, t_points in itertools.izip(nt_points_list, t_points_list): | ||
| 97 | + nt_ineff = [x['inefficiency'] for x in nt_points] | ||
| 98 | + t_ineff = [x['inefficiency'] for x in t_points] | ||
| 99 | + | ||
| 100 | + ax.plot(x_axis, nt_ineff, 'k') | ||
| 101 | + ax.plot(x_axis, t_ineff) | ||
| 102 | + | ||
| 103 | + | ||
| 104 | + ax = axes_array[1] | ||
| 105 | + ax.set_ylabel('Performance (ms)') | ||
| 106 | + | ||
| 107 | + for nt_points, t_points in itertools.izip(nt_points_list, t_points_list): | ||
| 108 | + nt_perf = [x['performance']/1e6 for x in nt_points] | ||
| 109 | + t_perf = [x['performance']/1e6 for x in t_points] | ||
| 110 | + | ||
| 111 | + ax.plot(x_axis, nt_perf, 'k') | ||
| 112 | + ax.plot(x_axis, t_perf) | ||
| 113 | + | ||
| 114 | + | ||
| 115 | + ax = axes_array[2] | ||
| 116 | + cpi = get_cpi(bmark, args.input_dir) | ||
| 117 | + | ||
| 118 | + ax.set_xlabel('Instructions (x10 Million)') | ||
| 119 | + ax.set_ylabel('CPI') | ||
| 120 | + | ||
| 121 | + ax.plot(x_axis, cpi) | ||
| 122 | + | ||
| 123 | + plt.tight_layout() | ||
| 124 | + fig.subplots_adjust(hspace=0.03) | ||
| 125 | + | ||
| 126 | + for ax in axes_array: | ||
| 127 | + ax.grid(True) | ||
| 128 | + ax.set_axisbelow(True) | ||
| 129 | + | ||
| 130 | + out_fname = os.path.join(os.path.join(os.path.join(args.output_dir, 'tuning_comparison'), bmark), '%.1f.jpg' % (inefficiency)) | ||
| 131 | + if not os.path.exists(os.path.dirname(out_fname)): | ||
| 132 | + os.makedirs(os.path.dirname(out_fname)) | ||
| 133 | + plt.savefig(out_fname, dpi=300) | ||
| 134 | + | ||
| 135 | +def main(argv): | ||
| 136 | + args = parse(argv) | ||
| 137 | + | ||
| 138 | + bmarks, labels = get_benchmarks(args) | ||
| 139 | + | ||
| 140 | + for bmark in bmarks: | ||
| 141 | + print 'BENCHMARK: %s' % (bmark) | ||
| 142 | + nt_points_list = [] | ||
| 143 | + t_points_list = [] | ||
| 144 | + for ineff in args.inefficiency: | ||
| 145 | + string = '\tinefficiency: %0.2f\n' % (ineff) | ||
| 146 | + | ||
| 147 | + nt_point, data, nt_performance, nt_energy = no_tuning(args, bmark, ineff, 3) | ||
| 148 | + string += '\t\tNO TUNING: %0.2fms %0.2fmJ\n' % (nt_performance / 1e6, nt_energy / 1e6) | ||
| 149 | + | ||
| 150 | + t_points, t_performance, t_energy = with_tuning(args, bmark, ineff, 3) | ||
| 151 | + string += '\t\tW. TUNING: %0.2fms %0.2fmJ\n' % (t_performance / 1e6, t_energy / 1e6) | ||
| 152 | + | ||
| 153 | + nt_points_list.append(data) | ||
| 154 | + t_points_list.append(t_points) | ||
| 155 | + print string | ||
| 156 | + plot(args, bmark, ineff, [data], [t_points]) | ||
| 157 | + # Uncomment this line if you want an aggregate plot per benchmark (Not very clean) | ||
| 158 | + #plot(args, bmark, nt_points_list, t_points_list) | ||
| 159 | + | ||
| 160 | +if __name__ == '__main__': | ||
| 161 | + main(sys.argv) |