# SPDX-FileCopyrightText: 2021 Guillaume DIDIER # # SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: MIT import os import sys import argparse import warnings import itertools from multiprocessing import Pool import numpy as np import pandas as pd import seaborn as sns from scipy import optimize import matplotlib.pyplot as plt import matplotlib.style as mplstyle mplstyle.use("fast") warnings.filterwarnings("ignore") print("warnings are filtered, enable them back if you are having some trouble") # TODO # args.path should be the root # with root-result_lite.csv.bz2 the result # and .stats.csv # root.slices a slice mapping - done # root.cores a core + socket mapping - done -> move to analyse csv ? # # Facet plot with actual dot cloud + plot the linear regression # each row is a slice # each row is an origin core # each column a helper core if applicable parser = argparse.ArgumentParser( prog=sys.argv[0], ) parser.add_argument("path", help="Path to the experiment files") parser.add_argument( "--no-plot", dest="no_plot", action="store_true", default=False, help="No visible plot (save figures to files)", ) parser.add_argument( "--rslice", dest="rslice", action="store_true", default=False, help="Create slice{} directories with segmented grid", ) parser.add_argument( "--no-slice-remap", dest="slice_remap", action="store_false", default=True, help="Don't remap the slices" ) args = parser.parse_args() img_dir = os.path.dirname(args.path) + "/figs/" os.makedirs(img_dir, exist_ok=True) assert os.path.exists(args.path + ".stats.csv") assert os.path.exists(args.path + ".cores.csv") if args.slice_remap: assert os.path.exists(args.path + ".slices.csv") stats = pd.read_csv( args.path + ".stats.csv", dtype={ "main_core": np.int8, "helper_core": np.int8, # "address": int, "hash": np.int8, # "time": np.int16, "clflush_remote_hit": np.float64, "clflush_shared_hit": np.float64, # "clflush_miss_f": np.int32, # "clflush_local_hit_f": np.int32, "clflush_miss_n": np.float64, "clflush_local_hit_n": np.float64, # "reload_miss": np.int32, # "reload_remote_hit": np.int32, # "reload_shared_hit": np.int32, # "reload_local_hit": np.int32 }, ) if args.slice_remap: slice_mapping = pd.read_csv(args.path + ".slices.csv") core_mapping = pd.read_csv(args.path + ".cores.csv") # print("core mapping:\n", core_mapping.to_string()) # print("slice mapping:\n", slice_mapping.to_string()) # print("core {} is mapped to '{}'".format(4, repr(core_mapping.iloc[4]))) min_time_miss = stats["clflush_miss_n"].min() max_time_miss = stats["clflush_miss_n"].max() def remap_core(key): def remap(core): remapped = core_mapping.iloc[core] return remapped[key] return remap def plot(filename, g=None): if args.no_plot: if g is not None: g.savefig(img_dir + filename) else: plt.savefig(img_dir + filename) # tikzplotlib.save( # img_dir+filename+".tex", # axis_width=r'0.175\textwidth', # axis_height=r'0.25\textwidth' # ) print(filename, "saved") plt.close() plt.show() stats["main_socket"] = stats["main_core"].apply(remap_core("socket")) stats["main_core_fixed"] = stats["main_core"].apply(remap_core("core")) stats["main_ht"] = stats["main_core"].apply(remap_core("hthread")) stats["helper_socket"] = stats["helper_core"].apply(remap_core("socket")) stats["helper_core_fixed"] = stats["helper_core"].apply(remap_core("core")) stats["helper_ht"] = stats["helper_core"].apply(remap_core("hthread")) # slice_mapping = {3: 0, 1: 1, 2: 2, 0: 3} if args.slice_remap: stats["slice_group"] = stats["hash"].apply( lambda h: slice_mapping["slice_group"].iloc[h] ) else: stats["slice_group"] = stats["hash"] graph_lower_miss = int((min_time_miss // 10) * 10) graph_upper_miss = int(((max_time_miss + 9) // 10) * 10) # print("Graphing from {} to {}".format(graph_lower_miss, graph_upper_miss)) # also explains remote # shared needs some thinking as there is something weird happening there. # # M 0 1 2 3 4 5 6 7 # # print(stats.head()) num_core = len(stats["main_core_fixed"].unique()) # print("Found {}".format(num_core)) def miss_topology(main_core_fixed, slice_group, C, h): return C + h * abs(main_core_fixed - slice_group) + h * abs(slice_group + 1) def miss_topology_df(x, C, h): func = lambda x, C, h: miss_topology(x["main_core_fixed"], x["slice_group"], C, h) return x.apply(func, args=(C, h), axis=1) memory = -1 gpu_if_any = num_core def exclusive_hit_topology_gpu(main_core, slice_group, helper_core, C, h1, h2): round_trip = gpu_if_any - memory if slice_group <= num_core / 2: # send message towards higher cores first if helper_core < slice_group: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(round_trip - (helper_core - memory)) ) else: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - slice_group) ) else: # send message toward lower cores first if helper_core > slice_group: r = C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - memory) else: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - slice_group) ) return r def exclusive_hit_topology_gpu_df(x, C, h1, h2): def func(x, C, h1, h2): return exclusive_hit_topology_gpu( x["main_core_fixed"], x["slice_group"], x["helper_core_fixed"], C, h1, h2 ) return x.apply(func, args=(C, h1, h2), axis=1) def exclusive_hit_topology_gpu2(main_core, slice_group, helper_core, C, h1, h2): round_trip = gpu_if_any + 1 - memory if slice_group <= num_core / 2: # send message towards higher cores first if helper_core < slice_group: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(round_trip - (helper_core - memory)) ) else: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - slice_group) ) else: # send message toward lower cores first if helper_core > slice_group: r = C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - memory) else: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - slice_group) ) return r def exclusive_hit_topology_gpu2_df(x, C, h1, h2): def func(x, C, h1, h2): return exclusive_hit_topology_gpu2( x["main_core_fixed"], x["slice_group"], x["helper_core_fixed"], C, h1, h2 ) return x.apply(func, args=(C, h1, h2), axis=1) # unlikely def exclusive_hit_topology_nogpu(main_core, slice_group, helper_core, C, h1, h2): round_trip = (num_core - 1) - memory if slice_group <= num_core / 2: # send message towards higher cores first if helper_core < slice_group: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(round_trip - (helper_core - memory)) ) else: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - slice_group) ) else: # send message toward lower cores first if helper_core > slice_group: r = C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - memory) else: r = ( C + h1 * abs(main_core - slice_group) + h2 * abs(helper_core - slice_group) ) return r def exclusive_hit_topology_nogpu_df(x, C, h1, h2): def func(x, C, h1, h2): return exclusive_hit_topology_nogpu( x["main_core_fixed"], x["slice_group"], x["helper_core_fixed"], C, h1, h2 ) return x.apply(func, args=(C, h1, h2), axis=1) def remote_hit_topology_2(x, C, h): main_core = x["main_core_fixed"] slice_group = x["slice_group"] helper_core = x["helper_core_fixed"] return ( C + h * abs(main_core - slice_group) + h * abs(slice_group - helper_core) + h * abs(helper_core - main_core) ) def shared_hit_topology_1(x, C, h): main_core = x["main_core_fixed"] slice_group = x["slice_group"] helper_core = x["helper_core_fixed"] return ( C + h * abs(main_core - slice_group) + h * max(abs(slice_group - main_core), abs(slice_group - helper_core)) ) def do_predictions(df): res_miss = optimize.curve_fit( miss_topology_df, df[["main_core_fixed", "slice_group"]], df["clflush_miss_n"] ) # print("Miss topology:") # print(res_miss) res_gpu = optimize.curve_fit( exclusive_hit_topology_gpu_df, df[["main_core_fixed", "slice_group", "helper_core_fixed"]], df["clflush_remote_hit"], ) # print("Exclusive hit topology (GPU):") # print(res_gpu) # res_gpu2 = optimize.curve_fit( # exclusive_hit_topology_gpu2_df, # df[["main_core_fixed", "slice_group", "helper_core_fixed"]], # df["clflush_remote_hit"] # ) # print("Exclusive hit topology (GPU2):") # print(res_gpu2) # res_no_gpu = optimize.curve_fit( # exclusive_hit_topology_nogpu_df, # df[["main_core_fixed", "slice_group", "helper_core_fixed"]], # df["clflush_remote_hit"] # ) # print("Exclusive hit topology (No GPU):") # print(res_no_gpu) df["predicted_miss"] = miss_topology_df(df, *(res_miss[0])) # df["predicted_remote_hit_no_gpu"] = exclusive_hit_topology_nogpu_df(df, *(res_no_gpu[0])) df["predicted_remote_hit_gpu"] = exclusive_hit_topology_gpu_df(df, *(res_gpu[0])) # df["predicted_remote_hit_gpu2"] = exclusive_hit_topology_gpu_df(df, *(res_gpu2[0])) df_A0 = df[df["main_core_fixed"] == 0] figure_A0 = sns.FacetGrid(df_A0, col="slice_group") figure_A0.map(sns.scatterplot, "helper_core_fixed", "clflush_remote_hit", color="r") figure_A0.map( sns.lineplot, "helper_core_fixed", "predicted_remote_hit_gpu", color="r" ) figure_A0.set_titles(col_template="$S$ = {col_name}") plot("medians_remote_hit.png") g2 = sns.FacetGrid(df, row="main_core_fixed", col="slice_group") g2.map(sns.scatterplot, "helper_core_fixed", "clflush_remote_hit", color="r") g2.map(sns.lineplot, "helper_core_fixed", "predicted_remote_hit_gpu", color="r") # g2.map(sns.lineplot, 'helper_core_fixed', 'predicted_remote_hit_gpu2', color="g") # g2.map(sns.lineplot, 'helper_core_fixed', 'predicted_remote_hit_no_gpu', color="g") plot("medians_remote_hit_grid.png", g=g2) def rslice(): for core in stats["main_core_fixed"].unique(): os.makedirs(img_dir + f"slices{core}", exist_ok=True) for slice_ in stats["slice_group"].unique(): df = stats[ (stats["slice_group"] == slice_) & (stats["main_core_fixed"] == core) ] fig = sns.scatterplot( df, x="helper_core_fixed", y="clflush_remote_hit", color="r" ) fig.set(title=f"main_core={core} slice={slice_}") plt.savefig(img_dir + f"slices{core}/" + str(slice_) + ".png") plt.close() def facet_grid( df, row, col, third, draw_fn=sns.scatterplot, shown=[ "clflush_shared_hit", "clflush_remote_hit", "clflush_local_hit_n", "clflush_miss_n", ], colors=["y", "r", "g", "b"], separate_hthreads=False, title=None, ): """ Creates a facet grid showing all points """ if separate_hthreads: colors=["y", "r", "g", "b"] for el in shown: for helper, main in itertools.product((0, 1), (0, 1)): df[el+f"_m{main}h{helper}"] = df[(df["main_ht"] == main) & (df["helper_ht"] == helper)][el] grid = sns.FacetGrid(df, row=row, col=col) for i, el in enumerate(shown): if separate_hthreads: for helper, main in itertools.product((0, 1), (0, 1)): grid.map(draw_fn, third, el+f"_m{main}h{helper}", color=colors[(helper+2*main) % len(colors)])# marker=['+', 'x'][helper]) else: grid.map(draw_fn, third, el, color=colors[i % len(colors)]) if title is not None: plot(title, g=grid) return grid def all_facets(df, pre="", post="", *args, **kwargs): """ df : panda dataframe pre, post: strings to add before and after the filename """ facet_grid( df, "helper_core_fixed", "main_core_fixed", "slice_group", title=f"{pre}facet_slice{post}.png", *args, **kwargs ) facet_grid( df, "helper_core_fixed", "slice_group", "main_core_fixed", title=f"{pre}facet_main{post}.png", *args, **kwargs ) facet_grid( df, "main_core_fixed", "slice_group", "helper_core_fixed", title=f"{pre}facet_helper{post}.png", *args, **kwargs ) def do_facet(main: int, helper: int, line: bool, metrics: str): """ - metrics: hit, miss or all """ df = stats.copy(deep=True) print(f"Doing all facets {main}x{helper} {metrics}") filtered_df = stats[ (stats["main_socket"] == main) & (stats["helper_socket"] == helper) ] method = "line" if line else "pt" shown = [] colors = [] if metrics == "hit" or metrics == "all": shown.append("clflush_remote_hit") colors.append("r") if metrics == "miss" or metrics == "all": shown.append("clflush_miss_n") colors.append("b") all_facets( filtered_df, pre=f"{metrics}_{method}_", post=f"_m{main}h{helper}", shown=shown, colors=colors, draw_fn=sns.lineplot if line else sns.scatterplot ) if args.rslice: rslice() # do_predictions(stats) # all_facets(stats, shown=["clflush_remote_hit"], colors=["r"]) with Pool(8) as pool: pool.starmap( do_facet, itertools.product( stats["main_socket"].unique(), stats["helper_socket"].unique(), (True, False), ("hit", "miss") ) )