Remove show_models.py

.gitignore

@@ -6,6 +6,7 @@
 #Cargo.lock
 # These are backup files generated by rustfmt
 **/*.rs.bk
+.vscode
 bootimage-kernel.iso
 bootimage-kerneliso
 kernel.sym
@@ -22,4 +23,5 @@ pxelinux.cfg
 serial.out
 memdisk
 **/venv
+**/.venv
 cache_utils/results*/

cache_utils/analyse_medians.py

@@ -213,23 +213,29 @@ def ha_dist(core, is_QPI):
     """
     if is_QPI:
         if core < 4:
-            return core
+            return core, 0
-        return 7-core
+        return 7-core, 1 # +1 for PCI
 
     if core < 4:
-        return 3-core
+        return 3-core, 0
-    return core-4
+    return core-4, 0
+
+def cclockwise_dist(source, dest):
+    base = (dest+8-source)%8
+    side_jump = 0
+    if source < 4 and dest >= 4:
+        side_jump = 1
+    elif source >= 4 and dest < 4:
+        side_jump = 2
+    return base, side_jump
 
 def cclockwise_ha_dist(core, is_QPI):
     """
     counter-clockwise distance to Home Agent
     """
     if is_QPI:
-        return 7-core
+        return cclockwise_dist(core, 7)
-
+    return cclockwise_dist(core, 3)
-    if core < 4:
-        return 3-core
-    return 11-core
 
 def miss_topology(main_core, slice_group, h, down_jump, top_jump, ini, ha_h):
     core, ring = slice_msg_distance(slice_group, main_core%8)
@@ -237,95 +243,81 @@ def miss_topology(main_core, slice_group, h, down_jump, top_jump, ini, ha_h):
     side_jump = 0
     side_jump += top_jump if ring == 2 else 0
     side_jump += down_jump if ring == 1 else 0
-    return (cclockwise_ha_dist(slice_group, False)//2)*ha_h+h*core + side_jump + ini
+    return (cclockwise_ha_dist(slice_group, False))*ha_h+h*core + side_jump + ini
 
 def miss_topology_df(x, h, down_jump, top_jump, ini, ha_h):
     func = lambda x, h, down_jump, top_jump, ini, ha_h: miss_topology(x["main_core_fixed"], x["slice_group"], h, down_jump, top_jump, ini, ha_h)
     return x.apply(func, args=(h, down_jump, top_jump, ini, ha_h), axis=1)
 
 
-def remote_hit_topology(main_core, helper_core, slice_group, C, h, H):
+def remote_hit_topology(main_core, helper_core, slice_group, const, core_jump, HA_jump):
-    core0, ring0, _ = slice_msg_distance(main_core, slice_group)
+    """
-    core1, ring1, _ = slice_msg_distance(helper_core, slice_group)
+    main_core
-    return C + h*(core0+core1) + H*(ring0+ring1)
+    -> local_slice
+    -> remote_slice
+    -> helper_core
+    -> remote_slice
+    -> local_slice
+    -> main_core
+    """
+    if main_core // 8 == helper_core // 8:
+        print("Can only do hit predictions for different socket", file=sys.stderr)
+        raise NotImplementedError
 
-def remote_hit_topology_df(x, C, h, H):
+    helper, main = helper_core%8, main_core%8
-    func = lambda x, C, h, H: remote_hit_topology(x["main_core_fixed"], x["helper_core_fixed"], x["slice_group"], C, h, H)
+    main_slice_local = slice_msg_distance(slice_group, main)
-    return x.apply(func, args=(C, h, H), axis=1)
+    slice_QPI = cclockwise_dist(0, slice_group) # clockwise
+    QPI_slice_r = cclockwise_dist(0, slice_group)
+    slice_r_helper = slice_msg_distance(slice_group, helper)
+
+    costs = (main_slice_local[0]+slice_QPI[0]+QPI_slice_r[0]+slice_r_helper[0], main_slice_local[1]+slice_QPI[1]+QPI_slice_r[1]+slice_r_helper[1])
+    return const+costs[0]*core_jump+costs[1]*HA_jump # may need some adjustments
+
+def remote_hit_topology_df(x, const, core_jump, HA_jump):
+    func = lambda x, const, core_jump, HA_jump: remote_hit_topology(x["main_core_fixed"], x["helper_core_fixed"], x["slice_group"], const, core_jump, HA_jump)
+    return x.apply(func, args=(const, core_jump, HA_jump), axis=1)
 
 
 def do_predictions(df):
     def plot_predicted_topo(col, row, x_ax, target, pred, df=df):
-        title_letter = {
+        titles = {
             "main_core_fixed": "A",
             "helper_core_fixed": "V",
-            "slice_group": "S"
+            "slice_group": "S",
-        }.get(col, col[0])
+            None: "None"
+        }
 
         figure_A0 = sns.FacetGrid(df, col=col, row=row)
-        figure_A0.map(sns.scatterplot, x_ax, pred, color="r")
+        figure_A0.map(sns.scatterplot, x_ax, target, color="g")
-        figure_A0.map(sns.scatterplot, x_ax, target, color="g", marker="+")
+        figure_A0.map(sns.scatterplot, x_ax, pred, color="r", marker="+")
-        figure_A0.set_titles(col_template="$"+title_letter+"$ = {col_name}")
+        figure_A0.set_titles(
+            col_template="$"+titles.get(col, col[0])+"$ = {col_name}",
+            row_template="$"+titles.get(row, row[0])+"$ = {row_name}"
+        )
         plot(f"medians_{pred}_{col}.png")
 
 
-
-    values = []
     main_socket, helper_socket = 0, 0
     dfc = df[(df["main_socket"] == main_socket) & (df["helper_socket"] == helper_socket)]
-    cores = sorted(list(dfc["main_core_fixed"].unique()))
-    slices = sorted(list(dfc["slice_group"].unique()))
     res_miss = optimize.curve_fit(
         miss_topology_df, dfc[["main_core_fixed", "slice_group"]], dfc["clflush_miss_n"]
     )
     print("Miss topology:")
     print(res_miss)
-    values.append(res_miss[0])
-
 
     dfc["predicted_miss"] = miss_topology_df(dfc, *(res_miss[0]))
     plot_predicted_topo("slice_group", None, "main_core_fixed", "clflush_miss_n", "predicted_miss", df=dfc)
     plot_predicted_topo("main_core_fixed", None, "slice_group", "clflush_miss_n", "predicted_miss", df=dfc)
 
-    for slice_ in slices:
+    main_socket, helper_socket = 0, 1
-        dfc = df[(df["slice_group"] == slice_) & (df["main_socket"] == main_socket) & (df["helper_socket"] == helper_socket)]
+    dfc = df[(df["main_socket"] == main_socket) & (df["helper_socket"] == helper_socket)]
-        res_miss = optimize.curve_fit(
-            miss_topology_df, dfc[["main_core_fixed", "slice_group"]], dfc["clflush_miss_n"]
-        )
-        values.append(res_miss[0])
-
-        dfc[f"predicted_miss_{slice_}"] = miss_topology_df(dfc, *(res_miss[0]))
-        plot_predicted_topo("slice_group", None, "main_core_fixed", "clflush_miss_n", f"predicted_miss_{slice_}", df=dfc)
-
-
-    print(list(values[0]))
-    print()
-    for i in values[1:]:
-        print(list(i))
-
-    values = []
-    for core in cores:
-        dfc = df[(df["main_core_fixed"] == core) & (df["main_socket"] == main_socket) & (df["helper_socket"] == helper_socket)]
-        res_miss = optimize.curve_fit(
-            miss_topology_df, dfc[["main_core_fixed", "slice_group"]], dfc["clflush_miss_n"]
-        )
-        values.append(res_miss[0])
-
-        dfc[f"predicted_miss_core{core}"] = miss_topology_df(dfc, *(res_miss[0]))
-        plot_predicted_topo("main_core_fixed", None, "slice_group", "clflush_miss_n", f"predicted_miss_core{core}", df=dfc)
-
-    for i in values:
-        print(list(i))
-    return
-
     res_remote_hit = optimize.curve_fit(
-        remote_hit_topology_df, df[["main_core_fixed", "helper_core_fixed", "slice_group"]], df["clflush_remote_hit"]
+        remote_hit_topology_df, dfc[["main_core_fixed", "helper_core_fixed", "slice_group"]], dfc["clflush_remote_hit"]
     )
     print("Remote hit topology:")
     print(res_remote_hit)
 
 
-
     df["diff_miss"] = df["clflush_miss_n"] - df["predicted_miss"]
     facet_grid(
         df, None, "main_core_fixed", "slice_group",
@@ -339,10 +331,20 @@ def do_predictions(df):
         shown=["diff_miss"],
         separate_hthreads=True
     )
+    dfc["predicted_remote_hit"] = remote_hit_topology_df(dfc, *(res_remote_hit[0]))
+    plot_predicted_topo("slice_group", "helper_core_fixed", "main_core_fixed", "clflush_remote_hit", "predicted_remote_hit", df=dfc)
+    plot_predicted_topo("main_core_fixed", "slice_group", "helper_core_fixed", "clflush_remote_hit", "predicted_remote_hit", df=dfc)
+    plot_predicted_topo("helper_core_fixed", "main_core_fixed", "slice_group", "clflush_remote_hit", "predicted_remote_hit", df=dfc)
 
-    # df["predicted_remote_hit"] = remote_hit_topology_df(df, *(res_remote_hit[0]))
+    for col in ["slice_group", "helper_core_fixed", "main_core_fixed"]:
-    # plot_predicted_topo("slice_group", "helper_core_fixed", "main_core_fixed", "clflush_remote_hit", "predicted_remote_hit")
+        for val in sorted(list(dfc[col].unique())):
-    # plot_predicted_topo("main_core_fixed", "helper_core_fixed", "slice_group", "clflush_remote_hit", "predicted_remote_hit")
+            df_temp = dfc[(dfc[col] == val)]
+            res_remote_hit = optimize.curve_fit(
+            remote_hit_topology_df, df_temp[["main_core_fixed", "helper_core_fixed", "slice_group"]], df_temp["clflush_remote_hit"]
+            )
+            df_temp[f"predicted_remote_hit_{col}={val}"] = remote_hit_topology_df(df_temp, *(res_remote_hit[0]))
+            plot_predicted_topo("slice_group", "helper_core_fixed", "main_core_fixed", "clflush_remote_hit", f"predicted_remote_hit_{col}={val}", df=df_temp)
+            plot_predicted_topo("main_core_fixed", "helper_core_fixed", "slice_group", "clflush_remote_hit", f"predicted_remote_hit_{col}={val}", df=df_temp)
 
 
 
@@ -374,6 +376,7 @@ def facet_grid(
         colors=["y", "r", "g", "b"],
         separate_hthreads=False,
         title=None,
+        letters=None
     ):
     """
     Creates a facet grid showing all points
@@ -400,6 +403,10 @@ def facet_grid(
         else:
             grid.map(draw_fn, third, el, color=colors[i % len(colors)], marker='+')
 
+    if letters is not None:
+        grid.set_titles(col_template="$"+letters[0]+"$ = {row_name}", row_template="$"+letters[1]+"$ = {col_name}")
+    
+
     if title is not None:
         plot(title, g=grid)
     return grid
@@ -465,27 +472,66 @@ if args.rslice:
     rslice()
 
 do_predictions(stats)
-#all_facets(stats, shown=["clflush_remote_hit"], colors=["r"], pre="hit")
+all_facets(stats, shown=["clflush_remote_hit"], colors=["r"], pre="hit_")
-#all_facets(stats, shown=["clflush_miss_n"], colors=["b"], pre="miss")
+all_facets(stats, shown=["clflush_miss_n"], colors=["b"], pre="miss_")
 
-#df=stats
+def compare_facing():
-#for m, h, s in itertools.product((0, 1), (0, 1), df["slice_group"].unique()):
+    df=stats
-#    dfc = df[(df["main_socket"] == m) & (df["main_core_fixed"]%8é == s) & (df["helper_socket"] == h)]
+    for m, h, s in itertools.product((0, 1), (0, 1), df["slice_group"].unique()):
-#
+        dfc = df[(df["main_socket"] == m) & (df["main_core_fixed"]%8 == s) & (df["helper_socket"] == h)]
-#    grid = sns.FacetGrid(dfc, row=None, col=None)
+
-#    grid.map(sns.scatterplot, "slice_group", "clflush_miss_n", marker="+")
+        grid = sns.FacetGrid(dfc, row=None, col=None)
-#
+        grid.map(sns.scatterplot, "slice_group", "clflush_miss_n", marker="+")
-#    plot(f"miss_m{m}h{h}m{s}", g=grid)
+
+        plot(f"miss_m{m}h{h}m{s}", g=grid)
 
 
-#with Pool(8) as pool:
+def isolate_sockets():
-#    pool.starmap(
+    with Pool(8) as pool:
-#        do_facet,
+        pool.starmap(
-#        itertools.product(
+            do_facet,
-#            stats["main_socket"].unique(),
+            itertools.product(
-#            stats["helper_socket"].unique(),
+                stats["main_socket"].unique(),
-#            (False, ),
+                stats["helper_socket"].unique(),
-#            ("hit", "miss")
+                (False, ),
-#        )
+                ("hit", "miss")
-#    )
+            )
+        )
 
+def superpose_sockets():
+    for main, same_socket in itertools.product(sorted(stats["main_core_fixed"].unique()), (True, False)):
+        df = stats[
+            (stats["slice_group"] == (main%8))
+            & (stats["main_core_fixed"] == main)
+            & ((stats["main_socket"] == stats["helper_socket"]) == same_socket)
+        ]
+        ax = sns.scatterplot(df, x="helper_core_fixed", y="clflush_remote_hit", marker="+", color="r")
+        ax.set_title(f"$S = {main%8}, V = {main}$")
+        plot(f"hit_{same_socket}_main{main:02d}.png")
+
+    df = stats[
+        (stats["slice_group"] == (stats["main_core_fixed"]%8))
+        & ((stats["main_core_fixed"]%8) == (stats["helper_core_fixed"]%8))
+        & (stats["main_socket"] != stats["helper_socket"])
+    ]
+    ax = sns.scatterplot(df, x="slice_group", y="clflush_remote_hit", marker="+", color="r")
+    plot(f"hit_same_slice.png")
+
+    stats["main_core_nosock"] = stats["main_core_fixed"]%8
+    stats["helper_core_nosock"] = stats["helper_core_fixed"]%8
+
+    facet_grid(
+        stats[(stats["main_socket"] != stats["helper_socket"])], "helper_core_nosock", "main_core_nosock", "slice_group",
+        title=f"hit_facet_slice_diff_socket.png",
+        separate_hthreads=True,
+        shown=["clflush_remote_hit"],
+        letters="VA"
+    )
+
+    facet_grid(
+        stats[(stats["main_socket"] == stats["helper_socket"])], "helper_core_nosock", "main_core_nosock", "slice_group",
+        title=f"hit_facet_slice_same_socket.png",
+        separate_hthreads=True,
+        letters="VA",
+        shown=["clflush_remote_hit"]
+    )

cache_utils/show_models.py

@@ -1,189 +0,0 @@
-"""
-Try some models and see what they look like
-Following this die could help https://en.wikichip.org/w/images/4/48/E5_v4_LCC.png
-Using the following naming convention:
-
-    ------
-    0    7
-    1    6
-    2    5
-    3    4
-    ------
-"""
-import matplotlib.pyplot as plt
-import seaborn as sns
-import pandas as pd
-import numpy as np
-import itertools
-import os
-
-nb_cores = 8
-nb_slices = 8
-num_core = nb_cores
-
-cores = list(range(nb_cores))
-slices = list(range(nb_slices))
-
-img_dir = os.getenv("PWD")+"/"
-def plot(filename, g=None):
-    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(img_dir + filename, "saved")
-    plt.close()
-
-def ring_distance(x0, x1):
-    """
-    return (a, b) where `a` is the core distance and `b` the larger "ring step"
-    it is possible that going from 0->7 costs one more than 3->4
-    """
-    dist = abs(x0-x1)
-    if x0 // (num_core/2) != x1 // (num_core/2):
-        # côté du coeur différent
-        return min((num_core-1-dist, 2), (dist-1, 1))
-    else:
-        return dist, 0
-
-def slice_msg_distance(source, dest):
-    """
-    Si l'expéditeur est à l'extrémité d'une des lignes, il envoie toujours dans le même sens
-    (vers toute sa ligne d'abord), sinon, il prend le chemin le plus court
-    le bonus correspond au fait que 0->7 puisse coûter 1 de plus que 3->4
-    """
-    dist = abs(source-dest)
-    if source // (num_core/2) == dest // (num_core/2):
-        return (dist, 0)
-
-    # Pour aller de l'autre côté
-    up, down = (num_core-1-dist, 2), (dist-1, 1)
-    if source in [0, 7]:
-        return down
-    if source in [3, 4] or source in [2, 5]:
-        return up
-    if source in [1, 6]:
-        return min(up, down)
-
-    raise IndexError
-
-def ha_dist(core, is_QPI):
-    """
-    distance to Home Agent
-    """
-    if is_QPI:
-        if core < 4:
-            return core, 0
-        return 7-core, 1 # +1 for PCI
-
-    if core < 4:
-        return 3-core, 0
-    return core-4, 0
-
-def cclockwise_dist(source, dest):
-    base = (dest+8-source)%8
-    side_jump = 0
-    if source < 4 and dest >= 4:
-        side_jump = 1
-    elif source >= 4 and dest < 4:
-        side_jump = 2
-    return base, side_jump
-
-def cclockwise_ha_dist(core, is_QPI):
-    """
-    counter-clockwise distance to Home Agent
-    """
-    if is_QPI:
-        return cclockwise_dist(core, 7)
-    return cclockwise_dist(core, 3)
-
-def no_QPI_dist(source, dest):
-    """
-    Path not using QPI hop
-    """
-    return abs(source-dest), 1 if source // 4 != dest //4 else 0
-
-
-def miss():
-    """
-    - ini : initial cost
-    - core_step : cost to go from one core to the following (eg 0 to 1)
-    - ring_step : cost to go from one line to the other (eg 0 to 7)
-
-    Issue: on a same socket, we observe always the first 4 or last 4 patterns, but not mixed
-    """
-    def miss_topo(main_core, slice, i, c, l, k):
-        x, y, z = slice_msg_distance(main_core, slice)
-        return i+c*x+l*y+k*z
-
-    ini, core_step, ring_step, other_ring_step = 4, 1, 5, 2
-    fig, axs = plt.subplots(nrows=1, ncols=8, figsize=(15, 5))
-
-    for i, slice in enumerate(range(8)):
-        axs[i].plot(cores, [miss_topo(main_core, slice, ini, core_step, ring_step, other_ring_step) for main_core in cores], "ro")
-        axs[i].set_title(f"slice_group = {slice}")
-        axs[i].set_ylabel("clflush_miss_n")
-        axs[i].set_xlabel("main_core_fixed")
-        axs[i].set_ylim([0, 25])
-
-    plt.tight_layout()
-    plt.show()
-
-
-def hit():
-    """
-    - ini : initial cost
-    - core_step : cost to go from one core to the following (eg 0 to 1)
-    - ring_step : cost to go from one line to the other (eg 0 to 7)
-
-    Issue: on a same socket, we observe always the first 4 or last 4 patterns, but not mixed
-    """
-    def hit_topo(main, helper, slice_g, i, c, l):
-        helper = helper%8
-
-        main_slice_local = slice_msg_distance(slice_g, main)
-        slice_QPI = cclockwise_dist(0, slice_g) # clockwise
-        QPI_slice_r = cclockwise_dist(0, slice_g)
-        slice_r_helper = slice_msg_distance(slice_g, helper)
-
-        costs = (main_slice_local[0]+slice_QPI[0]+QPI_slice_r[0]+slice_r_helper[0], main_slice_local[1]+slice_QPI[1]+QPI_slice_r[1]+slice_r_helper[1])
-        return ini+costs[0]*c+costs[1]*l
-
-
-    ini, core_step, ring_step = 12, 1, 1.5
-    fig, axs = plt.subplots(nrows=1, ncols=8, figsize=(15, 5))
-
-    # Define the ranges for x, y, z
-    main = range(8)
-    helper = range(8, 16)
-    slice_g = range(8)
-
-    # Create a DataFrame with all combinations of x, y, and z
-    data = pd.DataFrame([
-            (x, y, z) for z in slice_g
-            for x, y in itertools.product(main, helper)
-        ],
-        columns=['main', 'helper', 'slice_group']
-    )
-
-    # Define the function
-    def my_function(x):
-        return hit_topo(x["main"], x["helper"], x["slice_group"], ini, core_step, ring_step)
-
-    # Apply the function to create a new column
-    data['predicted_hit'] = data.apply(my_function, axis=1)
-
-    fig = sns.FacetGrid(data, col="main", row="helper")
-    fig.map(sns.scatterplot, "slice_group", "predicted_hit", color="r", marker="+")
-    fig.map(sns.scatterplot, "slice_group", "predicted_hit", color="r", marker="x")
-    fig.set_titles(col_template="$A$ = {col_name}", row_template="$V$ = {row_name}")
-
-    plot("model_hit.png", g=fig)
-
-
-
-hit()