{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "#### Analysis" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [], "source": [ "from typing import List\n", "import os\n", "import pickle\n", "import squarify\n", "\n", "import numpy as np\n", "import pandas as pd\n", "import seaborn as sns\n", "import matplotlib\n", "import matplotlib.pyplot as plt\n", "import matplotlib.patches as mpatches\n", "from matplotlib.lines import Line2D\n", "\n", "SAVEPATH = \"./figure\"\n", "TRACEPATH = \"./data/job_trace\"\n", "PKLPATH = \"./data/utilization/util_pkl\"\n", "\n", "sns.set_style(\"ticks\")\n", "font = {\n", " \"font.family\": \"Roboto\",\n", " \"font.size\": 12,\n", "}\n", "sns.set_style(font)\n", "paper_rc = {\n", " \"lines.linewidth\": 3,\n", " \"lines.markersize\": 10,\n", "}\n", "sns.set_context(\"paper\", font_scale=2, rc=paper_rc)\n", "cmp = sns.color_palette(\"tab10\")\n", "\n", "\n", "def autolabel(rects, ax, prec=1):\n", " \"\"\"Attach a text label above each bar in *rects*, displaying its height.\"\"\"\n", " for rect in rects:\n", " height = rect.get_height()\n", " ax.annotate(\n", " f\"{height:.{prec}f}\",\n", " xy=(rect.get_x() + rect.get_width() / 2, height),\n", " xytext=(0, 3), # 3 points vertical offset\n", " textcoords=\"offset points\",\n", " ha=\"center\",\n", " va=\"bottom\",\n", " size=16,\n", " )\n", "\n", "\n", "def calculate_num_cdf_customized_xaxis(df: pd.DataFrame, x_axis: List, key: str):\n", " \"\"\"\n", " Calculate quantity percentile CDF with customized threshold of x-axis, y-axis: 0-100%,\n", " \"\"\"\n", " # print(\"Parsing\")\n", " data = df[[key]].copy()\n", " data.dropna(inplace=True)\n", "\n", " y = [len(data[data[key] <= x]) / len(data) * 100 for x in x_axis]\n", "\n", " return y\n", "\n", "\n", "def calculate_sum_cdf_customized_xaxis(df: pd.DataFrame, x_axis: List, key: str, key_to_time=None):\n", " \"\"\"\n", " Calculate sum CDF with customized threshold of x-axis, y-axis: 0-100%,\n", " \"\"\"\n", " if key_to_time is not None:\n", " data = df[[key, key_to_time]].copy()\n", " data[\"new\"] = data[key] * data[key_to_time]\n", " else:\n", " data = df[[key]].copy()\n", " data[\"new\"] = data[key]\n", " data.dropna(inplace=True)\n", " sum = data[\"new\"].sum()\n", "\n", " y = [data[data[key] <= x][\"new\"].sum() / sum * 100 for x in x_axis]\n", "\n", " return y\n", "\n", "\n", "if not os.path.exists(SAVEPATH):\n", " os.makedirs(SAVEPATH)\n", "\n", "\n", "data_seren = pd.read_csv(f\"{TRACEPATH}/trace_seren.csv\")\n", "data_kalos = pd.read_csv(f\"{TRACEPATH}/trace_kalos.csv\")\n", "data_philly = pd.read_csv(f\"{TRACEPATH}/trace_previous_work/philly_trace.csv\")\n", "data_helios = pd.read_csv(f\"{TRACEPATH}/trace_previous_work/helios_trace.csv\")\n", "data_pai = pd.read_csv(f\"{TRACEPATH}/trace_previous_work/pai_trace.csv\")\n", "\n", "# A few further process\n", "data_pai.rename(columns={\"plan_cpu\": \"cpu_num\", \"plan_gpu\": \"gpu_num\", \"wait_time\": \"queue\", \"status\": \"state\"}, inplace=True)\n", "data_pai[[\"cpu_num\", \"gpu_num\"]] /= 100\n", "data_pai[\"state\"] = data_pai[\"state\"].map({\"Failed\": \"FAILED\"}) # Not suitable for final state analysis\n", "data_philly[\"state\"] = data_philly[\"state\"].map({\"Pass\": \"COMPLETED\", \"Failed\": \"FAILED\", \"Killed\": \"CANCELLED\"})" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### CDF: GPU Job Duration & Utilization" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "x = [2**i for i in range(0, 22)]\n", "y_gpu_seren = calculate_num_cdf_customized_xaxis(data_seren[data_seren[\"gpu_num\"] > 0], x_axis=x, key=\"duration\")\n", "y_gpu_kalos = calculate_num_cdf_customized_xaxis(data_kalos[data_kalos[\"gpu_num\"] > 0], x_axis=x, key=\"duration\")\n", "y_gpu_philly = calculate_num_cdf_customized_xaxis(data_philly[data_philly[\"gpu_num\"] > 0], x_axis=x, key=\"duration\")\n", "y_gpu_helios = calculate_num_cdf_customized_xaxis(data_helios[data_helios[\"gpu_num\"] > 0], x_axis=x, key=\"duration\")\n", "y_gpu_pai = calculate_num_cdf_customized_xaxis(data_pai[data_pai[\"gpu_num\"] > 0], x_axis=x, key=\"duration\")\n", "\n", "with open(f\"{PKLPATH}/util_gpu_seren.pkl\", \"rb\") as file:\n", " x1, y1, _, _, _, _, _, _, _, _ = pickle.load(file)\n", "\n", "with open(f\"{PKLPATH}/util_gpu_kalos.pkl\", \"rb\") as file:\n", " x4, y4, _, _ = pickle.load(file)\n", "\n", "with open(f\"{PKLPATH}/util_gpu_pai.pkl\", \"rb\") as file: # Collect via Antman\n", " x2, y2 = pickle.load(file)\n", "\n", "with open(f\"{PKLPATH}/util_gpu_philly.pkl\", \"rb\") as file:\n", " x3, y3 = pickle.load(file)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "linestyles = [\"-\", \"--\", \":\", \":\", \":\"]\n", "grid_params = dict(width_ratios=[1, 1])\n", "fig, (ax1, ax2) = plt.subplots(ncols=2, nrows=1, constrained_layout=True, figsize=(9, 3.75))\n", "\n", "ax1.plot(x, y_gpu_seren, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren\")\n", "ax1.plot(x, y_gpu_kalos, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos\")\n", "ax1.plot(x, y_gpu_philly, linestyles[2], linewidth=3, alpha=0.9, color=cmp[2], label=\"Philly\")\n", "ax1.plot(x, y_gpu_helios, linestyles[3], linewidth=3, alpha=0.9, color=cmp[3], label=\"Helios\")\n", "ax1.plot(x, y_gpu_pai, linestyles[3], linewidth=3, alpha=0.9, color=cmp[4], label=\"PAI\")\n", "\n", "ax2.plot(x1, y1, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren\")\n", "ax2.plot(x4, y4, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos\")\n", "ax2.plot(x2, y2, linestyles[2], linewidth=3, alpha=0.9, color=cmp[4], label=\"PAI\")\n", "ax2.plot(x3, y3, linestyles[2], linewidth=3, alpha=0.9, color=cmp[2], label=\"Philly\")\n", "\n", "ax1.set_xlabel(f\"(a) GPU Job Duration (s)\")\n", "ax1.set_ylabel(f\"CDF (%)\")\n", "ax1.set_xscale(\"log\")\n", "ax1.set_xticks([1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6])\n", "ax1.set_xlim(1, x[-1])\n", "ax1.set_ylim(-0.5, 100.8)\n", "ax1.grid(linestyle=\":\")\n", "\n", "ax2.set_xlabel(f\"(b) GPU Utilization (%)\")\n", "ax2.set_ylabel(f\"CDF (%)\")\n", "ax2.set_xlim(-0.8, 100.8)\n", "ax2.set_xticks([0, 25, 50, 75, 100])\n", "ax2.set_ylim(0, 100.8)\n", "ax2.grid(linestyle=\":\")\n", "\n", "handles, labels = ax1.get_legend_handles_labels()\n", "fig.legend(handles=handles, labels=labels, ncols=5, bbox_to_anchor=(0.1, 1.145), loc=2, columnspacing=1.5, handletextpad=0.5)\n", "\n", "sns.despine()\n", "fig.savefig(f\"{SAVEPATH}/cdf_job_duration_util.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### CDF: GPU Number" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "x = [i for i in range(0, 1025)]\n", "y_gpu_seren = calculate_num_cdf_customized_xaxis(data_seren[data_seren[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\")\n", "y_gpu_kalos = calculate_num_cdf_customized_xaxis(data_kalos[data_kalos[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\")\n", "y_gpu_philly = calculate_num_cdf_customized_xaxis(data_philly[data_philly[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\")\n", "y_gpu_helios = calculate_num_cdf_customized_xaxis(data_helios[data_helios[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\")\n", "y_gpu_pai = calculate_num_cdf_customized_xaxis(data_pai[data_pai[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\")\n", "\n", "y_gtime_seren = calculate_sum_cdf_customized_xaxis(\n", " data_seren[data_seren[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\", key_to_time=\"duration\"\n", ")\n", "y_gtime_kalos = calculate_sum_cdf_customized_xaxis(\n", " data_kalos[data_kalos[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\", key_to_time=\"duration\"\n", ")\n", "y_gtime_philly = calculate_sum_cdf_customized_xaxis(\n", " data_philly[data_philly[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\", key_to_time=\"duration\"\n", ")\n", "y_gtime_helios = calculate_sum_cdf_customized_xaxis(\n", " data_helios[data_helios[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\", key_to_time=\"duration\"\n", ")\n", "y_gtime_pai = calculate_sum_cdf_customized_xaxis(\n", " data_pai[data_pai[\"gpu_num\"] > 0], x_axis=x, key=\"gpu_num\", key_to_time=\"duration\"\n", ")" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "linestyles = [\"-\", \"--\", \":\", \":\", \":\"]\n", "grid_params = dict(width_ratios=[1, 1])\n", "fig, (ax1, ax2) = plt.subplots(ncols=2, nrows=1, constrained_layout=True, figsize=(9, 3.75))\n", "\n", "ax1.plot(x, y_gpu_seren, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren\")\n", "ax1.plot(x, y_gpu_kalos, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos\")\n", "ax1.plot(x, y_gpu_philly, linestyles[2], linewidth=3, alpha=0.9, color=cmp[2], label=\"Philly\")\n", "ax1.plot(x, y_gpu_helios, linestyles[3], linewidth=3, alpha=0.9, color=cmp[3], label=\"Helios\")\n", "ax1.plot(x, y_gpu_pai, linestyles[3], linewidth=3, alpha=0.9, color=cmp[4], label=\"PAI\")\n", "\n", "ax2.plot(x, y_gtime_seren, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren\")\n", "ax2.plot(x, y_gtime_kalos, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos\")\n", "ax2.plot(x, y_gtime_philly, linestyles[2], linewidth=3, alpha=0.9, color=cmp[2], label=\"Philly\")\n", "ax2.plot(x, y_gtime_helios, linestyles[3], linewidth=3, alpha=0.9, color=cmp[3], label=\"Helios\")\n", "ax2.plot(x, y_gtime_pai, linestyles[3], linewidth=3, alpha=0.9, color=cmp[4], label=\"PAI\")\n", "\n", "\n", "ax1.set_xlabel(f\"(a) Number of GPU\")\n", "ax1.set_ylabel(f\"CDF of Jobs (%)\")\n", "ax1.set_xscale(\"log\", base=2)\n", "ax1.set_xticks([2**i for i in range(0, 11, 2)])\n", "ax1.set_xticklabels(\n", " [2**i for i in range(0, 10, 2)]\n", " + [\n", " \"1024+\",\n", " ]\n", ")\n", "ax1.set_xlim(1, x[-1] + 1)\n", "ax1.set_ylim(-0.5, 100.8)\n", "ax1.grid(linestyle=\":\")\n", "\n", "ax2.set_xlabel(f\"(b) Number of GPU\")\n", "ax2.set_ylabel(f\"CDF of GPU Time (%)\")\n", "ax2.set_xscale(\"log\", base=2)\n", "ax2.set_xticks([2**i for i in range(0, 11, 2)])\n", "ax2.set_xticklabels(\n", " [2**i for i in range(0, 10, 2)]\n", " + [\n", " \"1024+\",\n", " ]\n", ")\n", "ax2.set_xlim(1, x[-1] + 50)\n", "ax2.set_ylim(-0.5, 100.8)\n", "ax2.grid(linestyle=\":\")\n", "\n", "handles, labels = ax1.get_legend_handles_labels()\n", "fig.legend(handles=handles, labels=labels, ncols=5, bbox_to_anchor=(0.1, 1.145), loc=2, columnspacing=1.5, handletextpad=0.5)\n", "\n", "sns.despine()\n", "fig.savefig(f\"{SAVEPATH}/cdf_job_gpunum.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Bar: Job Final State" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "df = pd.read_csv(\"./data/cluster_summary.csv\", index_col=\"id\")\n", "grid_params = dict(width_ratios=[1, 1])\n", "fig, (ax1, ax2) = plt.subplots(ncols=2, nrows=1, constrained_layout=True, figsize=(9, 3.75))\n", "\n", "x = np.arange(1, 3)\n", "width = 0.22\n", "p1 = ax1.bar(\n", " x - width,\n", " df.loc[[\"Seren\", \"Kalos\"], \"complete_rate_gpu\"] * 100,\n", " width,\n", " label=\"Completed\",\n", " alpha=0.8,\n", " linewidth=1,\n", " edgecolor=\"k\",\n", ")\n", "p2 = ax1.bar(\n", " x, df.loc[[\"Seren\", \"Kalos\"], \"cancel_rate_gpu\"] * 100, width, label=\"Canceled\", alpha=0.8, linewidth=1, edgecolor=\"k\"\n", ")\n", "p3 = ax1.bar(\n", " x + width, df.loc[[\"Seren\", \"Kalos\"], \"fail_rate_gpu\"] * 100, width, label=\"Failed\", alpha=0.8, linewidth=1, edgecolor=\"k\"\n", ")\n", "\n", "p4 = ax2.bar(\n", " x - width,\n", " df.loc[[\"Seren\", \"Kalos\"], \"complete_rate_gpu_time\"] * 100,\n", " width,\n", " label=\"Completed\",\n", " alpha=0.8,\n", " linewidth=1,\n", " edgecolor=\"k\",\n", ")\n", "p5 = ax2.bar(\n", " x, df.loc[[\"Seren\", \"Kalos\"], \"cancel_rate_gpu_time\"] * 100, width, label=\"Canceled\", alpha=0.8, linewidth=1, edgecolor=\"k\"\n", ")\n", "p6 = ax2.bar(\n", " x + width,\n", " df.loc[[\"Seren\", \"Kalos\"], \"fail_rate_gpu_time\"] * 100,\n", " width,\n", " label=\"Failed\",\n", " alpha=0.8,\n", " linewidth=1,\n", " edgecolor=\"k\",\n", ")\n", "\n", "autolabel(p1, ax1)\n", "autolabel(p2, ax1)\n", "autolabel(p3, ax1)\n", "autolabel(p4, ax2)\n", "autolabel(p5, ax2)\n", "autolabel(p6, ax2)\n", "\n", "ax1.set_xlabel(f\"(a) Job Count\")\n", "ax1.set_ylabel(f\"Fraction (%)\")\n", "ax1.set_xticks(x)\n", "ax1.set_xticklabels([\"Seren\", \"Kalos\"])\n", "ax1.set_xlim(0.5, 2.5)\n", "ax1.set_ylim(0, 100)\n", "ax1.grid(axis=\"y\", linestyle=\":\")\n", "\n", "ax2.set_xlabel(f\"(b) GPU Time\")\n", "ax2.set_ylabel(f\"Fraction (%)\")\n", "ax2.set_xticks(x)\n", "ax2.set_xticklabels([\"Seren\", \"Kalos\"])\n", "ax2.set_xlim(0.5, 2.5)\n", "ax2.set_ylim(0, 100)\n", "ax2.grid(axis=\"y\", linestyle=\":\")\n", "\n", "handles, labels = ax1.get_legend_handles_labels()\n", "fig.legend(handles=handles, labels=labels, ncols=5, bbox_to_anchor=(0.18, 1.145), loc=2)\n", "\n", "sns.despine()\n", "fig.savefig(f\"{SAVEPATH}/bar_job_state.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Treemap: Job Number Distribution" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "print(\"Processing Seren\")\n", "datas = data_seren[data_seren[\"gpu_num\"] > 0]\n", "\n", "job_type = [\"Eval\", \"Pretrain\", \"SFT\", \"MLLM\", \"Debug\", \"Other\"]\n", "df = pd.DataFrame(index=job_type, columns=[\"job_count\"]).fillna(0)\n", "df[\"job_count\"] = df.index.map(datas.groupby(\"type\").size()).astype(int)\n", "df[\"gtime\"] = df.index.map(datas.groupby(\"type\")[\"gpu_time\"].sum()).astype(int)\n", "\n", "total = df[\"job_count\"].sum()\n", "total_gtime = df[\"gtime\"].sum()\n", "\n", "df[\"count_percent\"] = df[\"job_count\"] / total * 100\n", "df[\"gtime_percent\"] = df[\"gtime\"] / total_gtime * 100\n", "\n", "# For plotting\n", "df[\"label\"] = [x + f\"\\n{df.at[x, 'count_percent']:.1f}%\" for x in list(df.index)]\n", "df[\"label_gtime\"] = [x + f\"\\n{df.at[x, 'gtime_percent']:.1f}%\" for x in list(df.index)]\n", "df[\"label_percent\"] = [f\"{df.at[x, 'count_percent']:.1f}%\" for x in list(df.index)]\n", "df[\"label_gtime_percent\"] = [f\"{df.at[x, 'gtime_percent']:.1f}%\" for x in list(df.index)]\n", "df_s = df.copy()\n", "\n", "print(\"Processing Kalos\")\n", "datak = data_kalos[data_kalos[\"gpu_num\"] > 0]\n", "\n", "job_type = [\"Eval\", \"Pretrain\", \"Debug\", \"Other\"]\n", "df = pd.DataFrame(index=job_type, columns=[\"job_count\"]).fillna(0)\n", "df[\"job_count\"] = df.index.map(datak.groupby(\"type\").size()).astype(int)\n", "df[\"gtime\"] = df.index.map(datak.groupby(\"type\")[\"gpu_time\"].sum()).astype(int)\n", "\n", "\n", "total = df[\"job_count\"].sum()\n", "total_gtime = df[\"gtime\"].sum()\n", "\n", "df[\"count_percent\"] = df[\"job_count\"] / total * 100\n", "df[\"gtime_percent\"] = df[\"gtime\"] / total_gtime * 100\n", "\n", "# For plotting\n", "df[\"label\"] = [x + f\"\\n{df.at[x, 'count_percent']:.1f}%\" for x in list(df.index)]\n", "df[\"label_gtime\"] = [x + f\"\\n{df.at[x, 'gtime_percent']:.1f}%\" for x in list(df.index)]\n", "df[\"label_percent\"] = [f\"{df.at[x, 'count_percent']:.1f}\\n%\" for x in list(df.index)]\n", "df[\"label_gtime_percent\"] = [f\"{df.at[x, 'gtime_percent']:.1f}\\n%\" for x in list(df.index)]\n", "df_k = df.copy()\n", "\n", "# For plotting\n", "df_k.at[\"Pretrain\", \"label_gtime_percent\"] = df_k.at[\"Pretrain\", \"label_gtime\"]\n", "df_k.at[\"Eval\", \"label_percent\"] = df_k.at[\"Eval\", \"label\"]\n", "df_k.at[\"Other\", \"label_percent\"] = \"\"\n", "df_k.at[\"Eval\", \"label_gtime_percent\"] = \" \"\n", "\n", "df_s.at[\"Pretrain\", \"label_gtime_percent\"] = df_s.at[\"Pretrain\", \"label_gtime\"]\n", "df_s.at[\"Eval\", \"label_percent\"] = df_s.at[\"Eval\", \"label\"]\n", "df_s.at[\"SFT\", \"label_percent\"] = df_s.at[\"SFT\", \"label\"]\n", "df_s.at[\"Other\", \"label_percent\"] = df_s.at[\"Other\", \"label\"]\n", "df_s.at[\"Pretrain\", \"label_percent\"] = \" \"\n", "df_s.at[\"Debug\", \"label_gtime_percent\"] = df_s.at[\"Debug\", \"label_gtime_percent\"].replace(\"%\", \"\\n%\")\n", "\n", "\n", "cmp_treemap = sns.color_palette(\"pastel\")\n", "label = df_s.index.to_list()\n", "df_s[\"color\"] = cmp_treemap[: len(df_s)]" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(ncols=2, nrows=2, constrained_layout=True, figsize=(9, 4.2))\n", "FONT = 15\n", "\n", "###### Fig 1 ######\n", "df_s.sort_values(by=\"count_percent\", ascending=False, inplace=True)\n", "squarify.plot(\n", " ax=ax1,\n", " sizes=list(df_s[\"job_count\"].values),\n", " label=df_s[\"label_percent\"],\n", " text_kwargs={\"fontsize\": FONT},\n", " color=df_s[\"color\"],\n", " bar_kwargs={\"alpha\": 0.8, \"linewidth\": 1, \"edgecolor\": \"k\"},\n", ")\n", "\n", "\n", "handles, labels = ax1.get_legend_handles_labels()\n", "handles_new = [handles[0], handles[-1], handles[1], handles[3], handles[4], handles[2]]\n", "fig.legend(\n", " handles=handles_new, labels=label, ncols=6, bbox_to_anchor=(0.0, 1.135), loc=2, columnspacing=0.82, handletextpad=0.2\n", ")\n", "\n", "###### Fig 2 ######\n", "df_s.sort_values(by=\"gtime\", ascending=False, inplace=True)\n", "squarify.plot(\n", " ax=ax2,\n", " sizes=list(df_s[\"gtime\"].values),\n", " label=df_s[\"label_gtime_percent\"],\n", " text_kwargs={\"fontsize\": FONT},\n", " color=df_s[\"color\"],\n", " bar_kwargs={\"alpha\": 0.8, \"linewidth\": 1, \"edgecolor\": \"k\"},\n", ")\n", "\n", "plt.tick_params(axis=\"both\", which=\"both\", bottom=False, top=False, left=False, right=False)\n", "\n", "ax1.set_xlabel(f\"(a) Job Count\", fontsize=16)\n", "ax2.set_xlabel(f\"(b) GPU Time\", fontsize=16)\n", "\n", "\n", "###### Fig 3 ######\n", "df_k.sort_values(by=\"count_percent\", ascending=False, inplace=True)\n", "df_k[\"color\"] = [df_s[\"color\"][job_name] for job_name in df_k.index]\n", "\n", "squarify.plot(\n", " ax=ax3,\n", " sizes=list(df_k[\"job_count\"].values),\n", " label=df_k[\"label_percent\"],\n", " text_kwargs={\"fontsize\": FONT},\n", " color=df_k[\"color\"],\n", " bar_kwargs={\"alpha\": 0.8, \"linewidth\": 1, \"edgecolor\": \"k\"},\n", ")\n", "\n", "###### Fig 4 ######\n", "df_k.sort_values(by=\"gtime\", ascending=False, inplace=True)\n", "squarify.plot(\n", " ax=ax4,\n", " sizes=list(df_k[\"gtime\"].values),\n", " label=df_k[\"label_gtime_percent\"],\n", " text_kwargs={\"fontsize\": FONT},\n", " color=df_k[\"color\"],\n", " bar_kwargs={\"alpha\": 0.8, \"linewidth\": 1, \"edgecolor\": \"k\"},\n", ")\n", "\n", "ax1.annotate(\n", " df_s.at[\"Pretrain\", \"label\"].split(\"\\n\")[1],\n", " xy=(97, 96),\n", " xytext=(90, 70),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=15,\n", ")\n", "\n", "ax3.annotate(\n", " df_k.at[\"Other\", \"label\"].split(\"\\n\")[1],\n", " xy=(98.5, 92),\n", " xytext=(80, 80),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=15,\n", ")\n", "\n", "ax4.annotate(\n", " df_k.at[\"Eval\", \"label_gtime\"].split(\"\\n\")[1],\n", " xy=(98.5, 93),\n", " xytext=(80, 80),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=15,\n", ")\n", "\n", "plt.tick_params(axis=\"both\", which=\"both\", bottom=False, top=False, left=False, right=False)\n", "\n", "ax3.set_xlabel(f\"(c) Job Count\", fontsize=16)\n", "ax4.set_xlabel(f\"(d) GPU Time\", fontsize=16, labelpad=8)\n", "\n", "ax1.set_xticks([])\n", "ax1.set_yticks([])\n", "ax2.set_xticks([])\n", "ax2.set_yticks([])\n", "ax3.set_xticks([])\n", "ax3.set_yticks([])\n", "ax4.set_xticks([])\n", "ax4.set_yticks([])\n", "\n", "ax1.text(0.015, 0.03, \"Seren\", transform=ax1.transAxes, size=18, fontweight=\"bold\")\n", "ax2.text(0.02, 0.03, \"Seren\", transform=ax2.transAxes, size=18, fontweight=\"bold\")\n", "ax3.text(0.02, 0.03, \"Kalos\", transform=ax3.transAxes, size=18, fontweight=\"bold\")\n", "ax4.text(0.02, 0.03, \"Kalos\", transform=ax4.transAxes, size=18, fontweight=\"bold\")\n", "fig.savefig(f\"{SAVEPATH}/treemap_job_dist.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### CDF: Duration and Queuing Delay of Different Type" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "\"\"\"\n", "(a) Seren Duration (b) Seren Queuing (c) Kalos Duration (d) Kalos Queuing\n", "\"\"\"\n", "\n", "# Duration part\n", "x = [2**i for i in range(0, 22)]\n", "y_gpu_seren_other = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Other\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_seren_debug = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Debug\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_seren_pretrain = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Pretrain\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_seren_eval = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Eval\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_seren_tuning = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"SFT\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_seren_mllm = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"MLLM\")], x_axis=x, key=\"duration\"\n", ")\n", "\n", "y_gpu_kalos_other = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Other\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_kalos_debug = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Debug\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_kalos_pretrain = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Pretrain\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_kalos_eval = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Eval\")], x_axis=x, key=\"duration\"\n", ")\n", "y_gpu_kalos_tuning = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"SFT\")], x_axis=x, key=\"duration\"\n", ")\n", "\n", "# Queuing part\n", "x2 = [2**i for i in range(0, 16)]\n", "y_que_s_other = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Other\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_s_debug = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Debug\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_s_pretrain = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Pretrain\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_s_eval = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"Eval\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_s_tuning = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"SFT\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_s_mllm = calculate_num_cdf_customized_xaxis(\n", " data_seren[(data_seren[\"gpu_num\"] > 0) & (data_seren[\"type\"] == \"MLLM\")], x_axis=x2, key=\"queue\"\n", ")\n", "\n", "y_que_ali_other = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Other\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_ali_debug = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Debug\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_ali_pretrain = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Pretrain\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_ali_eval = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"Eval\")], x_axis=x2, key=\"queue\"\n", ")\n", "y_que_ali_tuning = calculate_num_cdf_customized_xaxis(\n", " data_kalos[(data_kalos[\"gpu_num\"] > 0) & (data_kalos[\"type\"] == \"SFT\")], x_axis=x2, key=\"queue\"\n", ")" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "linestyles = [\"--\", \"-.\", \":\", \"--\", \"-.\", \":\"]\n", "grid_params = dict(width_ratios=[1, 1])\n", "fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(ncols=2, nrows=2, constrained_layout=True, figsize=(9, 7))\n", "\n", "# (a) Seren Duration\n", "ax1.plot(x, y_gpu_seren_eval, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Evaluation\")\n", "ax1.plot(x, y_gpu_seren_pretrain, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Pretrain\")\n", "ax1.plot(x, y_gpu_seren_tuning, linestyles[2], linewidth=3, alpha=0.9, color=cmp[2], label=\"SFT\")\n", "ax1.plot(x, y_gpu_seren_mllm, linestyles[0], linewidth=3, alpha=0.9, color=cmp[3], label=\"MLLM\")\n", "ax1.plot(x, y_gpu_seren_debug, linestyles[1], linewidth=3, alpha=0.9, color=cmp[4], label=\"Debug\")\n", "ax1.plot(x, y_gpu_seren_other, linestyles[2], linewidth=3, alpha=0.9, color=cmp[5], label=\"Other\")\n", "\n", "\n", "# (b) Seren Queuing\n", "ax2.plot(x2, y_que_s_eval, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Evaluation\")\n", "ax2.plot(x2, y_que_s_pretrain, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Pretrain\")\n", "ax2.plot(x2, y_que_s_tuning, linestyles[2], linewidth=3, alpha=0.9, color=cmp[2], label=\"SFT\")\n", "ax2.plot(x2, y_que_s_mllm, linestyles[0], linewidth=3, alpha=0.9, color=cmp[3], label=\"MLLM\")\n", "ax2.plot(x2, y_que_s_debug, linestyles[1], linewidth=3, alpha=0.9, color=cmp[4], label=\"Debug\")\n", "ax2.plot(x2, y_que_s_other, linestyles[2], linewidth=3, alpha=0.9, color=cmp[5], label=\"Other\")\n", "\n", "# (c) Kalos Duration\n", "ax3.plot(x, y_gpu_kalos_eval, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Evaluation\")\n", "ax3.plot(x, y_gpu_kalos_pretrain, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Pretrain\")\n", "ax3.plot(x, y_gpu_kalos_debug, linestyles[1], linewidth=3, alpha=0.9, color=cmp[4], label=\"Debug\")\n", "ax3.plot(x, y_gpu_kalos_other, linestyles[2], linewidth=3, alpha=0.9, color=cmp[5], label=\"Other\")\n", "\n", "\n", "# (d) Kalos Queuing\n", "ax4.plot(x2, y_que_ali_eval, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Evaluation\")\n", "ax4.plot(x2, y_que_ali_pretrain, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Pretrain\")\n", "ax4.plot(x2, y_que_ali_debug, linestyles[1], linewidth=3, alpha=0.9, color=cmp[4], label=\"Debug\")\n", "ax4.plot(x2, y_que_ali_other, linestyles[2], linewidth=3, alpha=0.9, color=cmp[5], label=\"Other\")\n", "\n", "ax1.set_xlabel(f\"(a) Job Duration (s)\")\n", "ax1.set_ylabel(f\"CDF (%)\")\n", "ax1.set_xscale(\"log\")\n", "ax1.set_xticks([1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6])\n", "ax1.set_xlim(1, x[-1])\n", "ax1.set_ylim(-0.5, 100.8)\n", "handles, labels = ax1.get_legend_handles_labels()\n", "fig.legend(handles=handles, labels=labels, ncols=6, bbox_to_anchor=(-0.01, 1.08), loc=2, columnspacing=0.9, handletextpad=0.2)\n", "ax1.grid(linestyle=\":\")\n", "\n", "ax2.set_xlabel(f\"(b) Job Queuing Delay (s)\")\n", "ax2.set_ylabel(f\"CDF (%)\")\n", "ax2.set_xscale(\"log\")\n", "ax2.set_xticks([1e0, 1e1, 1e2, 1e3, 1e4])\n", "ax2.set_xlim(1, x2[-1])\n", "ax2.set_ylim(-0.5, 100.8)\n", "ax2.grid(linestyle=\":\")\n", "\n", "ax3.set_xlabel(f\"(c) Job Duration (s)\")\n", "ax3.set_ylabel(f\"CDF (%)\")\n", "ax3.set_xscale(\"log\")\n", "ax3.set_xticks([1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6])\n", "ax3.set_xlim(1, x[-1])\n", "ax3.set_ylim(-0.5, 100.8)\n", "ax3.grid(linestyle=\":\")\n", "\n", "ax4.set_xlabel(f\"(d) Job Queuing Delay (s)\")\n", "ax4.set_ylabel(f\"CDF (%)\")\n", "ax4.set_xscale(\"log\")\n", "ax4.set_xticks([1e0, 1e1, 1e2, 1e3, 1e4])\n", "ax4.set_xlim(1, x2[-1])\n", "ax4.set_ylim(-0.5, 100.8)\n", "ax4.grid(linestyle=\":\")\n", "\n", "# 1 hour and 1 day\n", "ax1.axvline(x=3600, ls=\"--\", alpha=0.6, c=\"gray\", ymax=0.94, lw=1.5)\n", "ax1.axvline(x=3600 * 24, ls=\"--\", alpha=0.9, c=\"gray\", ymax=0.94, lw=1.5)\n", "ax3.axvline(x=3600, ls=\"--\", alpha=0.6, c=\"gray\", ymax=0.94, lw=1.5)\n", "ax3.axvline(x=3600 * 24, ls=\"--\", alpha=0.9, c=\"gray\", ymax=0.94, lw=1.5)\n", "\n", "sns.despine()\n", "ax1.text(0.78, 0.03, \"Seren\", transform=ax1.transAxes, size=20, fontweight=\"bold\")\n", "ax2.text(0.78, 0.03, \"Seren\", transform=ax2.transAxes, size=20, fontweight=\"bold\")\n", "ax3.text(0.78, 0.03, \"Kalos\", transform=ax3.transAxes, size=20, fontweight=\"bold\")\n", "ax4.text(0.78, 0.03, \"Kalos\", transform=ax4.transAxes, size=20, fontweight=\"bold\")\n", "\n", "fig.savefig(f\"{SAVEPATH}/cdf_job_duration_queue.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Box Plot: Request GPU number Different Type" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "cmap = sns.color_palette(\"pastel\")\n", "fig, (ax1, ax2) = plt.subplots(\n", " ncols=2,\n", " nrows=1,\n", " gridspec_kw={\"width_ratios\": [4.2, 3]},\n", " constrained_layout=True,\n", " figsize=(9, 3.75),\n", ")\n", "\n", "############ Fig 1 ############\n", "data_seren.sort_values(by=\"gpu_num\", ascending=False, inplace=True)\n", "data_seren[\"type\"].replace(\"SFT\", \"SFT\", inplace=True)\n", "\n", "x_ticks = [\n", " \"Eval\",\n", " \"Pretrain\",\n", " \"SFT\",\n", " \"MLLM\",\n", " \"Debug\",\n", " \"Other\",\n", "]\n", "\n", "flierprops = dict(marker=\".\", markerfacecolor=\"k\", markersize=2, linestyle=\"none\")\n", "sns.boxplot(\n", " x=\"type\",\n", " y=\"gpu_num\",\n", " data=data_seren,\n", " flierprops=flierprops,\n", " width=0.6,\n", " linewidth=2.2,\n", " saturation=2,\n", " palette=cmap,\n", " ax=ax1,\n", " order=x_ticks,\n", " boxprops=dict(alpha=1),\n", ")\n", "sns.color_palette(\"tab10\")\n", "ax1.set_xlabel(\"(a) Seren\")\n", "ax1.set_xticklabels(ax1.get_xticklabels(), rotation=0)\n", "ax1.set_ylabel(f\"Number of GPUs\")\n", "ax1.set_yscale(\"log\")\n", "ax1.grid(axis=\"y\", linestyle=\":\")\n", "\n", "\n", "############ Fig 2 ############\n", "data_kalos.sort_values(by=\"gpu_num\", ascending=False, inplace=True)\n", "data_kalos = data_kalos[data_kalos[\"type\"] != \"SFT\"]\n", "x_ticks_k = [\n", " \"Eval\",\n", " \"Pretrain\",\n", " \"Debug\",\n", " \"Other\",\n", "]\n", "my_pal = [cmap[0], cmap[1], cmap[4], cmap[5]]\n", "\n", "flierprops = dict(marker=\".\", markerfacecolor=\"k\", markersize=3, linestyle=\"none\")\n", "sns.boxplot(\n", " x=\"type\",\n", " y=\"gpu_num\",\n", " data=data_kalos,\n", " flierprops=flierprops,\n", " width=0.6,\n", " linewidth=2.2,\n", " saturation=2,\n", " palette=my_pal,\n", " ax=ax2,\n", " order=x_ticks_k,\n", " boxprops=dict(alpha=1),\n", ")\n", "sns.color_palette(\"tab10\")\n", "ax2.set_xlabel(\"(b) Kalos\")\n", "ax2.set_ylabel(None)\n", "ax2.set_xticklabels(ax2.get_xticklabels(), rotation=0)\n", "ax2.set_yscale(\"log\")\n", "ax2.grid(axis=\"y\", linestyle=\":\")\n", "\n", "sns.despine()\n", "fig.savefig(f\"{SAVEPATH}/box_gpu_num.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### CDF: Resource Utilization" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "with open(f\"{PKLPATH}/util_gpu_seren.pkl\", \"rb\") as file:\n", " _, _, x2, y2, x3, y3, x4, y4, x5, y5 = pickle.load(file)\n", "with open(f\"{PKLPATH}/util_gpu_kalos_full.pkl\", \"rb\") as file:\n", " _, _, x2_k, y2_k, x3_k, y3_k, x4_k, y4_k, x5_k, y5_k = pickle.load(file)\n", "with open(f\"{PKLPATH}/util_cpu_mem_seren.pkl\", \"rb\") as file:\n", " x6, y6, x7, y7 = pickle.load(file)\n", "with open(f\"{PKLPATH}/util_cpu_mem_kalos.pkl\", \"rb\") as file:\n", " x6_k, y6_k, x7_k, y7_k = pickle.load(file)\n", "with open(f\"{PKLPATH}/ib_seren.pkl\", \"rb\") as file:\n", " x8, y8, x9, y9 = pickle.load(file)\n", "\n", "x8 = x8 / x8.max() * 100\n", "x9 = x9 / x9.max() * 100\n", "\n", "linestyles = [\"--\", \":\", \"--\", \"-.\", \":\"]\n", "grid_params = dict(width_ratios=[1, 1])\n", "fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(ncols=2, nrows=2, constrained_layout=True, figsize=(9, 7))\n", "\n", "############ Fig 1: SM, Occupancy ############\n", "ax1.plot(x3, y3, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren SM Activity\")\n", "ax1.plot(x5, y5, linestyles[1], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren Occupancy\")\n", "ax1.plot(x3_k, y3_k, linestyles[0], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos SM Activity\")\n", "ax1.plot(x5_k, y5_k, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos Occupancy\")\n", "\n", "############ Fig 2: CPU mem usage, GPU mem usage ############\n", "ax2.plot(x7, y7, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren CPU Mem\")\n", "ax2.plot(x2, y2, linestyles[1], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren GPU Mem\")\n", "ax2.plot(x7_k, y7_k, linestyles[0], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos CPU Mem\")\n", "ax2.plot(x2_k, y2_k, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos GPU Mem\")\n", "\n", "############ Fig 3: CPU util ############\n", "ax3.plot(x6, y6, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren\")\n", "ax3.plot(x6_k, y6_k, linestyles[0], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos\")\n", "\n", "############ Fig 4: IB send, receive ############\n", "ax4.plot(x8, y8, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"IB Send\")\n", "ax4.plot(x9, y9, linestyles[1], linewidth=3, alpha=0.9, color=cmp[0], label=\"IB Receive\")\n", "\n", "ax1.set_xlabel(f\"(a) GPU DCGM Metric (%)\")\n", "ax1.set_ylabel(f\"CDF (%)\")\n", "ax1.set_xlim(-0.8, 100.8)\n", "ax1.set_ylim(0, 100.8)\n", "ax1.set_xticks([0, 25, 50, 75, 100])\n", "ax1.grid(linestyle=\":\")\n", "\n", "ax2.set_xlabel(f\"(b) Memory Footprint (%)\")\n", "ax2.set_ylabel(f\"CDF (%)\")\n", "ax2.set_xlim(-0.8, 100.8)\n", "ax2.set_xticks([0, 25, 50, 75, 100])\n", "ax2.set_ylim(0, 100.8)\n", "ax2.grid(linestyle=\":\")\n", "\n", "ax3.set_xlabel(f\"(c) CPU Utilization (%)\")\n", "ax3.set_ylabel(f\"CDF (%)\")\n", "ax3.set_xlim(-0.8, 100.8)\n", "ax3.set_xticks([0, 25, 50, 75, 100])\n", "ax3.set_ylim(0, 100.8)\n", "ax3.legend(loc=\"lower right\")\n", "ax3.grid(linestyle=\":\")\n", "\n", "ax4.set_xlabel(f\"(d) Network (%)\")\n", "ax4.set_ylabel(f\"CDF (%)\")\n", "ax4.set_xlim(-0.8, 100.8)\n", "ax4.set_xticks([0, 25, 50, 75, 100])\n", "ax4.set_ylim(0, 100.8)\n", "ax4.legend(loc=\"lower right\")\n", "ax4.grid(linestyle=\":\")\n", "sns.despine()\n", "\n", "\n", "S = mpatches.Patch(facecolor=cmp[0], alpha=0.9)\n", "K = mpatches.Patch(facecolor=cmp[1], alpha=0.9)\n", "A = (Line2D([0], [0], color=\"black\", lw=3, ls=\"--\"),)\n", "B = (Line2D([0], [0], color=\"black\", lw=3, ls=\":\"),)\n", "\n", "legend1 = ax1.legend([S, K], [\"Seren\", \"Kalos\"], bbox_to_anchor=(0.5, 0.62), loc=2, ncol=1, fontsize=17, frameon=False)\n", "\n", "ax1.add_artist(legend1)\n", "\n", "ax1.legend([A, B], [\"SM Activity\", \"TC Activity\"], bbox_to_anchor=(0.3, 0.36), loc=2, ncol=1)\n", "\n", "ax2.legend([A, B], [\"CPU Memory\", \"GPU Memory\"], bbox_to_anchor=(0.25, 0.36), loc=2, ncol=1)\n", "\n", "fig.savefig(f\"{SAVEPATH}/cdf_resource_util.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### CDF: Temperature" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# We use August data for GPU temperature and power\n", "with open(f\"{PKLPATH}/gpu_temp_seren.pkl\", \"rb\") as file:\n", " x, y1, x2, y2 = pickle.load(file)\n", "with open(f\"{PKLPATH}/gpu_temp_kalos.pkl\", \"rb\") as file:\n", " x1_k, y1_k, x2_k, y2_k = pickle.load(file)\n", "with open(f\"{PKLPATH}/gpu_power_seren.pkl\", \"rb\") as file:\n", " x3, y3 = pickle.load(file)\n", "with open(f\"{PKLPATH}/gpu_power_kalos.pkl\", \"rb\") as file:\n", " x3_k, y3_k = pickle.load(file)\n", "\n", "linestyles = [\"-\", \":\", \":\", \"-\"]\n", "fig, ax1 = plt.subplots(ncols=1, nrows=1, constrained_layout=True, figsize=(5, 3.75))\n", "\n", "############ Fig 1: Temperature ############\n", "ax1.plot(x, y1, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren GPU Temp\")\n", "ax1.plot(x2, y2, linestyles[1], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren GPU Mem Temp\")\n", "ax1.plot(x1_k, y1_k, linestyles[0], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos GPU Temp\")\n", "ax1.plot(x2_k, y2_k, linestyles[1], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos GPU Mem Temp\")\n", "\n", "ax1.set_xlabel(f\"Temperature (°C)\")\n", "ax1.set_ylabel(f\"CDF (%)\")\n", "ax1.set_xlim(20, 85)\n", "ax1.set_ylim(0, 100.8)\n", "ax1.grid(linestyle=\":\")\n", "\n", "S = mpatches.Patch(facecolor=cmp[0], alpha=0.9)\n", "K = mpatches.Patch(facecolor=cmp[1], alpha=0.9)\n", "A = (Line2D([0], [0], color=\"black\", lw=3, ls=\"-\"),)\n", "B = (Line2D([0], [0], color=\"black\", lw=3, ls=\":\"),)\n", "\n", "legend1 = ax1.legend([S, K], [\"Seren\", \"Kalos\"], bbox_to_anchor=(0.6, 0.62), loc=2, ncol=1, fontsize=17, frameon=False)\n", "\n", "ax1.add_artist(legend1)\n", "\n", "ax1.legend(\n", " [A, B],\n", " [\"GPU Temp.\", \"GMem Temp.\"],\n", " bbox_to_anchor=(0.4, 0.32),\n", " loc=2,\n", " ncol=1,\n", " fontsize=17,\n", ")\n", "\n", "sns.despine()\n", "fig.savefig(f\"{SAVEPATH}/cdf_temperature.pdf\", bbox_inches=\"tight\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### CDF: Power" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "with open(f\"{PKLPATH}/server_power.pkl\", \"rb\") as file:\n", " x1, y1, x2, y2 = pickle.load(file)\n", "with open(f\"{PKLPATH}/gpu_power_seren.pkl\", \"rb\") as file:\n", " x3, y3 = pickle.load(file)\n", "with open(f\"{PKLPATH}/gpu_power_kalos.pkl\", \"rb\") as file:\n", " x3_k, y3_k = pickle.load(file)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "linestyles = [\"--\", \":\", \":\", \"-\"]\n", "grid_params = dict(width_ratios=[1, 1])\n", "fig, (ax1, ax2) = plt.subplots(ncols=2, nrows=1, constrained_layout=True, figsize=(9, 3.75))\n", "\n", "############ Fig 1: GPU power ############\n", "ax1.plot(x3, y3, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"Seren\")\n", "ax1.plot(x3_k, y3_k, linestyles[0], linewidth=3, alpha=0.9, color=cmp[1], label=\"Kalos\")\n", "ax1.axvline(x=400, ls=\"--\", alpha=0.6, c=\"gray\", ymax=100, lw=1.5)\n", "ax1.annotate(\n", " \"A100 TDP\",\n", " xy=(400, 33),\n", " xytext=(420, 20),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=16,\n", ")\n", "ax1.annotate(\n", " \"Max=600\",\n", " xy=(600, 100),\n", " xytext=(430, 85),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=16,\n", ")\n", "\n", "############ Fig 2: Server power ############\n", "ax2.plot(x1, y1, linestyles[0], linewidth=3, alpha=0.9, color=cmp[0], label=\"GPU Node\")\n", "ax2.plot(x2, y2, linestyles[1], linewidth=3, alpha=0.9, color=cmp[0], label=\"CPU Node\")\n", "ax2.annotate(\n", " \"Max=960\",\n", " xy=(960, 100),\n", " xytext=(1200, 90),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=16,\n", ")\n", "ax2.annotate(\n", " \"Max=6550\",\n", " xy=(6550, 100),\n", " xytext=(4500, 70),\n", " arrowprops=dict(facecolor=\"black\", width=2.5, headwidth=8),\n", " color=\"black\",\n", " fontsize=16,\n", ")\n", "\n", "ax1.set_xlabel(f\"(a) GPU Power (W)\")\n", "ax1.set_ylabel(f\"CDF (%)\")\n", "ax1.set_xlim(-0.8, 610)\n", "ax1.set_ylim(0, 100.8)\n", "ax1.legend()\n", "ax1.grid(linestyle=\":\")\n", "ax1.xaxis.set_minor_locator(matplotlib.ticker.FixedLocator([60]))\n", "ax1.xaxis.set_minor_formatter(matplotlib.ticker.FixedFormatter([60]))\n", "ax1.tick_params(axis=\"x\", which=\"minor\", labelsize=15)\n", "\n", "ax2.set_xlabel(f\"(b) Server Power in Seren (W)\")\n", "ax2.set_ylabel(f\"CDF (%)\")\n", "ax2.set_xlim(-0.8, x1.max())\n", "ax2.set_ylim(0, 100.8)\n", "ax2.legend(loc=\"lower right\")\n", "ax2.grid(linestyle=\":\")\n", "ax2.xaxis.set_minor_locator(matplotlib.ticker.FixedLocator([520]))\n", "ax2.xaxis.set_minor_formatter(matplotlib.ticker.FixedFormatter([520]))\n", "ax2.tick_params(axis=\"x\", which=\"minor\", labelsize=15)\n", "sns.despine()\n", "\n", "fig.savefig(f\"{SAVEPATH}/cdf_power.pdf\", bbox_inches=\"tight\")" ] } ], "metadata": { "kernelspec": { "display_name": "base", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.9.16" }, "orig_nbformat": 4 }, "nbformat": 4, "nbformat_minor": 2 }