Cuda

.gitignore

@@ -1,3 +1,4 @@
 data
 build
 out.txt
+*.csv

Makefile

@@ -1,8 +1,8 @@
 CXX = mpic++
-CXXFLAGS = -std=c++17 -O2 -Wall -Wextra -Wno-cast-function-type
+CXXFLAGS = -std=c++17 -O2 -Wall -Wextra -Wno-cast-function-type -fopenmp
 
 NVCC = nvcc
-NVCCFLAGS = -O2 -Xcompiler -fPIC
+NVCCFLAGS = -O3 -std=c++17 -arch=sm_86 -Xcompiler -fPIC
 
 SRC_DIR = src
 BUILD_DIR = build

README.md

@@ -2,6 +2,16 @@
 
 [Kaggle Bitcoin Historical Data](https://www.kaggle.com/datasets/mczielinski/bitcoin-historical-data)
 
+Исходные данные хранят информацию по каждой минуте. Чтобы увеличить объём данных
+для более наглядной демонстрации эффективности параллельных вычислений
+и вычислений на GPU, с помощью линейной интерполяции данные были преобразованы
+из данных о каждой минуте в данные о каждых 10 секундах, то есть объём данных увеличился
+в 6 раз.
+
+```
+python3 upsample.py -i ./data/data.csv -o ./data/data_10s.csv -s 10
+```
+
 ## Задание
 
 Группируем данные по дням (Timestamp), за каждый день вычисляем среднюю цену 
@@ -13,7 +23,7 @@
 ## Сборка
 
 Проект обязательно должен быть расположен в общей директории для всех узлов,
-например, в `/mnt/shared/supercomputers/bitcoin-project/build`. 
+например, в `/mnt/shared/supercomputers/build`. 
 Перед запуском указать актуальный путь в `run.slurm`.
 
 ```sh

benchmark.py

@@ -0,0 +1,115 @@
+"""
+Запускает make run <number_of_runs> раз и считает статистику по времени выполнения.
+Тупо парсит out.txt и берём значение из строки "Total execution time: <time> sec".
+
+python benchmark.py <number_of_runs>
+"""
+
+import os
+import re
+import sys
+import time
+import subprocess
+import statistics
+
+N = int(sys.argv[1]) if len(sys.argv) > 1 else 10
+OUT = "out.txt"
+
+TIME_RE = re.compile(r"Total execution time:\s*([0-9]*\.?[0-9]+)\s*sec")
+JOB_RE  = re.compile(r"Submitted batch job\s+(\d+)")
+
+APPEAR_TIMEOUT = 300.0     # ждать появления out.txt
+FINISH_TIMEOUT = 3600.0    # ждать появления Total execution time (сек)
+POLL = 0.2                 # частота проверки файла
+
+def wait_for_exists(path: str, timeout: float):
+    t0 = time.time()
+    while not os.path.exists(path):
+        if time.time() - t0 > timeout:
+            raise TimeoutError(f"{path} did not appear within {timeout} seconds")
+        time.sleep(POLL)
+
+def try_read(path: str) -> str:
+    try:
+        with open(path, "r", encoding="utf-8", errors="replace") as f:
+            return f.read()
+    except FileNotFoundError:
+        return ""
+    except OSError:
+        # бывает, что файл на NFS в момент записи недоступен на чтение
+        return ""
+
+def wait_for_time_line(path: str, timeout: float) -> float:
+    t0 = time.time()
+    last_report = 0.0
+    while True:
+        txt = try_read(path)
+        matches = TIME_RE.findall(txt)
+        if matches:
+            return float(matches[-1])  # последняя встреченная строка
+
+        now = time.time()
+        if now - t0 > timeout:
+            tail = txt[-800:] if txt else "<empty>"
+            raise TimeoutError("Timed out waiting for 'Total execution time' line.\n"
+                               f"Last 800 chars of out.txt:\n{tail}")
+
+        # иногда полезно печатать прогресс раз в ~5 сек
+        if now - last_report > 5.0:
+            last_report = now
+            if txt:
+                # показать последнюю непустую строку
+                lines = [l for l in txt.splitlines() if l.strip()]
+                if lines:
+                    print(f"  waiting... last line: {lines[-1][:120]}", flush=True)
+                else:
+                    print("  waiting... (out.txt empty)", flush=True)
+            else:
+                print("  waiting... (out.txt not readable yet)", flush=True)
+
+        time.sleep(POLL)
+
+times = []
+
+for i in range(N):
+    print(f"Run {i+1}/{N} ...", flush=True)
+
+    # удаляем out.txt перед запуском
+    try:
+        os.remove(OUT)
+    except FileNotFoundError:
+        pass
+
+    # запускаем make run и забираем stdout (там будет Submitted batch job XXX)
+    res = subprocess.run(["make", "run"], capture_output=True, text=True)
+    out = (res.stdout or "") + "\n" + (res.stderr or "")
+
+    job_id = None
+    m = JOB_RE.search(out)
+    if m:
+        job_id = m.group(1)
+        print(f"  submitted job {job_id}", flush=True)
+    else:
+        print("  (job id not detected; will only watch out.txt)", flush=True)
+
+    # ждём появления out.txt и появления строки с Total execution time
+    wait_for_exists(OUT, APPEAR_TIMEOUT)
+    t = wait_for_time_line(OUT, FINISH_TIMEOUT)
+
+    times.append(t)
+    print(f"  time = {t:.3f} sec", flush=True)
+
+    # опционально удалить out.txt после парсинга
+    try:
+        os.remove(OUT)
+    except FileNotFoundError:
+        pass
+
+print("\n=== RESULTS ===")
+print(f"Runs:   {len(times)}")
+print(f"Mean:   {statistics.mean(times):.3f} sec")
+print(f"Median: {statistics.median(times):.3f} sec")
+print(f"Min:    {min(times):.3f} sec")
+print(f"Max:    {max(times):.3f} sec")
+if len(times) > 1:
+    print(f"Stddev: {statistics.stdev(times):.3f} sec")

data.ipynb

@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": 20,
    "id": "2acce44b",
    "metadata": {},
    "outputs": [],
@@ -12,7 +12,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 14,
+   "execution_count": 21,
    "id": "5ba70af7",
    "metadata": {},
    "outputs": [
@@ -111,7 +111,7 @@
        "7317758  0.410369  "
       ]
      },
-     "execution_count": 14,
+     "execution_count": 21,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -124,8 +124,8 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 19,
+   "execution_count": 23,
-   "id": "d4b22f3b",
+   "id": "3b320537",
    "metadata": {},
    "outputs": [
     {
@@ -150,67 +150,190 @@
        "    <tr style=\"text-align: right;\">\n",
        "      <th></th>\n",
        "      <th>Timestamp</th>\n",
-       "      <th>Low</th>\n",
-       "      <th>High</th>\n",
        "      <th>Open</th>\n",
+       "      <th>High</th>\n",
+       "      <th>Low</th>\n",
        "      <th>Close</th>\n",
+       "      <th>Volume</th>\n",
+       "      <th>Avg</th>\n",
        "    </tr>\n",
        "  </thead>\n",
        "  <tbody>\n",
        "    <tr>\n",
-       "      <th>5078</th>\n",
+       "      <th>7317754</th>\n",
-       "      <td>2025-11-26</td>\n",
+       "      <td>2025-11-30 23:55:00+00:00</td>\n",
-       "      <td>86304.0</td>\n",
+       "      <td>90405.0</td>\n",
-       "      <td>90646.0</td>\n",
+       "      <td>90452.0</td>\n",
-       "      <td>87331.0</td>\n",
+       "      <td>90403.0</td>\n",
-       "      <td>90477.0</td>\n",
+       "      <td>90452.0</td>\n",
+       "      <td>0.531700</td>\n",
+       "      <td>90427.5</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>5079</th>\n",
+       "      <th>7317755</th>\n",
-       "      <td>2025-11-27</td>\n",
+       "      <td>2025-11-30 23:56:00+00:00</td>\n",
-       "      <td>90091.0</td>\n",
+       "      <td>90452.0</td>\n",
-       "      <td>91926.0</td>\n",
+       "      <td>90481.0</td>\n",
-       "      <td>90476.0</td>\n",
+       "      <td>90420.0</td>\n",
-       "      <td>91325.0</td>\n",
+       "      <td>90420.0</td>\n",
+       "      <td>0.055547</td>\n",
+       "      <td>90450.5</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>5080</th>\n",
+       "      <th>7317756</th>\n",
-       "      <td>2025-11-28</td>\n",
+       "      <td>2025-11-30 23:57:00+00:00</td>\n",
-       "      <td>90233.0</td>\n",
+       "      <td>90412.0</td>\n",
-       "      <td>93091.0</td>\n",
+       "      <td>90458.0</td>\n",
-       "      <td>91326.0</td>\n",
+       "      <td>90396.0</td>\n",
-       "      <td>90913.0</td>\n",
+       "      <td>90435.0</td>\n",
+       "      <td>0.301931</td>\n",
+       "      <td>90427.0</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>5081</th>\n",
+       "      <th>7317757</th>\n",
-       "      <td>2025-11-29</td>\n",
+       "      <td>2025-11-30 23:58:00+00:00</td>\n",
-       "      <td>90216.0</td>\n",
+       "      <td>90428.0</td>\n",
-       "      <td>91179.0</td>\n",
+       "      <td>90428.0</td>\n",
-       "      <td>90913.0</td>\n",
+       "      <td>90362.0</td>\n",
-       "      <td>90832.0</td>\n",
+       "      <td>90362.0</td>\n",
-       "    </tr>\n",
+       "      <td>4.591653</td>\n",
-       "    <tr>\n",
+       "      <td>90395.0</td>\n",
-       "      <th>5082</th>\n",
+       "    </tr>\n",
-       "      <td>2025-11-30</td>\n",
+       "    <tr>\n",
+       "      <th>7317758</th>\n",
+       "      <td>2025-11-30 23:59:00+00:00</td>\n",
+       "      <td>90363.0</td>\n",
+       "      <td>90386.0</td>\n",
        "      <td>90362.0</td>\n",
-       "      <td>91969.0</td>\n",
-       "      <td>90832.0</td>\n",
        "      <td>90382.0</td>\n",
+       "      <td>0.410369</td>\n",
+       "      <td>90374.0</td>\n",
        "    </tr>\n",
        "  </tbody>\n",
        "</table>\n",
        "</div>"
       ],
       "text/plain": [
-       "       Timestamp      Low     High     Open    Close\n",
+       "                        Timestamp     Open     High      Low    Close  \\\n",
-       "5078  2025-11-26  86304.0  90646.0  87331.0  90477.0\n",
+       "7317754 2025-11-30 23:55:00+00:00  90405.0  90452.0  90403.0  90452.0   \n",
-       "5079  2025-11-27  90091.0  91926.0  90476.0  91325.0\n",
+       "7317755 2025-11-30 23:56:00+00:00  90452.0  90481.0  90420.0  90420.0   \n",
-       "5080  2025-11-28  90233.0  93091.0  91326.0  90913.0\n",
+       "7317756 2025-11-30 23:57:00+00:00  90412.0  90458.0  90396.0  90435.0   \n",
-       "5081  2025-11-29  90216.0  91179.0  90913.0  90832.0\n",
+       "7317757 2025-11-30 23:58:00+00:00  90428.0  90428.0  90362.0  90362.0   \n",
-       "5082  2025-11-30  90362.0  91969.0  90832.0  90382.0"
+       "7317758 2025-11-30 23:59:00+00:00  90363.0  90386.0  90362.0  90382.0   \n",
+       "\n",
+       "           Volume      Avg  \n",
+       "7317754  0.531700  90427.5  \n",
+       "7317755  0.055547  90450.5  \n",
+       "7317756  0.301931  90427.0  \n",
+       "7317757  4.591653  90395.0  \n",
+       "7317758  0.410369  90374.0  "
       ]
      },
-     "execution_count": 19,
+     "execution_count": 23,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "df['Avg'] = (df['Low'] + df['High']) / 2\n",
+    "df.tail()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 25,
+   "id": "4b1cd63c",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>Timestamp</th>\n",
+       "      <th>Avg</th>\n",
+       "      <th>OpenMin</th>\n",
+       "      <th>OpenMax</th>\n",
+       "      <th>CloseMin</th>\n",
+       "      <th>CloseMax</th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>5078</th>\n",
+       "      <td>2025-11-26</td>\n",
+       "      <td>88057.301736</td>\n",
+       "      <td>86312.0</td>\n",
+       "      <td>90574.0</td>\n",
+       "      <td>86323.0</td>\n",
+       "      <td>90574.0</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>5079</th>\n",
+       "      <td>2025-11-27</td>\n",
+       "      <td>91245.092708</td>\n",
+       "      <td>90126.0</td>\n",
+       "      <td>91888.0</td>\n",
+       "      <td>90126.0</td>\n",
+       "      <td>91925.0</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>5080</th>\n",
+       "      <td>2025-11-28</td>\n",
+       "      <td>91324.308681</td>\n",
+       "      <td>90255.0</td>\n",
+       "      <td>92970.0</td>\n",
+       "      <td>90283.0</td>\n",
+       "      <td>92966.0</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>5081</th>\n",
+       "      <td>2025-11-29</td>\n",
+       "      <td>90746.479514</td>\n",
+       "      <td>90265.0</td>\n",
+       "      <td>91158.0</td>\n",
+       "      <td>90279.0</td>\n",
+       "      <td>91179.0</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>5082</th>\n",
+       "      <td>2025-11-30</td>\n",
+       "      <td>91187.356250</td>\n",
+       "      <td>90363.0</td>\n",
+       "      <td>91940.0</td>\n",
+       "      <td>90362.0</td>\n",
+       "      <td>91940.0</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "</div>"
+      ],
+      "text/plain": [
+       "       Timestamp           Avg  OpenMin  OpenMax  CloseMin  CloseMax\n",
+       "5078  2025-11-26  88057.301736  86312.0  90574.0   86323.0   90574.0\n",
+       "5079  2025-11-27  91245.092708  90126.0  91888.0   90126.0   91925.0\n",
+       "5080  2025-11-28  91324.308681  90255.0  92970.0   90283.0   92966.0\n",
+       "5081  2025-11-29  90746.479514  90265.0  91158.0   90279.0   91179.0\n",
+       "5082  2025-11-30  91187.356250  90363.0  91940.0   90362.0   91940.0"
+      ]
+     },
+     "execution_count": 25,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -218,7 +341,13 @@
    "source": [
     "df_days = (\n",
     "    df.groupby(df[\"Timestamp\"].dt.date)\n",
-    "    .agg({\"Low\": \"min\", \"High\": \"max\", \"Open\": \"first\", \"Close\": \"last\"})\n",
+    "    .agg(\n",
+    "        Avg=(\"Avg\", \"mean\"),\n",
+    "        OpenMin=(\"Open\", \"min\"),\n",
+    "        OpenMax=(\"Open\", \"max\"),\n",
+    "        CloseMin=(\"Close\", \"min\"),\n",
+    "        CloseMax=(\"Close\", \"max\"),\n",
+    "    )\n",
     "    .reset_index()\n",
     ")\n",
     "df_days.tail()"
@@ -226,111 +355,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 21,
+   "execution_count": 26,
-   "id": "91823496",
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/html": [
-       "<div>\n",
-       "<style scoped>\n",
-       "    .dataframe tbody tr th:only-of-type {\n",
-       "        vertical-align: middle;\n",
-       "    }\n",
-       "\n",
-       "    .dataframe tbody tr th {\n",
-       "        vertical-align: top;\n",
-       "    }\n",
-       "\n",
-       "    .dataframe thead th {\n",
-       "        text-align: right;\n",
-       "    }\n",
-       "</style>\n",
-       "<table border=\"1\" class=\"dataframe\">\n",
-       "  <thead>\n",
-       "    <tr style=\"text-align: right;\">\n",
-       "      <th></th>\n",
-       "      <th>Timestamp</th>\n",
-       "      <th>Low</th>\n",
-       "      <th>High</th>\n",
-       "      <th>Open</th>\n",
-       "      <th>Close</th>\n",
-       "      <th>Avg</th>\n",
-       "    </tr>\n",
-       "  </thead>\n",
-       "  <tbody>\n",
-       "    <tr>\n",
-       "      <th>5078</th>\n",
-       "      <td>2025-11-26</td>\n",
-       "      <td>86304.0</td>\n",
-       "      <td>90646.0</td>\n",
-       "      <td>87331.0</td>\n",
-       "      <td>90477.0</td>\n",
-       "      <td>88475.0</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>5079</th>\n",
-       "      <td>2025-11-27</td>\n",
-       "      <td>90091.0</td>\n",
-       "      <td>91926.0</td>\n",
-       "      <td>90476.0</td>\n",
-       "      <td>91325.0</td>\n",
-       "      <td>91008.5</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>5080</th>\n",
-       "      <td>2025-11-28</td>\n",
-       "      <td>90233.0</td>\n",
-       "      <td>93091.0</td>\n",
-       "      <td>91326.0</td>\n",
-       "      <td>90913.0</td>\n",
-       "      <td>91662.0</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>5081</th>\n",
-       "      <td>2025-11-29</td>\n",
-       "      <td>90216.0</td>\n",
-       "      <td>91179.0</td>\n",
-       "      <td>90913.0</td>\n",
-       "      <td>90832.0</td>\n",
-       "      <td>90697.5</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>5082</th>\n",
-       "      <td>2025-11-30</td>\n",
-       "      <td>90362.0</td>\n",
-       "      <td>91969.0</td>\n",
-       "      <td>90832.0</td>\n",
-       "      <td>90382.0</td>\n",
-       "      <td>91165.5</td>\n",
-       "    </tr>\n",
-       "  </tbody>\n",
-       "</table>\n",
-       "</div>"
-      ],
-      "text/plain": [
-       "       Timestamp      Low     High     Open    Close      Avg\n",
-       "5078  2025-11-26  86304.0  90646.0  87331.0  90477.0  88475.0\n",
-       "5079  2025-11-27  90091.0  91926.0  90476.0  91325.0  91008.5\n",
-       "5080  2025-11-28  90233.0  93091.0  91326.0  90913.0  91662.0\n",
-       "5081  2025-11-29  90216.0  91179.0  90913.0  90832.0  90697.5\n",
-       "5082  2025-11-30  90362.0  91969.0  90832.0  90382.0  91165.5"
-      ]
-     },
-     "execution_count": 21,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "df_days[\"Avg\"] = (df_days[\"Low\"] + df_days[\"High\"]) / 2\n",
-    "df_days.tail()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 25,
    "id": "9a7b3310",
    "metadata": {},
    "outputs": [
@@ -358,6 +383,8 @@
        "      <th>start_date</th>\n",
        "      <th>end_date</th>\n",
        "      <th>min_open</th>\n",
+       "      <th>max_open</th>\n",
+       "      <th>min_close</th>\n",
        "      <th>max_close</th>\n",
        "      <th>start_avg</th>\n",
        "      <th>end_avg</th>\n",
@@ -366,76 +393,86 @@
        "  </thead>\n",
        "  <tbody>\n",
        "    <tr>\n",
-       "      <th>335</th>\n",
+       "      <th>316</th>\n",
        "      <td>2025-02-27</td>\n",
-       "      <td>2025-04-23</td>\n",
+       "      <td>2025-04-25</td>\n",
-       "      <td>76252.0</td>\n",
+       "      <td>74509.0</td>\n",
-       "      <td>94273.0</td>\n",
+       "      <td>95801.0</td>\n",
-       "      <td>84801.5</td>\n",
+       "      <td>74515.0</td>\n",
-       "      <td>93335.0</td>\n",
+       "      <td>95800.0</td>\n",
-       "      <td>0.100629</td>\n",
+       "      <td>85166.063889</td>\n",
+       "      <td>94303.907292</td>\n",
+       "      <td>0.107294</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>336</th>\n",
+       "      <th>317</th>\n",
-       "      <td>2025-04-24</td>\n",
+       "      <td>2025-04-26</td>\n",
-       "      <td>2025-05-09</td>\n",
+       "      <td>2025-05-11</td>\n",
-       "      <td>93730.0</td>\n",
+       "      <td>92877.0</td>\n",
-       "      <td>103261.0</td>\n",
+       "      <td>104971.0</td>\n",
-       "      <td>92867.5</td>\n",
+       "      <td>92872.0</td>\n",
-       "      <td>103341.0</td>\n",
+       "      <td>104965.0</td>\n",
-       "      <td>0.112779</td>\n",
+       "      <td>94500.950347</td>\n",
+       "      <td>104182.167708</td>\n",
+       "      <td>0.102446</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>337</th>\n",
+       "      <th>318</th>\n",
-       "      <td>2025-05-10</td>\n",
+       "      <td>2025-05-12</td>\n",
        "      <td>2025-07-11</td>\n",
-       "      <td>100990.0</td>\n",
+       "      <td>98384.0</td>\n",
-       "      <td>117579.0</td>\n",
+       "      <td>118833.0</td>\n",
-       "      <td>103915.0</td>\n",
+       "      <td>98382.0</td>\n",
-       "      <td>117032.5</td>\n",
+       "      <td>118839.0</td>\n",
-       "      <td>0.126233</td>\n",
+       "      <td>103569.791319</td>\n",
+       "      <td>117463.666667</td>\n",
+       "      <td>0.134150</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>338</th>\n",
+       "      <th>319</th>\n",
        "      <td>2025-07-12</td>\n",
        "      <td>2025-11-04</td>\n",
-       "      <td>106470.0</td>\n",
+       "      <td>98944.0</td>\n",
-       "      <td>124728.0</td>\n",
+       "      <td>126202.0</td>\n",
-       "      <td>117599.0</td>\n",
+       "      <td>98943.0</td>\n",
-       "      <td>103079.0</td>\n",
+       "      <td>126202.0</td>\n",
-       "      <td>0.123470</td>\n",
+       "      <td>117640.026389</td>\n",
+       "      <td>103712.985764</td>\n",
+       "      <td>0.118387</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>339</th>\n",
+       "      <th>320</th>\n",
        "      <td>2025-11-05</td>\n",
        "      <td>2025-11-18</td>\n",
-       "      <td>92112.0</td>\n",
+       "      <td>89291.0</td>\n",
-       "      <td>105972.0</td>\n",
+       "      <td>107343.0</td>\n",
-       "      <td>101737.5</td>\n",
+       "      <td>89286.0</td>\n",
-       "      <td>91471.0</td>\n",
+       "      <td>107343.0</td>\n",
-       "      <td>0.100912</td>\n",
+       "      <td>102514.621181</td>\n",
+       "      <td>91705.833333</td>\n",
+       "      <td>0.105437</td>\n",
        "    </tr>\n",
        "  </tbody>\n",
        "</table>\n",
        "</div>"
       ],
       "text/plain": [
-       "     start_date    end_date  min_open  max_close  start_avg   end_avg  \\\n",
+       "     start_date    end_date  min_open  max_open  min_close  max_close  \\\n",
-       "335  2025-02-27  2025-04-23   76252.0    94273.0    84801.5   93335.0   \n",
+       "316  2025-02-27  2025-04-25   74509.0   95801.0    74515.0    95800.0   \n",
-       "336  2025-04-24  2025-05-09   93730.0   103261.0    92867.5  103341.0   \n",
+       "317  2025-04-26  2025-05-11   92877.0  104971.0    92872.0   104965.0   \n",
-       "337  2025-05-10  2025-07-11  100990.0   117579.0   103915.0  117032.5   \n",
+       "318  2025-05-12  2025-07-11   98384.0  118833.0    98382.0   118839.0   \n",
-       "338  2025-07-12  2025-11-04  106470.0   124728.0   117599.0  103079.0   \n",
+       "319  2025-07-12  2025-11-04   98944.0  126202.0    98943.0   126202.0   \n",
-       "339  2025-11-05  2025-11-18   92112.0   105972.0   101737.5   91471.0   \n",
+       "320  2025-11-05  2025-11-18   89291.0  107343.0    89286.0   107343.0   \n",
        "\n",
-       "       change  \n",
+       "         start_avg        end_avg    change  \n",
-       "335  0.100629  \n",
+       "316   85166.063889   94303.907292  0.107294  \n",
-       "336  0.112779  \n",
+       "317   94500.950347  104182.167708  0.102446  \n",
-       "337  0.126233  \n",
+       "318  103569.791319  117463.666667  0.134150  \n",
-       "338  0.123470  \n",
+       "319  117640.026389  103712.985764  0.118387  \n",
-       "339  0.100912  "
+       "320  102514.621181   91705.833333  0.105437  "
       ]
      },
-     "execution_count": 25,
+     "execution_count": 26,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -455,8 +492,10 @@
     "        intervals.append({\n",
     "            \"start_date\": df_days.loc[start_idx, \"Timestamp\"],\n",
     "            \"end_date\": df_days.loc[i, \"Timestamp\"],\n",
-    "            \"min_open\": interval[\"Open\"].min(),\n",
+    "            \"min_open\": interval[\"OpenMin\"].min(),\n",
-    "            \"max_close\": interval[\"Close\"].max(),\n",
+    "            \"max_open\": interval[\"OpenMax\"].max(),\n",
+    "            \"min_close\": interval[\"CloseMin\"].min(),\n",
+    "            \"max_close\": interval[\"CloseMax\"].max(),\n",
     "            \"start_avg\": price_base,\n",
     "            \"end_avg\": price_now,\n",
     "            \"change\": change,\n",
@@ -470,6 +509,14 @@
     "df_intervals = pd.DataFrame(intervals)\n",
     "df_intervals.tail()"
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "07f1cd58",
+   "metadata": {},
+   "outputs": [],
+   "source": []
   }
  ],
  "metadata": {

run.slurm

@@ -2,9 +2,23 @@
 #SBATCH --job-name=btc
 #SBATCH --nodes=4
 #SBATCH --ntasks=4
+#SBATCH --cpus-per-task=2
 #SBATCH --output=out.txt
 
-# mpirun -np $SLURM_NTASKS ./build/bitcoin_app
+# Путь к файлу данных (должен существовать на всех узлах)
+export DATA_PATH="/mnt/shared/supercomputers/data/data_10s.csv"
+
+# Доли данных для каждого ранка (сумма определяет пропорции)
+export DATA_READ_SHARES="10,11,13,14"
+
+# Размер перекрытия в байтах для обработки границ строк
+export READ_OVERLAP_BYTES=131072
+
+# Интервал агрегации в секундах (60 = минуты, 600 = 10 минут, 86400 = дни)
+export AGGREGATION_INTERVAL=60
+
+# Использовать ли CUDA для агрегации (0 = нет, 1 = да)
+export USE_CUDA=1
 
 cd /mnt/shared/supercomputers/build
 mpirun -np $SLURM_NTASKS ./bitcoin_app

src/aggregation.cpp

@@ -0,0 +1,75 @@
+#include "aggregation.hpp"
+#include "utils.hpp"
+
+#include <algorithm>
+#include <cstdint>
+#include <limits>
+#include <vector>
+
+std::vector<PeriodStats> aggregate_periods(const std::vector<Record>& records) {
+    const int64_t interval = get_aggregation_interval();
+
+    std::vector<PeriodStats> result;
+    if (records.empty()) return result;
+
+    struct PeriodAccumulator {
+        double avg_sum = 0.0;
+        double open_min = std::numeric_limits<double>::max();
+        double open_max = std::numeric_limits<double>::lowest();
+        double close_min = std::numeric_limits<double>::max();
+        double close_max = std::numeric_limits<double>::lowest();
+        int64_t count = 0;
+
+        void add(const Record& r) {
+            const double avg = (r.low + r.high) / 2.0;
+            avg_sum += avg;
+            open_min = std::min(open_min, r.open);
+            open_max = std::max(open_max, r.open);
+            close_min = std::min(close_min, r.close);
+            close_max = std::max(close_max, r.close);
+            ++count;
+        }
+    };
+
+    PeriodIndex current_period =
+        static_cast<PeriodIndex>(records[0].timestamp) / interval;
+
+    PeriodAccumulator acc;
+    acc.add(records[0]);
+
+    for (size_t i = 1; i < records.size(); ++i) {
+        const Record& r = records[i];
+        const PeriodIndex period =
+            static_cast<PeriodIndex>(r.timestamp) / interval;
+
+        if (period != current_period) {
+            PeriodStats stats;
+            stats.period = current_period;
+            stats.avg = acc.avg_sum / static_cast<double>(acc.count);
+            stats.open_min = acc.open_min;
+            stats.open_max = acc.open_max;
+            stats.close_min = acc.close_min;
+            stats.close_max = acc.close_max;
+            stats.count = acc.count;
+            result.push_back(stats);
+
+            current_period = period;
+            acc = PeriodAccumulator{};
+        }
+
+        acc.add(r);
+    }
+
+    // последний период
+    PeriodStats stats;
+    stats.period = current_period;
+    stats.avg = acc.avg_sum / static_cast<double>(acc.count);
+    stats.open_min = acc.open_min;
+    stats.open_max = acc.open_max;
+    stats.close_min = acc.close_min;
+    stats.close_max = acc.close_max;
+    stats.count = acc.count;
+    result.push_back(stats);
+
+    return result;
+}

src/aggregation.hpp

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "record.hpp"
+#include "period_stats.hpp"
+#include <vector>
+
+// Агрегация записей по периодам на одном узле
+std::vector<PeriodStats> aggregate_periods(const std::vector<Record>& records);

src/csv_loader.cpp

@@ -2,45 +2,133 @@
 #include <fstream>
 #include <sstream>
 #include <iostream>
+#include <stdexcept>
 
-std::vector<Record> load_csv(const std::string& filename) {
+bool parse_csv_line(const std::string& line, Record& record) {
-    std::vector<Record> data;
+    if (line.empty()) {
-    std::ifstream file(filename);
+        return false;
-
-    if (!file.is_open()) {
-        throw std::runtime_error("Cannot open file: " + filename);
     }
     
-    std::string line;
-
-    // читаем первую строку (заголовок)
-    std::getline(file, line);
-
-    while (std::getline(file, line)) {
     std::stringstream ss(line);
     std::string item;
     
-        Record row;
+    try {
+        // timestamp
+        if (!std::getline(ss, item, ',') || item.empty()) return false;
+        record.timestamp = std::stod(item);
         
-        std::getline(ss, item, ',');
+        // open
-        row.timestamp = std::stod(item);
+        if (!std::getline(ss, item, ',') || item.empty()) return false;
+        record.open = std::stod(item);
         
-        std::getline(ss, item, ',');
+        // high
-        row.open = std::stod(item);
+        if (!std::getline(ss, item, ',') || item.empty()) return false;
+        record.high = std::stod(item);
         
-        std::getline(ss, item, ',');
+        // low
-        row.high = std::stod(item);
+        if (!std::getline(ss, item, ',') || item.empty()) return false;
+        record.low = std::stod(item);
         
-        std::getline(ss, item, ',');
+        // close
-        row.low = std::stod(item);
+        if (!std::getline(ss, item, ',') || item.empty()) return false;
+        record.close = std::stod(item);
         
-        std::getline(ss, item, ',');
+        // volume
-        row.close = std::stod(item);
+        if (!std::getline(ss, item, ',')) return false;
+        // Volume может быть пустым или содержать данные
+        if (item.empty()) {
+            record.volume = 0.0;
+        } else {
+            record.volume = std::stod(item);
+        }
         
-        std::getline(ss, item, ',');
+        return true;
-        row.volume = std::stod(item);
+    } catch (const std::exception&) {
+        return false;
+    }
+}
 
-        data.push_back(row);
+std::vector<Record> load_csv_parallel(int rank, int size) {
+    std::vector<Record> data;
+    
+    // Читаем настройки из переменных окружения
+    std::string data_path = get_data_path();
+    std::vector<int> shares = get_data_read_shares();
+    int64_t overlap_bytes = get_read_overlap_bytes();
+    
+    // Получаем размер файла
+    int64_t file_size = get_file_size(data_path);
+    
+    // Вычисляем диапазон байт для этого ранка
+    ByteRange range = calculate_byte_range(rank, size, file_size, shares, overlap_bytes);
+    
+    // Открываем файл и читаем нужный диапазон
+    std::ifstream file(data_path, std::ios::binary);
+    if (!file.is_open()) {
+        throw std::runtime_error("Cannot open file: " + data_path);
+    }
+    
+    // Переходим к началу диапазона
+    file.seekg(range.start);
+    
+    // Читаем данные в буфер
+    int64_t bytes_to_read = range.end - range.start;
+    std::vector<char> buffer(bytes_to_read);
+    file.read(buffer.data(), bytes_to_read);
+    int64_t bytes_read = file.gcount();
+    
+    file.close();
+    
+    // Преобразуем в строку для удобства парсинга
+    std::string content(buffer.data(), bytes_read);
+    
+    // Находим позицию начала первой полной строки
+    size_t parse_start = 0;
+    if (rank == 0) {
+        // Первый ранк: пропускаем заголовок (первую строку)
+        size_t header_end = content.find('\n');
+        if (header_end != std::string::npos) {
+            parse_start = header_end + 1;
+        }
+    } else {
+        // Остальные ранки: начинаем с первого \n (пропускаем неполную строку)
+        size_t first_newline = content.find('\n');
+        if (first_newline != std::string::npos) {
+            parse_start = first_newline + 1;
+        }
+    }
+    
+    // Находим позицию конца последней полной строки
+    size_t parse_end = content.size();
+    if (rank != size - 1) {
+        // Не последний ранк: ищем последний \n
+        size_t last_newline = content.rfind('\n');
+        if (last_newline != std::string::npos && last_newline > parse_start) {
+            parse_end = last_newline;
+        }
+    }
+    
+    // Парсим строки
+    size_t pos = parse_start;
+    while (pos < parse_end) {
+        size_t line_end = content.find('\n', pos);
+        if (line_end == std::string::npos || line_end > parse_end) {
+            line_end = parse_end;
+        }
+        
+        std::string line = content.substr(pos, line_end - pos);
+        
+        // Убираем \r если есть (Windows line endings)
+        if (!line.empty() && line.back() == '\r') {
+            line.pop_back();
+        }
+        
+        Record record;
+        if (parse_csv_line(line, record)) {
+            data.push_back(record);
+        }
+        
+        pos = line_end + 1;
     }
     
     return data;

src/csv_loader.hpp

@@ -2,5 +2,14 @@
 #include <string>
 #include <vector>
 #include "record.hpp"
+#include "utils.hpp"
 
-std::vector<Record> load_csv(const std::string& filename);
+// Параллельное чтение CSV файла для MPI
+// rank - номер текущего ранка
+// size - общее количество ранков
+// Возвращает вектор записей, прочитанных этим ранком
+std::vector<Record> load_csv_parallel(int rank, int size);
+
+// Парсинг одной строки CSV в Record
+// Возвращает true если парсинг успешен
+bool parse_csv_line(const std::string& line, Record& record);

src/gpu_loader.cpp

@@ -1,12 +1,133 @@
 #include "gpu_loader.hpp"
 #include <dlfcn.h>
+#include <iostream>
+#include <cstdint>
 
-gpu_is_available_fn load_gpu_is_available() {
+// Структура результата GPU (должна совпадать с gpu_plugin.cu)
-    void* h = dlopen("./libgpu_compute.so", RTLD_NOW | RTLD_LOCAL);
+struct GpuPeriodStats {
-    if (!h) return nullptr;
+    int64_t period;
+    double avg;
+    double open_min;
+    double open_max;
+    double close_min;
+    double close_max;
+    int64_t count;
+};
 
-    auto fn = (gpu_is_available_fn)dlsym(h, "gpu_is_available");
+// Типы функций из GPU плагина
-    if (!fn) return nullptr;
+using gpu_is_available_fn = int (*)();
 
-    return fn;
+using gpu_aggregate_periods_fn = int (*)(
+    const double* h_timestamps,
+    const double* h_open,
+    const double* h_high,
+    const double* h_low,
+    const double* h_close,
+    int num_ticks,
+    int64_t interval,
+    GpuPeriodStats** h_out_stats,
+    int* out_num_periods
+);
+
+using gpu_free_results_fn = void (*)(GpuPeriodStats*);
+
+static void* get_gpu_lib_handle() {
+    static void* h = dlopen("./libgpu_compute.so", RTLD_NOW | RTLD_LOCAL);
+    return h;
+}
+
+bool gpu_is_available() {
+    void* h = get_gpu_lib_handle();
+    if (!h) return false;
+    
+    auto fn = reinterpret_cast<gpu_is_available_fn>(dlsym(h, "gpu_is_available"));
+    if (!fn) return false;
+    
+    return fn() != 0;
+}
+
+bool aggregate_periods_gpu(
+    const std::vector<Record>& records,
+    int64_t aggregation_interval,
+    std::vector<PeriodStats>& out_stats)
+{
+    if (records.empty()) {
+        out_stats.clear();
+        return true;
+    }
+    
+    void* h = get_gpu_lib_handle();
+    if (!h) {
+        std::cerr << "GPU: Failed to load libgpu_compute.so" << std::endl;
+        return false;
+    }
+    
+    auto aggregate_fn = reinterpret_cast<gpu_aggregate_periods_fn>(
+        dlsym(h, "gpu_aggregate_periods"));
+    auto free_fn = reinterpret_cast<gpu_free_results_fn>(
+        dlsym(h, "gpu_free_results"));
+    
+    if (!aggregate_fn || !free_fn) {
+        std::cerr << "GPU: Failed to load functions from plugin" << std::endl;
+        return false;
+    }
+    
+    int num_ticks = static_cast<int>(records.size());
+    
+    // Конвертируем AoS в SoA
+    std::vector<double> timestamps(num_ticks);
+    std::vector<double> open(num_ticks);
+    std::vector<double> high(num_ticks);
+    std::vector<double> low(num_ticks);
+    std::vector<double> close(num_ticks);
+    
+    for (int i = 0; i < num_ticks; i++) {
+        timestamps[i] = records[i].timestamp;
+        open[i] = records[i].open;
+        high[i] = records[i].high;
+        low[i] = records[i].low;
+        close[i] = records[i].close;
+    }
+    
+    // Вызываем GPU функцию
+    GpuPeriodStats* gpu_stats = nullptr;
+    int num_periods = 0;
+    
+    int result = aggregate_fn(
+        timestamps.data(),
+        open.data(),
+        high.data(),
+        low.data(),
+        close.data(),
+        num_ticks,
+        aggregation_interval,
+        &gpu_stats,
+        &num_periods
+    );
+    
+    if (result != 0) {
+        std::cerr << "GPU: Aggregation failed with code " << result << std::endl;
+        return false;
+    }
+    
+    // Конвертируем результат в PeriodStats
+    out_stats.clear();
+    out_stats.reserve(num_periods);
+    
+    for (int i = 0; i < num_periods; i++) {
+        PeriodStats ps;
+        ps.period = gpu_stats[i].period;
+        ps.avg = gpu_stats[i].avg;
+        ps.open_min = gpu_stats[i].open_min;
+        ps.open_max = gpu_stats[i].open_max;
+        ps.close_min = gpu_stats[i].close_min;
+        ps.close_max = gpu_stats[i].close_max;
+        ps.count = gpu_stats[i].count;
+        out_stats.push_back(ps);
+    }
+    
+    // Освобождаем память
+    free_fn(gpu_stats);
+    
+    return true;
 }

src/gpu_loader.hpp

@@ -1,4 +1,15 @@
 #pragma once
-using gpu_is_available_fn = int (*)();
+#include "period_stats.hpp"
+#include "record.hpp"
+#include <vector>
 
-gpu_is_available_fn load_gpu_is_available();
+// Проверка доступности CUDA
+bool gpu_is_available();
+
+// Агрегация периодов на GPU
+// Возвращает true если успешно, false если GPU недоступен или ошибка
+bool aggregate_periods_gpu(
+    const std::vector<Record>& records,
+    int64_t aggregation_interval,
+    std::vector<PeriodStats>& out_stats
+);

src/gpu_plugin.cu

@@ -1,8 +1,430 @@
 #include <cuda_runtime.h>
+#include <cub/cub.cuh>
+#include <cstdint>
+#include <cfloat>
+#include <cstdio>
+#include <ctime>
+#include <string>
+#include <sstream>
+#include <iomanip>
+
+// ============================================================================
+// Структуры данных
+// ============================================================================
+
+// SoA (Structure of Arrays) для входных данных на GPU
+struct GpuTicksSoA {
+    double* timestamp;
+    double* open;
+    double* high;
+    double* low;
+    double* close;
+    int n;
+};
+
+// Результат агрегации одного периода
+struct GpuPeriodStats {
+    int64_t period;
+    double avg;
+    double open_min;
+    double open_max;
+    double close_min;
+    double close_max;
+    int64_t count;
+};
+
+// ============================================================================
+// Вспомогательные функции
+// ============================================================================
+
+static double get_time_ms() {
+    struct timespec ts;
+    clock_gettime(CLOCK_MONOTONIC, &ts);
+    return ts.tv_sec * 1000.0 + ts.tv_nsec / 1000000.0;
+}
+
+#define CUDA_CHECK(call) do { \
+    cudaError_t err = call; \
+    if (err != cudaSuccess) { \
+        printf("CUDA error at %s:%d: %s\n", __FILE__, __LINE__, cudaGetErrorString(err)); \
+        return -1; \
+    } \
+} while(0)
+
+// ============================================================================
+// Kernel: вычисление period_id для каждого тика
+// ============================================================================
+
+__global__ void compute_period_ids_kernel(
+    const double* __restrict__ timestamps,
+    int64_t* __restrict__ period_ids,
+    int n,
+    int64_t interval)
+{
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) {
+        period_ids[idx] = static_cast<int64_t>(timestamps[idx]) / interval;
+    }
+}
+
+// ============================================================================
+// Kernel: агрегация одного периода (один блок на период)
+// ============================================================================
+
+__global__ void aggregate_periods_kernel(
+    const double* __restrict__ open,
+    const double* __restrict__ high,
+    const double* __restrict__ low,
+    const double* __restrict__ close,
+    const int64_t* __restrict__ unique_periods,
+    const int* __restrict__ offsets,
+    const int* __restrict__ counts,
+    int num_periods,
+    GpuPeriodStats* __restrict__ out_stats)
+{
+    int period_idx = blockIdx.x;
+    if (period_idx >= num_periods) return;
+    
+    int offset = offsets[period_idx];
+    int count = counts[period_idx];
+    
+    // Используем shared memory для редукции внутри блока
+    __shared__ double s_avg_sum;
+    __shared__ double s_open_min;
+    __shared__ double s_open_max;
+    __shared__ double s_close_min;
+    __shared__ double s_close_max;
+    
+    // Инициализация shared memory первым потоком
+    if (threadIdx.x == 0) {
+        s_avg_sum = 0.0;
+        s_open_min = DBL_MAX;
+        s_open_max = -DBL_MAX;
+        s_close_min = DBL_MAX;
+        s_close_max = -DBL_MAX;
+    }
+    __syncthreads();
+    
+    // Локальные аккумуляторы для каждого потока
+    double local_avg_sum = 0.0;
+    double local_open_min = DBL_MAX;
+    double local_open_max = -DBL_MAX;
+    double local_close_min = DBL_MAX;
+    double local_close_max = -DBL_MAX;
+    
+    // Каждый поток обрабатывает свою часть тиков
+    for (int i = threadIdx.x; i < count; i += blockDim.x) {
+        int tick_idx = offset + i;
+        double avg = (low[tick_idx] + high[tick_idx]) / 2.0;
+        local_avg_sum += avg;
+        local_open_min = min(local_open_min, open[tick_idx]);
+        local_open_max = max(local_open_max, open[tick_idx]);
+        local_close_min = min(local_close_min, close[tick_idx]);
+        local_close_max = max(local_close_max, close[tick_idx]);
+    }
+    
+    // Редукция с использованием атомарных операций
+    atomicAdd(&s_avg_sum, local_avg_sum);
+    atomicMin(reinterpret_cast<unsigned long long*>(&s_open_min), 
+              __double_as_longlong(local_open_min));
+    atomicMax(reinterpret_cast<unsigned long long*>(&s_open_max),
+              __double_as_longlong(local_open_max));
+    atomicMin(reinterpret_cast<unsigned long long*>(&s_close_min),
+              __double_as_longlong(local_close_min));
+    atomicMax(reinterpret_cast<unsigned long long*>(&s_close_max),
+              __double_as_longlong(local_close_max));
+    
+    __syncthreads();
+    
+    // Первый поток записывает результат
+    if (threadIdx.x == 0) {
+        GpuPeriodStats stats;
+        stats.period = unique_periods[period_idx];
+        stats.avg = s_avg_sum / static_cast<double>(count);
+        stats.open_min = s_open_min;
+        stats.open_max = s_open_max;
+        stats.close_min = s_close_min;
+        stats.close_max = s_close_max;
+        stats.count = count;
+        out_stats[period_idx] = stats;
+    }
+}
+
+// ============================================================================
+// Простой kernel для агрегации (один поток на период)
+// Используется когда периодов много и тиков в каждом мало
+// ============================================================================
+
+__global__ void aggregate_periods_simple_kernel(
+    const double* __restrict__ open,
+    const double* __restrict__ high,
+    const double* __restrict__ low,
+    const double* __restrict__ close,
+    const int64_t* __restrict__ unique_periods,
+    const int* __restrict__ offsets,
+    const int* __restrict__ counts,
+    int num_periods,
+    GpuPeriodStats* __restrict__ out_stats)
+{
+    int period_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (period_idx >= num_periods) return;
+    
+    int offset = offsets[period_idx];
+    int count = counts[period_idx];
+    
+    double avg_sum = 0.0;
+    double open_min = DBL_MAX;
+    double open_max = -DBL_MAX;
+    double close_min = DBL_MAX;
+    double close_max = -DBL_MAX;
+    
+    for (int i = 0; i < count; i++) {
+        int tick_idx = offset + i;
+        double avg = (low[tick_idx] + high[tick_idx]) / 2.0;
+        avg_sum += avg;
+        open_min = min(open_min, open[tick_idx]);
+        open_max = max(open_max, open[tick_idx]);
+        close_min = min(close_min, close[tick_idx]);
+        close_max = max(close_max, close[tick_idx]);
+    }
+    
+    GpuPeriodStats stats;
+    stats.period = unique_periods[period_idx];
+    stats.avg = avg_sum / static_cast<double>(count);
+    stats.open_min = open_min;
+    stats.open_max = open_max;
+    stats.close_min = close_min;
+    stats.close_max = close_max;
+    stats.count = count;
+    out_stats[period_idx] = stats;
+}
+
+// ============================================================================
+// Проверка доступности GPU
+// ============================================================================
 
 extern "C" int gpu_is_available() {
     int n = 0;
     cudaError_t err = cudaGetDeviceCount(&n);
     if (err != cudaSuccess) return 0;
+    if (n > 0) {
+        cudaFree(0);  // Форсируем инициализацию контекста
+    }
     return (n > 0) ? 1 : 0;
 }
+
+// ============================================================================
+// Главная функция агрегации на GPU
+// ============================================================================
+
+extern "C" int gpu_aggregate_periods(
+    const double* h_timestamps,
+    const double* h_open,
+    const double* h_high,
+    const double* h_low,
+    const double* h_close,
+    int num_ticks,
+    int64_t interval,
+    GpuPeriodStats** h_out_stats,
+    int* out_num_periods)
+{
+    if (num_ticks == 0) {
+        *h_out_stats = nullptr;
+        *out_num_periods = 0;
+        return 0;
+    }
+    
+    std::ostringstream output;
+    double total_start = get_time_ms();
+    
+    // ========================================================================
+    // Шаг 1: Выделение памяти и копирование данных на GPU
+    // ========================================================================
+    double step1_start = get_time_ms();
+    
+    double* d_timestamps = nullptr;
+    double* d_open = nullptr;
+    double* d_high = nullptr;
+    double* d_low = nullptr;
+    double* d_close = nullptr;
+    int64_t* d_period_ids = nullptr;
+    
+    size_t ticks_bytes = num_ticks * sizeof(double);
+    
+    CUDA_CHECK(cudaMalloc(&d_timestamps, ticks_bytes));
+    CUDA_CHECK(cudaMalloc(&d_open, ticks_bytes));
+    CUDA_CHECK(cudaMalloc(&d_high, ticks_bytes));
+    CUDA_CHECK(cudaMalloc(&d_low, ticks_bytes));
+    CUDA_CHECK(cudaMalloc(&d_close, ticks_bytes));
+    CUDA_CHECK(cudaMalloc(&d_period_ids, num_ticks * sizeof(int64_t)));
+    
+    CUDA_CHECK(cudaMemcpy(d_timestamps, h_timestamps, ticks_bytes, cudaMemcpyHostToDevice));
+    CUDA_CHECK(cudaMemcpy(d_open, h_open, ticks_bytes, cudaMemcpyHostToDevice));
+    CUDA_CHECK(cudaMemcpy(d_high, h_high, ticks_bytes, cudaMemcpyHostToDevice));
+    CUDA_CHECK(cudaMemcpy(d_low, h_low, ticks_bytes, cudaMemcpyHostToDevice));
+    CUDA_CHECK(cudaMemcpy(d_close, h_close, ticks_bytes, cudaMemcpyHostToDevice));
+    
+    double step1_ms = get_time_ms() - step1_start;
+    
+    // ========================================================================
+    // Шаг 2: Вычисление period_id для каждого тика
+    // ========================================================================
+    double step2_start = get_time_ms();
+    
+    const int BLOCK_SIZE = 256;
+    int num_blocks = (num_ticks + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    
+    compute_period_ids_kernel<<<num_blocks, BLOCK_SIZE>>>(
+        d_timestamps, d_period_ids, num_ticks, interval);
+    CUDA_CHECK(cudaGetLastError());
+    CUDA_CHECK(cudaDeviceSynchronize());
+    
+    double step2_ms = get_time_ms() - step2_start;
+    
+    // ========================================================================
+    // Шаг 3: RLE (Run-Length Encode) для нахождения уникальных периодов
+    // ========================================================================
+    double step3_start = get_time_ms();
+    
+    int64_t* d_unique_periods = nullptr;
+    int* d_counts = nullptr;
+    int* d_num_runs = nullptr;
+    
+    CUDA_CHECK(cudaMalloc(&d_unique_periods, num_ticks * sizeof(int64_t)));
+    CUDA_CHECK(cudaMalloc(&d_counts, num_ticks * sizeof(int)));
+    CUDA_CHECK(cudaMalloc(&d_num_runs, sizeof(int)));
+    
+    // Определяем размер временного буфера для CUB
+    void* d_temp_storage = nullptr;
+    size_t temp_storage_bytes = 0;
+    
+    cub::DeviceRunLengthEncode::Encode(
+        d_temp_storage, temp_storage_bytes,
+        d_period_ids, d_unique_periods, d_counts, d_num_runs,
+        num_ticks);
+    
+    CUDA_CHECK(cudaMalloc(&d_temp_storage, temp_storage_bytes));
+    
+    cub::DeviceRunLengthEncode::Encode(
+        d_temp_storage, temp_storage_bytes,
+        d_period_ids, d_unique_periods, d_counts, d_num_runs,
+        num_ticks);
+    CUDA_CHECK(cudaGetLastError());
+    
+    // Копируем количество уникальных периодов
+    int num_periods = 0;
+    CUDA_CHECK(cudaMemcpy(&num_periods, d_num_runs, sizeof(int), cudaMemcpyDeviceToHost));
+    
+    cudaFree(d_temp_storage);
+    d_temp_storage = nullptr;
+    
+    double step3_ms = get_time_ms() - step3_start;
+    
+    // ========================================================================
+    // Шаг 4: Exclusive Scan для вычисления offsets
+    // ========================================================================
+    double step4_start = get_time_ms();
+    
+    int* d_offsets = nullptr;
+    CUDA_CHECK(cudaMalloc(&d_offsets, num_periods * sizeof(int)));
+    
+    temp_storage_bytes = 0;
+    cub::DeviceScan::ExclusiveSum(
+        d_temp_storage, temp_storage_bytes,
+        d_counts, d_offsets, num_periods);
+    
+    CUDA_CHECK(cudaMalloc(&d_temp_storage, temp_storage_bytes));
+    
+    cub::DeviceScan::ExclusiveSum(
+        d_temp_storage, temp_storage_bytes,
+        d_counts, d_offsets, num_periods);
+    CUDA_CHECK(cudaGetLastError());
+    
+    cudaFree(d_temp_storage);
+    
+    double step4_ms = get_time_ms() - step4_start;
+    
+    // ========================================================================
+    // Шаг 5: Агрегация периодов
+    // ========================================================================
+    double step5_start = get_time_ms();
+    
+    GpuPeriodStats* d_out_stats = nullptr;
+    CUDA_CHECK(cudaMalloc(&d_out_stats, num_periods * sizeof(GpuPeriodStats)));
+    
+    // Используем простой kernel (один поток на период)
+    // т.к. обычно тиков в периоде немного
+    int agg_blocks = (num_periods + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    
+    aggregate_periods_simple_kernel<<<agg_blocks, BLOCK_SIZE>>>(
+        d_open, d_high, d_low, d_close,
+        d_unique_periods, d_offsets, d_counts,
+        num_periods, d_out_stats);
+    CUDA_CHECK(cudaGetLastError());
+    CUDA_CHECK(cudaDeviceSynchronize());
+    
+    double step5_ms = get_time_ms() - step5_start;
+    
+    // ========================================================================
+    // Шаг 6: Копирование результатов на CPU
+    // ========================================================================
+    double step6_start = get_time_ms();
+    
+    GpuPeriodStats* h_stats = new GpuPeriodStats[num_periods];
+    CUDA_CHECK(cudaMemcpy(h_stats, d_out_stats, num_periods * sizeof(GpuPeriodStats), 
+                          cudaMemcpyDeviceToHost));
+    
+    double step6_ms = get_time_ms() - step6_start;
+    
+    // ========================================================================
+    // Шаг 7: Освобождение GPU памяти
+    // ========================================================================
+    double step7_start = get_time_ms();
+    
+    cudaFree(d_timestamps);
+    cudaFree(d_open);
+    cudaFree(d_high);
+    cudaFree(d_low);
+    cudaFree(d_close);
+    cudaFree(d_period_ids);
+    cudaFree(d_unique_periods);
+    cudaFree(d_counts);
+    cudaFree(d_offsets);
+    cudaFree(d_num_runs);
+    cudaFree(d_out_stats);
+    
+    double step7_ms = get_time_ms() - step7_start;
+    
+    // ========================================================================
+    // Итого
+    // ========================================================================
+    double total_ms = get_time_ms() - total_start;
+    
+    // Формируем весь вывод одной строкой
+    output << "  GPU aggregation (" << num_ticks << " ticks, interval=" << interval << " sec):\n";
+    output << "    1. Malloc + H->D copy:  " << std::fixed << std::setprecision(3) << std::setw(7) << step1_ms << " ms\n";
+    output << "    2. Compute period_ids:  " << std::setw(7) << step2_ms << " ms\n";
+    output << "    3. RLE (CUB):           " << std::setw(7) << step3_ms << " ms (" << num_periods << " periods)\n";
+    output << "    4. Exclusive scan:      " << std::setw(7) << step4_ms << " ms\n";
+    output << "    5. Aggregation kernel:  " << std::setw(7) << step5_ms << " ms\n";
+    output << "    6. D->H copy:           " << std::setw(7) << step6_ms << " ms\n";
+    output << "    7. Free GPU memory:     " << std::setw(7) << step7_ms << " ms\n";
+    output << "    GPU TOTAL:              " << std::setw(7) << total_ms << " ms\n";
+    
+    // Выводим всё одним принтом
+    printf("%s", output.str().c_str());
+    fflush(stdout);
+    
+    *h_out_stats = h_stats;
+    *out_num_periods = num_periods;
+    
+    return 0;
+}
+
+// ============================================================================
+// Освобождение памяти результатов
+// ============================================================================
+
+extern "C" void gpu_free_results(GpuPeriodStats* stats) {
+    delete[] stats;
+}

src/intervals.cpp

@@ -0,0 +1,320 @@
+#include "intervals.hpp"
+#include "utils.hpp"
+#include <mpi.h>
+#include <algorithm>
+#include <cmath>
+#include <fstream>
+#include <iomanip>
+#include <sstream>
+#include <ctime>
+#include <limits>
+
+// Вспомогательная структура для накопления min/max в интервале
+struct IntervalAccumulator {
+    PeriodIndex start_period;
+    double start_avg;
+    double open_min;
+    double open_max;
+    double close_min;
+    double close_max;
+    
+    void init(const PeriodStats& p) {
+        start_period = p.period;
+        start_avg = p.avg;
+        open_min = p.open_min;
+        open_max = p.open_max;
+        close_min = p.close_min;
+        close_max = p.close_max;
+    }
+    
+    void update(const PeriodStats& p) {
+        open_min = std::min(open_min, p.open_min);
+        open_max = std::max(open_max, p.open_max);
+        close_min = std::min(close_min, p.close_min);
+        close_max = std::max(close_max, p.close_max);
+    }
+    
+    Interval finalize(const PeriodStats& end_period, double change) const {
+        Interval iv;
+        iv.start_period = start_period;
+        iv.end_period = end_period.period;
+        iv.start_avg = start_avg;
+        iv.end_avg = end_period.avg;
+        iv.change = change;
+        iv.open_min = std::min(open_min, end_period.open_min);
+        iv.open_max = std::max(open_max, end_period.open_max);
+        iv.close_min = std::min(close_min, end_period.close_min);
+        iv.close_max = std::max(close_max, end_period.close_max);
+        return iv;
+    }
+};
+
+// Упакованная структура PeriodStats для MPI передачи (8 doubles)
+struct PackedPeriodStats {
+    double period;   // PeriodIndex as double
+    double avg;
+    double open_min;
+    double open_max;
+    double close_min;
+    double close_max;
+    double count;    // int64_t as double
+    double valid;    // флаг валидности (1.0 = valid, 0.0 = invalid)
+    
+    void pack(const PeriodStats& ps) {
+        period = static_cast<double>(ps.period);
+        avg = ps.avg;
+        open_min = ps.open_min;
+        open_max = ps.open_max;
+        close_min = ps.close_min;
+        close_max = ps.close_max;
+        count = static_cast<double>(ps.count);
+        valid = 1.0;
+    }
+    
+    PeriodStats unpack() const {
+        PeriodStats ps;
+        ps.period = static_cast<PeriodIndex>(period);
+        ps.avg = avg;
+        ps.open_min = open_min;
+        ps.open_max = open_max;
+        ps.close_min = close_min;
+        ps.close_max = close_max;
+        ps.count = static_cast<int64_t>(count);
+        return ps;
+    }
+    
+    bool is_valid() const { return valid > 0.5; }
+    void set_invalid() { valid = 0.0; }
+};
+
+IntervalResult find_intervals_parallel(
+    const std::vector<PeriodStats>& periods,
+    int rank, int size,
+    double threshold)
+{
+    IntervalResult result;
+    result.compute_time = 0.0;
+    result.wait_time = 0.0;
+    
+    if (periods.empty()) {
+        if (rank < size - 1) {
+            PackedPeriodStats invalid;
+            invalid.set_invalid();
+            MPI_Send(&invalid, 8, MPI_DOUBLE, rank + 1, 0, MPI_COMM_WORLD);
+        }
+        return result;
+    }
+    
+    double compute_start = MPI_Wtime();
+    
+    size_t process_until = (rank == size - 1) ? periods.size() : periods.size() - 1;
+    
+    IntervalAccumulator acc;
+    size_t start_idx = 0;
+    bool have_pending_interval = false;
+    
+    if (rank > 0) {
+        double wait_start = MPI_Wtime();
+        
+        PackedPeriodStats received;
+        MPI_Recv(&received, 8, MPI_DOUBLE, rank - 1, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+        
+        result.wait_time = MPI_Wtime() - wait_start;
+        compute_start = MPI_Wtime();
+        
+        if (received.is_valid()) {
+            PeriodStats prev_period = received.unpack();
+            
+            for (start_idx = 0; start_idx < periods.size(); start_idx++) {
+                if (periods[start_idx].period > prev_period.period) {
+                    break;
+                }
+            }
+            
+            if (start_idx < process_until) {
+                acc.init(prev_period);
+                have_pending_interval = true;
+                
+                for (size_t i = start_idx; i < process_until; i++) {
+                    acc.update(periods[i]);
+                    
+                    double change = std::abs(periods[i].avg - acc.start_avg) / acc.start_avg;
+                    
+                    if (change >= threshold) {
+                        result.intervals.push_back(acc.finalize(periods[i], change));
+                        have_pending_interval = false;
+                        
+                        start_idx = i + 1;
+                        if (start_idx < process_until) {
+                            acc.init(periods[start_idx]);
+                            have_pending_interval = true;
+                        }
+                    }
+                }
+            }
+        } else {
+            if (process_until > 0) {
+                acc.init(periods[0]);
+                have_pending_interval = true;
+                start_idx = 0;
+            }
+        }
+    } else {
+        if (process_until > 0) {
+            acc.init(periods[0]);
+            have_pending_interval = true;
+            start_idx = 0;
+        }
+    }
+    
+    if (rank == 0 && have_pending_interval) {
+        for (size_t i = 1; i < process_until; i++) {
+            acc.update(periods[i]);
+            
+            double change = std::abs(periods[i].avg - acc.start_avg) / acc.start_avg;
+            
+            if (change >= threshold) {
+                result.intervals.push_back(acc.finalize(periods[i], change));
+                have_pending_interval = false;
+                
+                start_idx = i + 1;
+                if (start_idx < process_until) {
+                    acc.init(periods[start_idx]);
+                    have_pending_interval = true;
+                }
+            }
+        }
+    }
+    
+    if (rank == size - 1 && have_pending_interval && !periods.empty()) {
+        const auto& last_period = periods.back();
+        double change = std::abs(last_period.avg - acc.start_avg) / acc.start_avg;
+        result.intervals.push_back(acc.finalize(last_period, change));
+    }
+    
+    result.compute_time = MPI_Wtime() - compute_start;
+    
+    if (rank < size - 1) {
+        PackedPeriodStats to_send;
+        
+        if (have_pending_interval) {
+            PeriodStats start_period;
+            start_period.period = acc.start_period;
+            start_period.avg = acc.start_avg;
+            start_period.open_min = acc.open_min;
+            start_period.open_max = acc.open_max;
+            start_period.close_min = acc.close_min;
+            start_period.close_max = acc.close_max;
+            start_period.count = 0;
+            to_send.pack(start_period);
+        } else if (periods.size() >= 2) {
+            to_send.pack(periods[periods.size() - 2]);
+        } else {
+            to_send.set_invalid();
+        }
+        
+        MPI_Send(&to_send, 8, MPI_DOUBLE, rank + 1, 0, MPI_COMM_WORLD);
+    }
+    
+    return result;
+}
+
+double collect_intervals(
+    std::vector<Interval>& local_intervals,
+    int rank, int size)
+{
+    double wait_time = 0.0;
+    
+    if (rank == 0) {
+        for (int r = 1; r < size; r++) {
+            double wait_start = MPI_Wtime();
+            
+            int count;
+            MPI_Recv(&count, 1, MPI_INT, r, 1, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+            
+            if (count > 0) {
+                std::vector<double> buffer(count * 9);
+                MPI_Recv(buffer.data(), count * 9, MPI_DOUBLE, r, 2, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+                
+                for (int i = 0; i < count; i++) {
+                    Interval iv;
+                    iv.start_period = static_cast<PeriodIndex>(buffer[i * 9 + 0]);
+                    iv.end_period = static_cast<PeriodIndex>(buffer[i * 9 + 1]);
+                    iv.open_min = buffer[i * 9 + 2];
+                    iv.open_max = buffer[i * 9 + 3];
+                    iv.close_min = buffer[i * 9 + 4];
+                    iv.close_max = buffer[i * 9 + 5];
+                    iv.start_avg = buffer[i * 9 + 6];
+                    iv.end_avg = buffer[i * 9 + 7];
+                    iv.change = buffer[i * 9 + 8];
+                    local_intervals.push_back(iv);
+                }
+            }
+            
+            wait_time += MPI_Wtime() - wait_start;
+        }
+        
+        std::sort(local_intervals.begin(), local_intervals.end(),
+            [](const Interval& a, const Interval& b) {
+                return a.start_period < b.start_period;
+            });
+    } else {
+        int count = static_cast<int>(local_intervals.size());
+        MPI_Send(&count, 1, MPI_INT, 0, 1, MPI_COMM_WORLD);
+        
+        if (count > 0) {
+            std::vector<double> buffer(count * 9);
+            for (int i = 0; i < count; i++) {
+                const auto& iv = local_intervals[i];
+                buffer[i * 9 + 0] = static_cast<double>(iv.start_period);
+                buffer[i * 9 + 1] = static_cast<double>(iv.end_period);
+                buffer[i * 9 + 2] = iv.open_min;
+                buffer[i * 9 + 3] = iv.open_max;
+                buffer[i * 9 + 4] = iv.close_min;
+                buffer[i * 9 + 5] = iv.close_max;
+                buffer[i * 9 + 6] = iv.start_avg;
+                buffer[i * 9 + 7] = iv.end_avg;
+                buffer[i * 9 + 8] = iv.change;
+            }
+            MPI_Send(buffer.data(), count * 9, MPI_DOUBLE, 0, 2, MPI_COMM_WORLD);
+        }
+    }
+    
+    return wait_time;
+}
+
+std::string period_index_to_datetime(PeriodIndex period) {
+    int64_t interval = get_aggregation_interval();
+    time_t ts = static_cast<time_t>(period) * interval;
+    struct tm* tm_info = gmtime(&ts);
+    
+    std::ostringstream oss;
+    oss << std::setfill('0') 
+        << (tm_info->tm_year + 1900) << "-"
+        << std::setw(2) << (tm_info->tm_mon + 1) << "-"
+        << std::setw(2) << tm_info->tm_mday << " "
+        << std::setw(2) << tm_info->tm_hour << ":"
+        << std::setw(2) << tm_info->tm_min << ":"
+        << std::setw(2) << tm_info->tm_sec;
+    
+    return oss.str();
+}
+
+void write_intervals(const std::string& filename, const std::vector<Interval>& intervals) {
+    std::ofstream out(filename);
+    
+    out << std::fixed << std::setprecision(2);
+    out << "start_datetime,end_datetime,open_min,open_max,close_min,close_max,start_avg,end_avg,change\n";
+    
+    for (const auto& iv : intervals) {
+        out << period_index_to_datetime(iv.start_period) << ","
+            << period_index_to_datetime(iv.end_period) << ","
+            << iv.open_min << ","
+            << iv.open_max << ","
+            << iv.close_min << ","
+            << iv.close_max << ","
+            << iv.start_avg << ","
+            << iv.end_avg << ","
+            << std::setprecision(6) << iv.change << "\n";
+    }
+}

src/intervals.hpp

@@ -0,0 +1,44 @@
+#pragma once
+
+#include "period_stats.hpp"
+#include <vector>
+#include <string>
+
+// Интервал с изменением >= threshold
+struct Interval {
+    PeriodIndex start_period;
+    PeriodIndex end_period;
+    double open_min;
+    double open_max;
+    double close_min;
+    double close_max;
+    double start_avg;
+    double end_avg;
+    double change;
+};
+
+// Результат параллельного построения интервалов
+struct IntervalResult {
+    std::vector<Interval> intervals;
+    double compute_time;  // время вычислений
+    double wait_time;     // время ожидания данных от предыдущего ранка
+};
+
+// Параллельное построение интервалов с использованием MPI
+IntervalResult find_intervals_parallel(
+    const std::vector<PeriodStats>& periods,
+    int rank, int size,
+    double threshold = 0.10
+);
+
+// Сбор интервалов со всех ранков на ранк 0
+double collect_intervals(
+    std::vector<Interval>& local_intervals,
+    int rank, int size
+);
+
+// Вывод интервалов в файл
+void write_intervals(const std::string& filename, const std::vector<Interval>& intervals);
+
+// Преобразование PeriodIndex в строку даты/времени
+std::string period_index_to_datetime(PeriodIndex period);

src/main.cpp

@@ -1,87 +1,115 @@
 #include <mpi.h>
 #include <iostream>
 #include <vector>
-#include <map>
+#include <iomanip>
 
 #include "csv_loader.hpp"
-#include "utils.hpp"
 #include "record.hpp"
+#include "period_stats.hpp"
+#include "aggregation.hpp"
+#include "intervals.hpp"
+#include "utils.hpp"
 #include "gpu_loader.hpp"
 
-// Функция: отобрать записи для конкретного ранга
-std::vector<Record> select_records_for_rank(
-    const std::map<long long, std::vector<Record>>& days,
-    const std::vector<long long>& day_list)
-{
-    std::vector<Record> out;
-    for (auto d : day_list) {
-        auto it = days.find(d);
-        if (it != days.end()) {
-            const auto& vec = it->second;
-            out.insert(out.end(), vec.begin(), vec.end());
-        }
-    }
-    return out;
-}
-
 int main(int argc, char** argv) {
     MPI_Init(&argc, &argv);
+    double total_start = MPI_Wtime();
 
     int rank, size;
     MPI_Comm_rank(MPI_COMM_WORLD, &rank);
     MPI_Comm_size(MPI_COMM_WORLD, &size);
 
-    std::vector<Record> local_records;
+    // Проверяем доступность GPU
+    bool use_cuda = get_use_cuda();
+    bool have_gpu = gpu_is_available();
+    bool use_gpu = use_cuda && have_gpu;
+    
+    std::cout << "Rank " << rank 
+              << ": USE_CUDA=" << use_cuda 
+              << ", GPU available=" << have_gpu 
+              << ", using " << (use_gpu ? "GPU" : "CPU")
+              << std::endl;
+
+    // Параллельное чтение данных
+    double read_start = MPI_Wtime();
+    std::vector<Record> records = load_csv_parallel(rank, size);
+    double read_time = MPI_Wtime() - read_start;
+
+    std::cout << "Rank " << rank 
+              << ": read " << records.size() << " records"
+              << " in " << std::fixed << std::setprecision(3) << read_time << " sec"
+              << std::endl;
+
+    // Агрегация по периодам
+    double agg_start = MPI_Wtime();
+    std::vector<PeriodStats> periods;
+    
+    if (use_gpu) {
+        int64_t interval = get_aggregation_interval();
+        if (!aggregate_periods_gpu(records, interval, periods)) {
+            std::cerr << "Rank " << rank << ": GPU aggregation failed, falling back to CPU" << std::endl;
+            periods = aggregate_periods(records);
+        }
+    } else {
+        periods = aggregate_periods(records);
+    }
+    
+    double agg_time = MPI_Wtime() - agg_start;
+
+    std::cout << "Rank " << rank 
+              << ": aggregated " << periods.size() << " periods"
+              << " [" << (periods.empty() ? 0 : periods.front().period) 
+              << ".." << (periods.empty() ? 0 : periods.back().period) << "]"
+              << " in " << std::fixed << std::setprecision(3) << agg_time << " sec"
+              << std::endl;
+
+    // Удаляем крайние периоды (могут быть неполными из-за параллельного чтения)
+    trim_edge_periods(periods, rank, size);
+
+    std::cout << "Rank " << rank 
+              << ": after trim " << periods.size() << " periods"
+              << " [" << (periods.empty() ? 0 : periods.front().period) 
+              << ".." << (periods.empty() ? 0 : periods.back().period) << "]"
+              << std::endl;
+
+    // Параллельное построение интервалов
+    IntervalResult iv_result = find_intervals_parallel(periods, rank, size);
+
+    std::cout << "Rank " << rank 
+              << ": found " << iv_result.intervals.size() << " intervals"
+              << ", compute " << std::fixed << std::setprecision(6) << iv_result.compute_time << " sec"
+              << ", wait " << iv_result.wait_time << " sec"
+              << std::endl;
+
+    // Сбор интервалов на ранке 0
+    double collect_wait = collect_intervals(iv_result.intervals, rank, size);
 
     if (rank == 0) {
-        std::cout << "Rank 0 loading CSV..." << std::endl;
+        std::cout << "Rank 0: collected " << iv_result.intervals.size() << " total intervals"
-
+                  << ", wait " << std::fixed << std::setprecision(3) << collect_wait << " sec"
-        // Запускаем из build
+                  << std::endl;
-        auto records = load_csv("../data/data.csv");
-
-        auto days = group_by_day(records);
-        auto parts = split_days(days, size);
-
-        // Рассылаем данные
-        for (int r = 0; r < size; r++) {
-            auto vec = select_records_for_rank(days, parts[r]);
-
-            if (r == 0) {  
-                // себе не отправляем — сразу сохраняем
-                local_records = vec;
-                continue;
     }
 
-            int count = vec.size();
+    // Запись результатов в файл (только ранк 0)
-            MPI_Send(&count, 1, MPI_INT, r, 0, MPI_COMM_WORLD);
+    if (rank == 0) {
-            MPI_Send(vec.data(), count * sizeof(Record), MPI_BYTE, r, 1, MPI_COMM_WORLD);
+        double write_start = MPI_Wtime();
-        }
+        write_intervals("result.csv", iv_result.intervals);
-    }
+        double write_time = MPI_Wtime() - write_start;
-    else {
-        // Принимает данные
-        int count = 0;
-        MPI_Recv(&count, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
 
-        local_records.resize(count);
+        std::cout << "Rank 0: wrote result.csv"
-        MPI_Recv(local_records.data(), count * sizeof(Record),
+                  << " in " << std::fixed << std::setprecision(3) << write_time << " sec"
-                 MPI_BYTE, 0, 1, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+                  << std::endl;
     }
 
+    // Вывод общего времени выполнения
     MPI_Barrier(MPI_COMM_WORLD);
-
+    double total_time = MPI_Wtime() - total_start;
-    std::cout << "Rank " << rank << " received "
+    if (rank == 0) {
-              << local_records.size() << " records" << std::endl;
+        std::cout << "Total execution time: "
-
+                  << std::fixed << std::setprecision(3)
-
+                  << total_time << " sec" << std::endl;
-    auto gpu_is_available = load_gpu_is_available();
-    int have_gpu = 0;
-    if (gpu_is_available) {
-        std::cout << "Rank " << rank << " dll loaded" << std::endl;
-        have_gpu = gpu_is_available();
     }
 
-    std::cout << "Rank " << rank << ": gpu_available=" << have_gpu << "\n";
-
     MPI_Finalize();
     return 0;
 }

src/mpi_utils.cpp

@@ -1,11 +0,0 @@
-#include "mpi_utils.hpp"
-#include <mpi.h>
-#include <iostream>
-
-void mpi_print_basic() {
-    int rank, size;
-    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
-    MPI_Comm_size(MPI_COMM_WORLD, &size);
-
-    std::cout << "Hello from rank " << rank << " of " << size << std::endl;
-}

src/mpi_utils.hpp

@@ -1,2 +0,0 @@
-#pragma once
-void mpi_print_basic();

src/period_stats.hpp

@@ -0,0 +1,15 @@
+#pragma once
+#include <cstdint>
+
+using PeriodIndex = int64_t;
+
+// Агрегированные данные за один период
+struct PeriodStats {
+    PeriodIndex period;   // индекс периода (timestamp / AGGREGATION_INTERVAL)
+    double avg;           // среднее значение (Low + High) / 2 по всем записям
+    double open_min;      // минимальный Open за период
+    double open_max;      // максимальный Open за период
+    double close_min;     // минимальный Close за период
+    double close_max;     // максимальный Close за период
+    int64_t count;        // количество записей, по которым агрегировали
+};

src/utils.cpp

@@ -1,25 +1,105 @@
 #include "utils.hpp"
+#include <fstream>
+#include <sstream>
+#include <stdexcept>
+#include <numeric>
 
-std::map<DayIndex, std::vector<Record>> group_by_day(const std::vector<Record>& recs) {
+int get_num_cpu_threads() {
-    std::map<DayIndex, std::vector<Record>> days;
+    const char* env_threads = std::getenv("NUM_CPU_THREADS");
-
+    int num_cpu_threads = 1;
-    for (const auto& r : recs) {
+    if (env_threads) {
-        DayIndex day = static_cast<DayIndex>(r.timestamp) / 86400;
+        num_cpu_threads = std::atoi(env_threads);
-        days[day].push_back(r);
+        if (num_cpu_threads < 1) num_cpu_threads = 1;
+    }
+    return num_cpu_threads;
 }
 
-    return days;
+std::string get_env(const char* name) {
+    const char* env = std::getenv(name);
+    if (!env) {
+        throw std::runtime_error(std::string("Environment variable not set: ") + name);
+    }
+    return std::string(env);
 }
 
-std::vector<std::vector<DayIndex>> split_days(const std::map<DayIndex, std::vector<Record>>& days, int parts) {
+std::string get_data_path() {
-    std::vector<std::vector<DayIndex>> out(parts);
+    return get_env("DATA_PATH");
-
-    int i = 0;
-    for (auto& kv : days) {
-        out[i % parts].push_back(kv.first);
-        i++;
 }
 
-    return out;
+std::vector<int> get_data_read_shares() {
+    std::vector<int> shares;
+    std::stringstream ss(get_env("DATA_READ_SHARES"));
+    std::string item;
+    while (std::getline(ss, item, ',')) {
+        shares.push_back(std::stoi(item));
+    }
+    return shares;
 }
 
+int64_t get_read_overlap_bytes() {
+    return std::stoll(get_env("READ_OVERLAP_BYTES"));
+}
+
+int64_t get_aggregation_interval() {
+    return std::stoll(get_env("AGGREGATION_INTERVAL"));
+}
+
+bool get_use_cuda() {
+    return std::stoi(get_env("USE_CUDA")) != 0;
+}
+
+int64_t get_file_size(const std::string& path) {
+    std::ifstream file(path, std::ios::binary | std::ios::ate);
+    if (!file.is_open()) {
+        throw std::runtime_error("Cannot open file: " + path);
+    }
+    return static_cast<int64_t>(file.tellg());
+}
+
+ByteRange calculate_byte_range(int rank, int size, int64_t file_size,
+                               const std::vector<int>& shares, int64_t overlap_bytes) {
+    std::vector<int> effective_shares;
+    if (shares.size() == static_cast<size_t>(size)) {
+        effective_shares = shares;
+    } else {
+        effective_shares.assign(size, 1);
+    }
+    
+    int total_shares = std::accumulate(effective_shares.begin(), effective_shares.end(), 0);
+    int64_t bytes_per_share = file_size / total_shares;
+    
+    int64_t base_start = 0;
+    for (int i = 0; i < rank; i++) {
+        base_start += bytes_per_share * effective_shares[i];
+    }
+    
+    int64_t base_end = base_start + bytes_per_share * effective_shares[rank];
+    
+    ByteRange range;
+    
+    if (rank == 0) {
+        range.start = 0;
+        range.end = std::min(base_end + overlap_bytes, file_size);
+    } else if (rank == size - 1) {
+        range.start = std::max(base_start - overlap_bytes, static_cast<int64_t>(0));
+        range.end = file_size;
+    } else {
+        range.start = std::max(base_start - overlap_bytes, static_cast<int64_t>(0));
+        range.end = std::min(base_end + overlap_bytes, file_size);
+    }
+    
+    return range;
+}
+
+void trim_edge_periods(std::vector<PeriodStats>& periods, int rank, int size) {
+    if (periods.empty()) return;
+    
+    if (rank == 0) {
+        periods.pop_back();
+    } else if (rank == size - 1) {
+        periods.erase(periods.begin());
+    } else {
+        periods.pop_back();
+        periods.erase(periods.begin());
+    }
+}

src/utils.hpp

@@ -1,11 +1,33 @@
 #pragma once
 
 #include "record.hpp"
+#include "period_stats.hpp"
 #include <map>
 #include <vector>
+#include <string>
+#include <cstdlib>
+#include <cstdint>
 
-using DayIndex = long long;
+// Чтение переменных окружения
+int get_num_cpu_threads();
+std::string get_data_path();
+std::vector<int> get_data_read_shares();
+int64_t get_read_overlap_bytes();
+int64_t get_aggregation_interval();
+bool get_use_cuda();
 
-std::map<DayIndex, std::vector<Record>> group_by_day(const std::vector<Record>& recs);
+// Структура для хранения диапазона байт для чтения
-std::vector<std::vector<DayIndex>> split_days(const std::map<DayIndex, std::vector<Record>>& days, int parts);
+struct ByteRange {
+    int64_t start;
+    int64_t end;  // exclusive
+};
 
+// Вычисляет диапазон байт для конкретного ранка
+ByteRange calculate_byte_range(int rank, int size, int64_t file_size,
+                               const std::vector<int>& shares, int64_t overlap_bytes);
+
+// Получение размера файла
+int64_t get_file_size(const std::string& path);
+
+// Удаляет крайние периоды, которые могут быть неполными из-за параллельного чтения
+void trim_edge_periods(std::vector<PeriodStats>& periods, int rank, int size);

upsample.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+import argparse
+import sys
+import time
+
+
+def parse_row(line: str):
+    # Timestamp,Open,High,Low,Close,Volume
+    ts, o, h, l, c, v = line.split(',')
+    return int(float(ts)), float(o), float(h), float(l), float(c), float(v)
+
+
+def fmt_row(ts, o, h, l, c, v):
+    return f"{ts},{o:.2f},{h:.2f},{l:.2f},{c:.2f},{v:.8f}\n"
+
+
+def count_lines_fast(path: str) -> int:
+    with open(path, "rb") as f:
+        return sum(1 for _ in f) - 1  # минус header
+
+
+def main(inp, out, step, flush_every):
+    # считаем количество строк для прогресса
+    total_lines = count_lines_fast(inp)
+    print(f"Total input rows: {total_lines:,}", file=sys.stderr)
+
+    start_time = time.time()
+    processed = 0
+    last_report = start_time
+
+    with open(inp, "r", buffering=8 * 1024 * 1024) as fin, \
+         open(out, "w", buffering=8 * 1024 * 1024) as fout:
+
+        fin.readline()  # пропускаем header
+        fout.write("Timestamp,Open,High,Low,Close,Volume\n")
+
+        first = fin.readline()
+        if not first:
+            return
+
+        prev = parse_row(first.strip())
+
+        out_buf = []
+        out_rows = 0
+
+        for line in fin:
+            line = line.strip()
+            if not line:
+                continue
+
+            cur = parse_row(line)
+
+            t1, o1, h1, l1, c1, v1 = prev
+            t2, o2, h2, l2, c2, v2 = cur
+
+            dt = t2 - t1
+            steps = dt // step
+
+            if steps > 0:
+                do = o2 - o1
+                dh = h2 - h1
+                dl = l2 - l1
+                dc = c2 - c1
+                dv = v2 - v1
+
+                inv = 1.0 / steps
+                for i in range(steps):
+                    a = i * inv
+                    out_buf.append(fmt_row(
+                        t1 + i * step,
+                        o1 + do * a,
+                        h1 + dh * a,
+                        l1 + dl * a,
+                        c1 + dc * a,
+                        v1 + dv * a
+                    ))
+
+                out_rows += steps
+
+            prev = cur
+            processed += 1
+
+            # прогресс
+            if processed % 100_000 == 0:
+                now = time.time()
+                if now - last_report >= 0.5:
+                    pct = processed * 100.0 / total_lines
+                    elapsed = now - start_time
+                    speed = processed / elapsed if elapsed > 0 else 0
+                    eta = (total_lines - processed) / speed if speed > 0 else 0
+
+                    print(
+                        f"\rprocessed: {processed:,} / {total_lines:,} "
+                        f"({pct:5.1f}%) | "
+                        f"out ~ {out_rows:,} | "
+                        f"{speed:,.0f} rows/s | "
+                        f"ETA {eta/60:5.1f} min",
+                        end="",
+                        file=sys.stderr,
+                        flush=True,
+                    )
+                    last_report = now
+
+            # сброс буфера
+            if out_rows >= flush_every:
+                fout.write("".join(out_buf))
+                out_buf.clear()
+                out_rows = 0
+
+        # остатки
+        if out_buf:
+            fout.write("".join(out_buf))
+
+        # последнюю строку пишем как есть
+        t, o, h, l, c, v = prev
+        fout.write(fmt_row(t, o, h, l, c, v))
+
+    total_time = time.time() - start_time
+    print(
+        f"\nDone in {total_time/60:.1f} min",
+        file=sys.stderr
+    )
+
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser()
+    ap.add_argument("-i", "--input", required=True)
+    ap.add_argument("-o", "--output", required=True)
+    ap.add_argument("-s", "--step", type=int, default=10)
+    ap.add_argument("--flush-every", type=int, default=200_000)
+    args = ap.parse_args()
+
+    if args.step <= 0:
+        raise SystemExit("step must be > 0")
+
+    main(args.input, args.output, args.step, args.flush_every)