linux - PySpark 速度 Ubuntu 与 Windows

Question

我有一个 PySpark 示例作业，它是 PageRank 算法的一个版本。代码如下:

from __future__ import print_function
from operator import add
import timeit
from pyspark.sql import SparkSession

# Normalize a list of pairs(url, rank) to 1
def normalize(ranks):
    norm = sum([rank for u, rank in ranks])
    ranks = [(u, rank / norm) for (u, rank) in ranks ]
    return sorted(ranks, key=lambda x: x[1], reverse=True)

def pagerank_2(edgeList, n, niter):
    # Loads all URLs from input file and initialize their neighbors.
    m = edgeList.groupByKey().cache()
    s = 0.85

    # Loads all URLs with other URL(s) link to from input file 
    # and initialize ranks of them to one.
    q = spark.sparkContext.range(n).map(lambda x: (x, 1.0)).cache()
    r = spark.sparkContext.range(n).map(lambda x: (x, 0.0)).cache()

    # Calculates and updates URL ranks continuously 
    # using PageRank algorithm.
    for iteration in range(niter):
        # Calculates URL contributions to the rank of other URLs.
        # Add URL ranks based on neighbor contributions.
        # Do not forget to add missing values in q and set to 0.0
        q = q.fullOuterJoin(m)\
             .flatMap(lambda x: (x[1][1] and [(u, x[1][0]/len(x[1][1])) for u in x[1][1]]) or [])\
             .reduceByKey(add)\
             .rightOuterJoin(r)\
             .mapValues(lambda x: (x[0] or 0)*s + (1-s))
        print("iteration = ", iteration)

    # Collects all URL ranks and dump them to console after normalization
    ranks = normalize(q.collect())
    print(ranks[0:10])


if __name__ == "__main__":

    spark = SparkSession\
            .builder\
            .master('local[*]')\
            .appName("SparkPageRank")\
            .config('spark.driver.allowMultipleContexts', 'true')\
            .config('spark.sql.warehouse.dir', 'file:///C:/Home/Org/BigData/python/BE4/') \
            .config('spark.sql.shuffle.partitions', '10')\
            .getOrCreate()

    spark.sparkContext.setLogLevel('WARN')

    g = [(0, 1), (0, 5), (1, 2), (1, 3), (2, 3),
         (2, 4), (2, 5), (3, 0), (5, 0), (5, 2)]
    n = 6
    edgeList = spark.sparkContext.parallelize(g)
    print(timeit.timeit('pagerank_2(edgeList, 6, 10)', number=1, globals=globals()))

节点从 0 到 n-1 编号。 edgeList 参数是一个 RDD，其中包含节点对(也称为边)列表。

我在 Windows 10(Anaconda、Spark 2.1.0、winutils)上以本地模式运行它。这项工作被分配为 2896 个任务，这些任务都很轻。

我的问题是运行时间。以上面的例子为例:

• Windows 10:>4000 万!
• 适用于 Linux 的 Windows 子系统 (Ubuntu 14.04):30 秒

该计算机是笔记本电脑核心 i7-4702HQ、16Gb 内存、512Gb SSD。在启动过程中，Windows 比 Linux 慢，但慢了 50 倍？肯定有办法缩小这种差距？

我已为所有相关文件禁用 Windows Defender:java 目录、python 目录等。关于要看什么还有其他想法吗？

感谢任何线索。

Answer 1

也许关键是 local[*] 这意味着

Run Spark locally with as many worker threads as logical cores on your machine.

尝试使用诸如local[10]