<dependencies>

        <dependency>

            <groupId>junit</groupId>

            <artifactId>junit</artifactId>

            <version>3.8.1</version>

            <scope>test</scope>

        </dependency>

        <!-- https://mvnrepository.com/artifact/org.apache.spark/spark-core -->

        <dependency>

            <groupId>org.apache.spark</groupId>

            <artifactId>spark-core_2.12</artifactId>

            <version>3.0.0-preview</version>

        </dependency>

        <!-- https://mvnrepository.com/artifact/org.apache.spark/spark-launcher -->

        <dependency>

            <groupId>org.apache.spark</groupId>

            <artifactId>spark-launcher_2.12</artifactId>

            <version>3.0.0-preview</version>

        </dependency>

        <dependency>

            <groupId>org.apache.spark</groupId>

            <artifactId>spark-sql_2.12</artifactId>

            <version>3.0.0-preview</version>

        </dependency>

    </dependencies>

注意：目前阿里云镜像部分包还没有（2019年11月10日，spark-launcher_2.12下载没有），可以用国外的。

测试代码：

object SparkPi {

  def main(args: Array[String]): Unit = {

    val spark = SparkSession

      .builder

      .appName("Spark Pi")

      .master("local[2]")

      .config("spark.driver.resource.gpu.discoveryScript", "D:\\gpu.bat")

      .config("spark.worker.resource.gpu.discoveryScript", "D:\\gpu.bat")

      .config("spark.driver.resource.gpu.amount", 1)

      .config("spark.executor.resource.gpu.amount", 1)

      .config("spark.worker.resource.gpu.amount", 1)

      .getOrCreate()

    val slices = if (args.length > 0) args(0).toInt else 2

    val n = math.min(100000L * slices, Int.MaxValue).toInt // avoid overflow

    val count = spark.sparkContext.parallelize(1 until n, slices).map { i =>

      val x = random * 2 - 1

      val y = random * 2 - 1

      if (x * x + y * y <= 1) 1 else 0

    }.reduce(_ + _)

    println(s"Pi is roughly ${4.0 * count / (n - 1)}")

    spark.stop()

  }

}

其中，gpu.bat的内容如下：

@echo off

echo {"name": "gpu", "addresses": ["0"]}

运行日志如下：

我以为可以成功调用GPU，查看任务管理器里面的GPU显示，并没有发现，最后搜索代码，在"spark-3.0.0-preview\core\src\main\scala\org\apache\spark\scheduler\local\LocalSchedulerBackend.scala"(85,56)显示如下：

def reviveOffers(): Unit = {

    // local mode doesn't support extra resources like GPUs right now

    val offers = IndexedSeq(new WorkerOffer(localExecutorId, localExecutorHostname, freeCores,

      Some(rpcEnv.address.hostPort)))

    for (task <- scheduler.resourceOffers(offers).flatten) {

      freeCores -= scheduler.CPUS_PER_TASK

      executor.launchTask(executorBackend, task)

    }

  }

注释：local mode doesn't support extra resources like GPUs right now

心一凉，本来打算搭建standalone模式，最后看了一下window的搞不了，Linux的得个虚拟机了，比较笨资源有限，就暂不试了。