类似的并发任务，运行时间却大相径庭(Java)

Question

我有一个任务要在oracle数据库表中插入800万行。我把这个任务扩展到CPU的8个核心上，这里是示例代码。

final long start = System.currentTimeMillis();
final int batchSize = 800000;
final int nCore = Runtime.getRuntime().availableProcessors();
final int batchPerCore = batchSize / nCore;
final CountDownLatch taskCountDown = new CountDownLatch(nCore);
final CountDownLatch kickoffLatch = new CountDownLatch(1);
final ExecutorService es = Executors.newFixedThreadPool(nCore);

Class.forName("oracle.jdbc.driver.OracleDriver");

for (int n = 0; n < nCore; n++) {

    es.submit(() -> {

        log.info("Thread {} starts working on {} insertion jobs.", Thread.currentThread().getName(), batchPerCore);
        try (Connection conn = DriverManager.getConnection(DB_URL, USER_NAME, PASSWORD)) {

            Assert.assertNotNull("connection established", conn);
            conn.setAutoCommit(false);
            PreparedStatement pstmt = conn.prepareStatement(PREP_STATEMENT);


            for (int i = 0; i < batchPerCore; i++) {
                initPrepTranStatement(pstmt);
                pstmt.addBatch();

            }

            int[] resultTran = pstmt.executeBatch();
            Assert.assertEquals(resultTran.length, batchPerCore);
            conn.commit();
            log.info("Thread {} completed its job!", Thread.currentThread().getName());
            taskCountDown.countDown();

        } catch (SQLException e) {
            e.printStackTrace();
        }

    });

} // ~ for ~

es.shutdown();

kickoffLatch.countDown();

taskCountDown.await();

log.info("It takes {} milliseconds to complete all {} tasks for {} insertions each. ",
        (System.currentTimeMillis()-start), nCore, batchPerCore);

当它进行时，它显示3个任务完成了 在下午2: 09运行课程， 但其他的仍然是待定的 在写这篇文章时（3: 29PM）。

这里是日志。

Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-7 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-2 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-5 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-8 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-4 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-3 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-6 starts working on 100000 insertion jobs.
Apr 27, 2020 1:25:17 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-1 starts working on 100000 insertion jobs.
Apr 27, 2020 2:09:53 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-4 completed its job!
Apr 27, 2020 2:09:53 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-6 completed its job!
Apr 27, 2020 2:09:53 PM com.db.loader.DBFlexDataLoadingPerfTest lambda$4
INFO: Thread pool-1-thread-7 completed its job!

在这种情况下使用DriverManager并不是一个好主意，我应该使用DataSource实现，尽管如此，它并不相关，也不影响本次测试的目的。

鉴于所有的任务在工作性质和运行方式上都是一样的，我无法理解为什么有些任务严重落后于其他任务，更有可能只是挂在那里。

没有设置特别的JVM参数。机器是Windows 10 pro，8核i7-6700，32G内存。

Answer 1

线程是不确定的。它不能保证所有的任务都会被所有线程均匀地执行。在最悲观的情况下，这些任务甚至可以被序列化（不太可能，但不是不可能）。这一切都取决于JVM的实现和操作系统的调度。

除此之外，还有很多实现上的依赖性因素，这些因素在提供的代码片段中是不可见的（例如，驱动或数据库处理并行批次提交）。