如何在Java中并行执行for-each循环中的语句？

Question

我有一段代码如下所示：

public List<Restaurant> getAllRestaurants() {
    List<Restaurant> restaurants = getRestaurants().subList(0, 7); // This takes 234 ms to execute on average.    

    // There are 7 items in the restaurants list
    for (Restaurant restaurant : restaurants) {
        PlacesAPIResponse response = callGooglePlacesAPI(restaurant); // A call to the Google API should take 520ms for a given restaurant
        restaurant.setRating(response.getRating());
    }
    return restaurants;
}

如果我如图所示在 for-each 循环中执行上述语句，我预计该方法的总时间为 234ms + (7*520)ms = 3874ms，因为语句是按顺序运行的。这太慢了，所以我想并行化 for-each 循环中的语句，以便我为列表中的每个餐厅同时调用 Google Places API。理论上，响应时间应该是234ms + max(API call for Restaurant 1, ..., API call for Restaurant 7) = 234ms + 520ms = 754ms，因为对 Google API 的调用是并行进行的。

根据this link (Java 8: Parallel FOR loop)，我应该可以像这样使用parallelStream()同时执行语句：

long startTime = System.currentTimeMillis();
restaurants.parallelStream().forEach(restaurant -> {
    PlacesAPIResponse response = callGooglePlacesAPI(restaurant);
    restaurant.setRating(response.getRating());
});
long endTime = System.currentTimeMillis();
System.out.println("Calling Google Places API took " + (endTime - startTime) + " milliseconds");

这似乎为每家餐厅并行调用 Google Places API，但现在每次调用 Google Places API 似乎都需要越来越多的时间。这是我的时间戳的输出：

getRestaurants() took 234 milliseconds
Took 335 milliseconds to call Google Places API for Restaurant 1
Took 337 milliseconds to call Google Places API for Restaurant 2
Took 671 milliseconds to call Google Places API for Restaurant 3
Took 742 milliseconds to call Google Places API for Restaurant 4
Took 1086 milliseconds to call Google Places API for Restaurant 5
Took 1116 milliseconds to call Google Places API for Restaurant 6
Took 1470 milliseconds to call Google Places API for Restaurant 7
Calling Google Places API took 1473 milliseconds

1734ms 比我预期的 754ms 大得多。我已经尝试过并行流以及 ExecutorService 来同时调用 Google Places API，但我似乎无法获得所需的响应时间。谁能指出我正确的方向？谢谢。

编辑：这是我对 ExecutorService 的尝试，根据这篇文章(Is there a easy way to parallelize a foreach loop in java?)：

startTime = System.currentTimeMillis();
ExecutorService exe = Executors.newFixedThreadPool(2);   // 2 can be changed of course
for (Restaurant restaurant : restaurants) {
    exe.submit(() -> {
        PlacesAPIResponse response = callGooglePlacesAPI(restaurant); // A call to the Google API should take 520ms for a given restaurant
        restaurant.setRating(response.getRating());
    });
}    

exe.shutdown();
try {
    exe.awaitTermination(Long.MAX_VALUE, TimeUnit.NANOSECONDS);
} catch (InterruptedException e) {
    e.printStackTrace();
}    

endTime = System.currentTimeMillis();
System.out.println("Calling Google Places API took " + (endTime - startTime) + " milliseconds");
return restaurants; 

这是我的时间戳的输出：

getRestaurants() took 234 milliseconds
Took 464 milliseconds to call Google Places API for Restaurant 1
Took 575 milliseconds to call Google Places API for Restaurant 2
Took 452 milliseconds to call Google Places API for Restaurant 3
Took 420 milliseconds to call Google Places API for Restaurant 4
Took 414 milliseconds to call Google Places API for Restaurant 5
Took 444 milliseconds to call Google Places API for Restaurant 6
Took 422 milliseconds to call Google Places API for Restaurant 7
Calling Google Places API took 1757 milliseconds

这个方法的响应时间仍然是234ms + 1757 ms而不是234ms + 575ms，我不明白为什么。

Answer 1

这里最好使用 executorService 并为它们提供任务作为单独的 Runnable()。

或者你可以在这里使用 Future。

Answer 2

这是很久以前的事了，但我想原因在于您选择的线程池大小。线程池大小为 2 意味着您只能并行执行两个作业。剩余的作业排队，直到线程被释放。因此，您执行 Google Places API 的计算将类似于max(464+452+414+422, 575+420+444) = max(1752, 1439) = 1752，接近实际值。这很好解释here。

Answer 3

我猜你的瓶颈是与互联网或 Google Places 服务器的连接，而不是你的循环。服务器识别相同的 IP 地址，因此将您的请求排队以保护自己免受拒绝服务攻击。 这意味着您的循环并行运行，但互联网请求堆叠在服务器上，这就是为什么每个请求越来越需要更多时间才能得到响应和返回。 为避免这种情况，您需要类似 bot net（从不同计算机发送每个查询）之类的东西，或者 Google Places 可能会向您出售用于并行请求的特殊连接。