CompletableFuture异步多任务最佳实践

简介

CompletableFuture相比于Java8的并行流，对于处理并发的IO密集型任务有着得天独厚的优势：

在流式编程下，支持构建任务流时即可执行任务。
CompletableFuture任务支持提交到自定义线程池，调优方便。

本文所有案例都会基于这样的一个需求，某网站有多个商家,用户会在不同的店铺查看同一件商品，只要用户在提供给商店对应的产品名称，商店就会返回对应产品的最终价格。
该需求有几个注意点:

每次到一家商店查询时，同一时间只能查询一个商品。
允许用户同一时间在系统里查询多个商店的一个商品。
查询的商品售价是需要耗时的，平均500ms-2500ms不等。
为了更直观了解代码性能，我们的例子用户每次会去家左右的店铺分别查询一件商品。

在这里插入图片描述

这里介绍一下，这些功能所要用到的类，首先是商家类，核心方法getPrice会根据用户传入的产品名称返回售价，代码执行时会休眠一段时间返回用户产品最终价格。

/**
 * 商店
 */
public class Shop {

    private final String name;
    private final Random random;

    public Shop(String name) {
        this.name = name;
        random = new Random(name.charAt(0) * name.charAt(1) * name.charAt(2));
    }

    /**
     * 传入产品名称返回对应价格和折扣代码
     * @param product 产品名称
     * @return
     */
    public String getPrice(String product) {
        double price = calculatePrice(product);
        Discount.Code code = Discount.Code.values()[random.nextInt(Discount.Code.values().length)];
        return name + ":" + price + ":" + code;
    }


    

    /**
     * 获取商品价格
     * @param product
     * @return
     */
    public double calculatePrice(String product) {
        try {
            TimeUnit.MILLISECONDS.sleep(RandomUtil.randomInt(500,2500));
        } catch (InterruptedException e) {
            e.printStackTrace();
        }
        return (int)(random.nextDouble() * product.charAt(0) + product.charAt(1));
    }

    public String getName() {
        return name;
    }
}

同时我们也给出了18个商家的列表。

private final static List<Shop> shops = Arrays.asList(new Shop("Nike"),
            new Shop("Apple"),
            new Shop("Coca-Cola"),
            new Shop("Amazon"),
            new Shop("Samsung"),
            new Shop("McDonald's"),
            new Shop("Mercedes-Benz"),
            new Shop("Google"),
            new Shop("Louis Vuitton"),
            new Shop("Chanel"),
            new Shop("Gucci"),
            new Shop("Adidas"),
            new Shop("Pepsi"),
            new Shop("Ford"),
            new Shop("Microsoft"),
            new Shop("Rolex"),
            new Shop("Ferrari"),
            new Shop("IKEA")
    );

实现

顺序流

第一版我们先用顺序流实现改需求，代码很简单用stream遍历商家并调用getPrice获得结果，最终存到List中。

public static void main(String[] args) {
        long begin = System.currentTimeMillis();
        List<String> resultList = shops.stream()
                .map(shop -> shop.getName() + " price is " + shop.getPrice("Air Jordan"))
                .collect(Collectors.toList());
        long end = System.currentTimeMillis();
        System.out.println("执行结束,执行结果:" + JSONUtil.toJsonStr(resultList) + " 执行耗时:" + (end - begin)+"ms");
    }

最终输出结果如下，可以看到用了快30s。

执行结束,执行结果:["Nike price is Nike:112.0:GOLD","Apple price is Apple:135.0:DIAMOND","Coca-Cola price is Coca-Cola:111.0:DIAMOND","Amazon price is Amazon:155.0:GOLD","Samsung price is Samsung:165.0:SILVER","McDonald's price is McDonald's:164.0:PLATINUM","Mercedes-Benz price is Mercedes-Benz:111.0:DIAMOND","Google price is Google:121.0:NONE","Louis Vuitton price is Louis Vuitton:168.0:PLATINUM","Chanel price is Chanel:131.0:NONE","Gucci price is Gucci:132.0:SILVER","Adidas price is Adidas:133.0:PLATINUM","Pepsi price is Pepsi:158.0:SILVER","Ford price is Ford:113.0:PLATINUM","Microsoft price is Microsoft:156.0:GOLD","Rolex price is Rolex:126.0:GOLD","Ferrari price is Ferrari:163.0:DIAMOND","IKEA price is IKEA:150.0:NONE"] 执行耗时:27409ms

并行流

对此我们采用并行流尝试以多线程的形式执行任务，所以我们将stream改为parallelStream。

public static void main(String[] args) {
        long begin = System.currentTimeMillis();
        List<String> resultList = shops.parallelStream()
                .map(shop -> shop.getName() + " price is " + shop.getPrice("Air Jordan"))
                .collect(Collectors.toList());
        long end = System.currentTimeMillis();
        System.out.println("执行结束,执行结果:" + JSONUtil.toJsonStr(resultList) + " 执行耗时:" + ((end - begin)/1000)+"s");
    }

可以看到耗时仅仅5s,这里补充一下笔者的机器信息，笔者的CPU是6核的，所以并行流执行时会有6个线程在同时工作。

执行结束,执行结果:["Nike price is Nike:112.0:GOLD","Apple price is Apple:135.0:DIAMOND","Coca-Cola price is Coca-Cola:111.0:DIAMOND","Amazon price is Amazon:155.0:GOLD","Samsung price is Samsung:165.0:SILVER","McDonald's price is McDonald's:164.0:PLATINUM","Mercedes-Benz price is Mercedes-Benz:111.0:DIAMOND","Google price is Google:121.0:NONE","Louis Vuitton price is Louis Vuitton:168.0:PLATINUM","Chanel price is Chanel:131.0:NONE","Gucci price is Gucci:132.0:SILVER","Adidas price is Adidas:133.0:PLATINUM","Pepsi price is Pepsi:158.0:SILVER","Ford price is Ford:113.0:PLATINUM","Microsoft price is Microsoft:156.0:GOLD","Rolex price is Rolex:126.0:GOLD","Ferrari price is Ferrari:163.0:DIAMOND","IKEA price is IKEA:150.0:NONE"] 执行耗时:5077ms

使用CompletableFuture执行异步多查询任务

我们给出了CompletableFuture执行多IO查询任务的代码示例，可以看到代码的执行流程大致为:

遍历商家。
提交异步查询任务。
调用join()，注意这里的join和CompletableFuture的get方法作用是一样的,都是阻塞获取查询结果，唯一的区别就是join方法签名没有抛异常，所以无需try-catch处理。

 public static void main(String[] args) {
        long begin = System.currentTimeMillis();
        List<String> resultList = shops.stream()
                        .map(shop -> CompletableFuture.supplyAsync(() -> shop.getName() + " price is "
                                + shop.getPrice("Air Jordan")))//CompletableFuture提交价格查询任务
                        .map(CompletableFuture::join) //用join阻塞获取结果
                        .collect(Collectors.toList());//组成列表
        long end = System.currentTimeMillis();
        System.out.println("执行结束,执行结果:" + JSONUtil.toJsonStr(resultList) + " 执行耗时:" + (end - begin) + "ms");
    }

可以看到执行结果为31s，查询效率还不如顺序流。

执行结束,执行结果:["Nike price is Nike:112.0:GOLD","Apple price is Apple:135.0:DIAMOND","Coca-Cola price is Coca-Cola:111.0:DIAMOND","Amazon price is Amazon:155.0:GOLD","Samsung price is Samsung:165.0:SILVER","McDonald's price is McDonald's:164.0:PLATINUM","Mercedes-Benz price is Mercedes-Benz:111.0:DIAMOND","Google price is Google:121.0:NONE","Louis Vuitton price is Louis Vuitton:168.0:PLATINUM","Chanel price is Chanel:131.0:NONE","Gucci price is Gucci:132.0:SILVER","Adidas price is Adidas:133.0:PLATINUM","Pepsi price is Pepsi:158.0:SILVER","Ford price is Ford:113.0:PLATINUM","Microsoft price is Microsoft:156.0:GOLD","Rolex price is Rolex:126.0:GOLD","Ferrari price is Ferrari:163.0:DIAMOND","IKEA price is IKEA:150.0:NONE"] 执行耗时:31097ms

原因很简单，我们本次的流操作执行会构成下面这样一张图，可以看到查询价格的操作是有耗时的，随后我们又调用了join方法使得流的后续步骤被阻塞，最终CompletableFuture用的和顺序流一样。

在这里插入图片描述

分解流优化使用CompletableFuture

上述我们提到过，之所以慢是因为join阻塞了流的操作，所以提升效率的方式就是不要让join阻塞流的操作。所以笔者将流拆成了两个。
如下图，第一个流负责提交任务，即遍历每一个商家并将查询价格的任务提交出去，期间不阻塞，最终会生成一个CompletableFuture的List。

紧接着我们遍历上一个流生成的List<CompletableFuture>，调用join方法阻塞获取结果，因为上一个流操作提交任务时不阻塞，所以每个任务一提交时就可能已经在执行了，所以join方法获取结果的耗时也会相对短一些。

在这里插入图片描述

所以我们的代码最后改造成了这样:

public static void main(String[] args) {
        long begin = System.currentTimeMillis();
        List<CompletableFuture<String>> completableFutureList = shops.stream()
                .map(shop -> CompletableFuture.supplyAsync(() -> shop.getName() + " price is " + shop.getPrice("Air Jordan")))
                .collect(Collectors.toList());//CompletableFuture提交价格查询任务

        List<String> resultList = completableFutureList.stream()
                .map(CompletableFuture::join) //用join阻塞获取结果
                .collect(Collectors.toList());//组成列表

        long end = System.currentTimeMillis();
        System.out.println("执行结束,执行结果:" + JSONUtil.toJsonStr(resultList) + " 执行耗时:" + (end - begin) + "ms");
    }

执行结果如下，可以看到代码耗时差不多也是5s和并行流差不多，原因很简单，线程池默认用6个，对于IO密集型任务来说显然是不够的。

执行结束,执行结果:["Nike price is Nike:112.0:GOLD","Apple price is Apple:135.0:DIAMOND","Coca-Cola price is Coca-Cola:111.0:DIAMOND","Amazon price is Amazon:155.0:GOLD","Samsung price is Samsung:165.0:SILVER","McDonald's price is McDonald's:164.0:PLATINUM","Mercedes-Benz price is Mercedes-Benz:111.0:DIAMOND","Google price is Google:121.0:NONE","Louis Vuitton price is Louis Vuitton:168.0:PLATINUM","Chanel price is Chanel:131.0:NONE","Gucci price is Gucci:132.0:SILVER","Adidas price is Adidas:133.0:PLATINUM","Pepsi price is Pepsi:158.0:SILVER","Ford price is Ford:113.0:PLATINUM","Microsoft price is Microsoft:156.0:GOLD","Rolex price is Rolex:126.0:GOLD","Ferrari price is Ferrari:163.0:DIAMOND","IKEA price is IKEA:150.0:NONE"] 执行耗时:5633ms

CompletableFuture使用自定义线程池

《Java并发编程实战》一书中，Brian Goetz和合著者们为线程池大小的优化提供了不少中肯的建议。这非常重要，如果线程池中线程的数量过多，最终它们会竞争稀缺的处理器和内存资源，浪费大量的时间在上下文切换上。反之，如果线程的数目过少，正如你的应用所面临的情况，处理器的一些核可能就无法充分利用。BrianGoetz建议，线程池大小与处理器的利用率之比可以使用下面的公式进行估算：
N(threads) = N(CPU)* U(CPU) * (1+ W/C)
其中： N(CPU)是处理器的核的数目，可以通过 Runtime.getRuntime().available Processors() 得到。U(CPU)是期望的 CPU利用率(该值应该介于 0和 1之间) W/C是等待时间与计算时间的比率。

我们的CPU核心数为6,我们希望的CPU利用率为1,而等待时间按照这种的计算应该是1250ms而计算时间可以忽略不计，所以W/C差不多可以换算为1000。
最终我们计算结果为:

N(threads) = N(CPU)* U(CPU) * (1+ W/C)
		   = 6*1*1000
		   =6000

很明显6000个线程非常不合理，所以我们使用了和商店数差不多的线程数，所以我们将线程设置为18(这也是个大概的数字，具体情况还需要经过压测进行增减)。

最终代码写成这样。

 private static ThreadPoolExecutor threadPool = new ThreadPoolExecutor(20,
            20,
            1,
            TimeUnit.MINUTES,
            new ArrayBlockingQueue<>(100),
            new ThreadFactoryBuilder().setNamePrefix("threadPool-%d").build());


    public static void main(String[] args) {
        long begin = System.currentTimeMillis();
        List<CompletableFuture<String>> completableFutureList = shops.stream()
                .map(shop -> CompletableFuture.supplyAsync(() -> shop.getName() + " price is " + shop.getPrice("Air Jordan"), threadPool))
                .collect(Collectors.toList());//CompletableFuture提交价格查询任务

        List<String> resultList = completableFutureList.stream()
                .map(CompletableFuture::join) //用join阻塞获取结果
                .collect(Collectors.toList());//组成列表

        long end = System.currentTimeMillis();
        System.out.println("执行结束,执行结果:" + JSONUtil.toJsonStr(resultList) + " 执行耗时:" + (end - begin) + "ms");
        threadPool.shutdownNow();
    }

输出结果如下，可以看到查询效率有了质的飞跃。

执行结束,执行结果:["Nike price is Nike:112.0:GOLD","Apple price is Apple:135.0:DIAMOND","Coca-Cola price is Coca-Cola:111.0:DIAMOND","Amazon price is Amazon:155.0:GOLD","Samsung price is Samsung:165.0:SILVER","McDonald's price is McDonald's:164.0:PLATINUM","Mercedes-Benz price is Mercedes-Benz:111.0:DIAMOND","Google price is Google:121.0:NONE","Louis Vuitton price is Louis Vuitton:168.0:PLATINUM","Chanel price is Chanel:131.0:NONE","Gucci price is Gucci:132.0:SILVER","Adidas price is Adidas:133.0:PLATINUM","Pepsi price is Pepsi:158.0:SILVER","Ford price is Ford:113.0:PLATINUM","Microsoft price is Microsoft:156.0:GOLD","Rolex price is Rolex:126.0:GOLD","Ferrari price is Ferrari:163.0:DIAMOND","IKEA price is IKEA:150.0:NONE"] 执行耗时:1893ms

并行流一定比CompletableFuture烂吗？

如果是计算密集型的任务，使用stream是最佳姿势，因为密集型需要一直计算，加多少个线程都无济于事，使用stream简单使用了。
而对于io密集型的任务，例如上文这种大量查询都需要干等的任务，使用CompletableFuture是最佳姿势了，通过自定义线程创建比cpu核心数更多的线程来提高工作效率才是较好的解决方案