apache ignite系列（五）：分布式计算

标签：any   故障   作业   try   实体   result   static   序列   cut   

ignite分布式计算

在ignite中，有传统的MapReduce模型的分布式计算，也有基于分布式存储的并置计算，当数据分散到不同的节点上时，根据提供的并置键，计算会传播到数据所在的节点进行计算，再结合数据并置，相关联的数据存储在相同节点，这样可以避免在计算过程中涉及到大量的数据移动，有效保证计算的性能。

ignite分布式计算的主要特点如下：

1.分布式闭包:

Ignite计算网格可以对集群或者集群组内的任何闭包进行广播和负载平衡，包括纯Java的runnables和callables

ComputeTestController.java

    /** broadCast测试*/
    @RequestMapping("/broadcast")
    String broadcastTest(HttpServletRequest request, HttpServletResponse response) {
//        IgniteCompute compute = ignite.compute(ignite.cluster().forRemotes());  //只传播远程节点
        IgniteCompute compute = ignite.compute();
        compute.broadcast(() -> System.out.println("Hello Node: " + ignite.cluster().localNode().id()));
        return "all executed.";
    }

    /** call和run测试 */
    @RequestMapping("/call")
    public @ResponseBody
    String callTest(HttpServletRequest request, HttpServletResponse response) {
        Collection<IgniteCallable<Integer>> calls = new ArrayList<>();

        /** call */
        System.out.println("-----------call-----------");
        for(String word : "How many characters".split(" ")) {
            calls.add(word::length);
//            calls.add(() -> word.length());
        }
        Collection<Integer> res = ignite.compute().call(calls);
        int total = res.stream().mapToInt(Integer::intValue).sum();
        System.out.println(String.format("the total lengths of all words is [%s].", total));

        /** run */
        System.out.println("-----------run-----------");
        for (String word : "Print words on different cluster nodes".split(" ")) {
            ignite.compute().run(() -> System.out.println(word));
        }

        /** async call */
        System.out.println("-----------async call-----------");
        IgniteCompute asyncCompute =  ignite.compute().withAsync();
        asyncCompute.call(calls);
        asyncCompute.future().listen(fut -> {
            Collection<Integer> result = (Collection<Integer>)fut.get();
            int t = result.stream().mapToInt(Integer::intValue).sum();
            System.out.println("Total number of characters: " + total);
        });

        /** async run */
        System.out.println("-----------async run-----------");
        Collection<ComputeTaskFuture<?>> futs = new ArrayList<>();
        asyncCompute = ignite.compute().withAsync();
        for (String word : "Print words on different cluster nodes".split(" ")) {
            asyncCompute.run(() -> System.out.println(word));
            futs.add(asyncCompute.future());
        }
        futs.stream().forEach(ComputeTaskFuture::get);

        return "all executed.";
    }

    /** apply测试 */
    @RequestMapping("/apply")
    public @ResponseBody
    String applyTest(HttpServletRequest request, HttpServletResponse response) {
        /** apply */
        System.out.println("-----------apply-----------");
        IgniteCompute compute = ignite.compute();
        Collection<Integer> res = compute.apply(
                String::length,
                Arrays.asList("How many characters".split(" "))
        );
        int total = res.stream().mapToInt(Integer::intValue).sum();
        System.out.println(String.format("the total lengths of all words is [%s].", total));

        /** async apply */
        IgniteCompute asyncCompute = ignite.compute().withAsync();
        res = asyncCompute.apply(
                String::length,
                Arrays.asList("How many characters".split(" "))
        );
        asyncCompute.future().listen(fut -> {
            int t = ((Collection<Integer>)fut.get()).stream().mapToInt(Integer::intValue).sum();
            System.out.println(String.format("Total number of characters: " + total));
        });

        return "all executed.";
    }

2. MapReduce:

在ignite中MapReduce的实现是ComputeTask，其主要方法是map()和reduce()，map()可以控制任务映射到节点的过程，而reduce()则是对最终计算结果集的一个处理。ComputeTask有两个主要实现ComputeTaskAdapter和ComputeTaskSplitAdapter, 主要的区别在于ComputeTaskAdapter需要手动实现map()方法，而ComputeTaskSplitAdapter可以自动映射任务。

ComputeTaskAdapter：

    /**ComputeTaskAdapter*/
    @RequestMapping("/taskMap")
    public @ResponseBody
    String taskMapTest(HttpServletRequest request, HttpServletResponse response) {
        /**ComputeTaskMap*/
        int cnt = ignite.compute().execute(MapExampleCharacterCountTask.class, "Hello Ignite Enable World!");

        System.out.println(String.format(">>> Total number of characters in the phrase is %s.", cnt));

        return "all executed.";
    }

    private static class MapExampleCharacterCountTask extends ComputeTaskAdapter<String, Integer> {
        /**节点映射*/
        @Override
        public Map<? extends ComputeJob, ClusterNode> map(List<ClusterNode> nodes, String arg) throws IgniteException {
            Map<ComputeJob, ClusterNode> map = new HashMap<>();
            Iterator<ClusterNode> it = nodes.iterator();
            for (final String word : arg.split(" ")) {
                // If we used all nodes, restart the iterator.
                if (!it.hasNext()) {
                    it = nodes.iterator();
                }
                ClusterNode node = it.next();
                map.put(new ComputeJobAdapter() {
                    @Override
                    public Object execute() throws IgniteException {
                        System.out.println("-------------------------------------");
                        System.out.println(String.format(">>> Printing [%s] on this node from ignite job.", word));
                        return word.length();
                    }
                }, node);
            }
            return map;
        }
        /**结果汇总*/
        @Override
        public Integer reduce(List<ComputeJobResult> results) throws IgniteException {
            int sum = 0;
            for (ComputeJobResult res : results) {
                sum += res.<Integer>getData();
            }
            return sum;
        }
    }

运行结果:

-------------------------------------
>>> Printing [Ignite] on this node from ignite job.
-------------------------------------
>>> Printing [World!] on this node from ignite job.
>>> Total number of characters in the phrase is 23.

ComputeTaskSplitAdapter：

    /**ComputeTaskSplitAdapter*/
    @RequestMapping("/taskSplit")
    public @ResponseBody
    String taskSplitTest(HttpServletRequest request, HttpServletResponse response) {
        /**ComputeTaskSplitAdapter(自动映射) */
        int result = ignite.compute().execute(SplitExampleDistributedCompute.class, null);
        System.out.println(String.format(">>> result: [%s]", result));

        return "all executed.";
    }

    private static class SplitExampleDistributedCompute extends ComputeTaskSplitAdapter<String, Integer> {

        @Override
        protected Collection<? extends ComputeJob> split(int gridSize, String arg) throws IgniteException {
            Collection<ComputeJob> jobs = new LinkedList<>();
            jobs.add(new ComputeJobAdapter() {
                @Override
                public Object execute() throws IgniteException {
//                    IgniteCache<Long, Student> cache = Ignition.ignite().cache(CacheKeyConstant.STUDENT);
                    IgniteCache<Long, BinaryObject> cache = Ignition.ignite().cache(CacheKeyConstant.STUDENT).withKeepBinary();
                    /**普通查询*/
                    String sql_query = "name = ? and email = ?";
//                    SqlQuery<Long, Student> cSqlQuery = new SqlQuery<>(Student.class, sql_query);
                    SqlQuery<Long, BinaryObject> cSqlQuery = new SqlQuery<>(Student.class, sql_query);
                    cSqlQuery.setReplicatedOnly(true).setArgs("student_54", "student_54gmail.com");
//                  List<Cache.Entry<Long, Student>> result = cache.query(cSqlQuery).getAll();
                    List<Cache.Entry<Long, BinaryObject>> result = cache.query(cSqlQuery).getAll();
                    System.out.println("--------------------");
                    result.stream().map(x -> {
                        Integer studId = x.getValue().field("studId");
                        String name = x.getValue().field("name");
                        return String.format("name=[%s], studId=[%s].", name, studId);
                    }).forEach(System.out::println);
                    System.out.println(String.format("the query size is [%s].", result.size()));
                    return result.size();
                }
            });
            return jobs;
        }

        @Override
        public Integer reduce(List<ComputeJobResult> results) throws IgniteException {
            int sum = results.stream().mapToInt(x -> x.<Integer>getData()).sum();
            return sum;
        }
    }

运行结果:

--------------------
name=[student_54], studId=[54].
the query size is [1].
>>> result: [1]

MapReduce的局限性:

MapReduce适合解决并行和批处理的场景，不适合串行，迭代和递归一类无法并行和分割任务的场景。

分布式计算存在的问题以及注意点

   在使用ignite的分布式计算功能的时候，如果用到了缓存, 并且缓存value不是平台类型(java基础类型)，则需要考虑反序列化的问题。

现有两种解决方案:

• 部署缓存实体类包到ignite节点

缓存实体类得实现Serializable接口，并且得指定serialVersionUID

serialVersionUID表示实体类的当前版本，每个实现Serializable接口的类都有，如果没有的设置该值，java序列化机制会帮你默认生成一个。最好在使用serializable接口时，设定serialVersionUID为某个值，不然当在传输的某一端修改实体类时，serialVersionUID会被虚拟机设置成一个新的值，造成两端的serialVersionUID不一致会发生异常。

public class Student implements Serializable {

    private static final long serialVersionUID = -5941489737545326242L;
    ....
}

将实体类打包成普通jar包，并放在$IGNITE_HOME/libs/路径下面:

注意：打包的时候不能打包成spring-boot的可执行包，要打包成普通jar包，这样相关类才能正常加载。当然如果集群里的节点均为应用节点，则可以不用考虑这个问题。

• 使用二进制对象对缓存进行操作

  Ignite默认使用反序列化值作为最常见的使用场景，要启用BinaryObject处理，需要获得一个IgniteCache的实例然后使用withKeepBinary()方法。启用之后，如果可能，这个标志会确保从缓存返回的对象都是BinaryObject格式的。

 IgniteCache<Long, BinaryObject> cache = ignite.cache("student").withKeepBinary();
 BinaryObject obj = cache.get(k);  //获取二进制对象
 String name = obj.<String>field("name");  //读取二进制对象属性值<使用field方法>

3.并置计算：

affinityCall(...)和affinityRun(...)方法使作业和缓存着数据的节点位于一处，换句话说，给定缓存名字和关系键，这些方法会试图在指定的缓存中定位键所在的节点，然后在那里执行作业。

并置的两种类型以及区别:

ComputeTestController.class

    /**并置计算测试*/
    @RequestMapping("/affinity")
    public @ResponseBody
    String affinityTest(HttpServletRequest request, HttpServletResponse response) {

        /** affinityRun call */
        System.out.println("-----------affinityRun call-----------");
        IgniteCompute compute = ignite.compute();
//        IgniteCompute compute = ignite.compute(ignite.cluster().forRemotes());
        for(int key = 0; key < 100; key++) {
//            final long k = key;
            //生成随机k值
            final long k = IntStream.generate(() -> (int)(System.nanoTime() % 100)).limit(1).findFirst().getAsInt();
            compute.affinityRun(CacheKeyConstant.STUDENT, k, () -> {
                IgniteCache<Long, BinaryObject> cache = ignite.cache(CacheKeyConstant.STUDENT).withKeepBinary();
                BinaryObject obj = cache.get(k);
                if(obj!=null) {
                    System.out.println(String.format("Co-located[key= %s, value= %s]", k, obj.<String>field("name")));
                }
            });
        }

        IgniteCache<Long, BinaryObject> cache = ignite.cache(CacheKeyConstant.STUDENT).withKeepBinary();
        cache.forEach(lo -> compute.affinityRun(CacheKeyConstant.STUDENT, lo.getKey(), () -> {
            System.out.println(lo.getValue().<String>field("name"));
        }));

        return "all executed.";
    }

运行结果:

-----------affinityRun call-----------
student_495
student_496
student_498
...

至此，ignite分布式计算完毕。

apache ignite系列（五）：分布式计算

标签：any   故障   作业   try   实体   result   static   序列   cut   

原文地址：https://www.cnblogs.com/cord/p/9431867.html