上一篇讲了普通轮询、加权轮询的两种实现方式，重点讲了平滑加权轮询算法，并在文末留下了悬念：节点出现分配失败时降低有效权重值；成功时提高有效权重值（但不能大于weight值）。

本文在平滑加权轮询算法的基础上讲，还没弄懂的可以看上一篇文章。

现在来模拟实现：平滑加权轮询算法的降权和提权

1.两个关键点

节点宕机时，降低有效权重值；

节点正常时，提高有效权重值（但不能大于weight值）；

注意：降低或提高权重都是针对有效权重。

2.代码实现

2.1.服务节点类

package com.yty.loadbalancingalgorithm.wrr;

/**
 * String ip：负载IP
 * final Integer weight：权重，保存配置的权重
 * Integer effectiveWeight：有效权重，轮询的过程权重可能变化
 * Integer currentWeight：当前权重，比对该值大小获取节点
 *   第一次加权轮询时：currentWeight = weight = effectiveWeight
 *   后面每次加权轮询时：currentWeight 的值都会不断变化，其他权重不变
 * Boolean isAvailable：是否存活
 */
public class ServerNode implements Comparable<ServerNode>{
    private String ip;
    private final Integer weight;
    private Integer effectiveWeight;
    private Integer currentWeight;
    private Boolean isAvailable;

    public ServerNode(String ip, Integer weight){
        this(ip,weight,true);
    }
    public ServerNode(String ip, Integer weight,Boolean isAvailable){
        this.ip = ip;
        this.weight = weight;
        this.effectiveWeight = weight;
        this.currentWeight = weight;
        this.isAvailable = isAvailable;
    }

    public String getIp() {
        return ip;
    }

    public void setIp(String ip) {
        this.ip = ip;
    }

    public Integer getWeight() {
        return weight;
    }

    public Integer getEffectiveWeight() {
        return effectiveWeight;
    }

    public void setEffectiveWeight(Integer effectiveWeight) {
        this.effectiveWeight = effectiveWeight;
    }

    public Integer getCurrentWeight() {
        return currentWeight;
    }

    public void setCurrentWeight(Integer currentWeight) {
        this.currentWeight = currentWeight;
    }

    public Boolean isAvailable() {
        return isAvailable;
    }
    public void setIsAvailable(Boolean isAvailable){
        this.isAvailable = isAvailable;
    }

    // 每成功一次，恢复有效权重1，不超过配置的起始权重
    public void onInvokeSuccess(){
        if(effectiveWeight < weight) effectiveWeight++;
    }
    // 每失败一次，有效权重减少1，无底线的减少
    public void onInvokeFault(){
        effectiveWeight--;
    }

    @Override
    public int compareTo(ServerNode node) {
        return currentWeight > node.currentWeight ? 1 : (currentWeight.equals(node.currentWeight) ? 0 : -1);
    }

    @Override
    public String toString() {
        return "{ip='" + ip + "', weight=" + weight + ", effectiveWeight=" + effectiveWeight
                + ", currentWeight=" + currentWeight + ", isAvailable=" + isAvailable + "}";
    }
}

2.2.平滑轮询算法降权和提权

package com.yty.loadbalancingalgorithm.wrr;

import java.util.ArrayList;
import java.util.List;

/**
 * 加权轮询算法：加入存活状态，降权使宕机权重降低，从而不会被选中
 */
public class WeightedRoundRobinAvailable {

    private static List<ServerNode> serverNodes = new ArrayList<>();
    // 准备模拟数据
    static {
        serverNodes.add(new ServerNode("192.168.1.101",1));// 默认为true
        serverNodes.add(new ServerNode("192.168.1.102",3,false));
        serverNodes.add(new ServerNode("192.168.1.103",2));
    }

    /**
     * 按照当前权重（currentWeight）最大值获取IP
     * @return ServerNode
     */
    public ServerNode selectNode(){
        if (serverNodes.size() <= 0) return null;
        if (serverNodes.size() == 1)
            return (serverNodes.get(0).isAvailable()) ? serverNodes.get(0) : null;
        
        // 权重之和
        Integer totalWeight = 0;
        ServerNode nodeOfMaxWeight = null; // 保存轮询选中的节点信息
        synchronized (WeightedRoundRobinAvailable.class){
            StringBuffer sb1 = new StringBuffer();
            StringBuffer sb2 = new StringBuffer();
            sb1.append(Thread.currentThread().getName()+"==加权轮询--[当前权重]值的变化："+printCurrentWeight(serverNodes));
            // 有限权重总和可能发生变化
            for(ServerNode serverNode : serverNodes){
                totalWeight += serverNode.getEffectiveWeight();
            }

            // 选出当前权重最大的节点
            ServerNode tempNodeOfMaxWeight = serverNodes.get(0);
            for (ServerNode serverNode : serverNodes) {
                if (serverNode.isAvailable()) {
                    serverNode.onInvokeSuccess();//提权
                    sb2.append(Thread.currentThread().getName()+"==[正常节点]："+serverNode+"\n");
                } else {
                    serverNode.onInvokeFault();//降权
                    sb2.append(Thread.currentThread().getName()+"==[宕机节点]："+serverNode+"\n");
                }

                tempNodeOfMaxWeight = tempNodeOfMaxWeight.compareTo(serverNode) > 0 ? tempNodeOfMaxWeight : serverNode;
            }
            // 必须new个新的节点实例来保存信息，否则引用指向同一个堆实例，后面的set操作将会修改节点信息
            nodeOfMaxWeight = new ServerNode(tempNodeOfMaxWeight.getIp(),tempNodeOfMaxWeight.getWeight(),tempNodeOfMaxWeight.isAvailable());
            nodeOfMaxWeight.setEffectiveWeight(tempNodeOfMaxWeight.getEffectiveWeight());
            nodeOfMaxWeight.setCurrentWeight(tempNodeOfMaxWeight.getCurrentWeight());

            // 调整当前权重比：按权重（effectiveWeight）的比例进行调整，确保请求分发合理。
            tempNodeOfMaxWeight.setCurrentWeight(tempNodeOfMaxWeight.getCurrentWeight() - totalWeight);
            sb1.append(" -> "+printCurrentWeight(serverNodes));

            serverNodes.forEach(serverNode -> serverNode.setCurrentWeight(serverNode.getCurrentWeight()+serverNode.getEffectiveWeight()));

            sb1.append(" -> "+printCurrentWeight(serverNodes));
            System.out.print(sb2);  //所有节点的当前信息
            System.out.println(sb1); //打印当前权重变化过程
        }
        return nodeOfMaxWeight;
    }

    // 格式化打印信息
    private String printCurrentWeight(List<ServerNode> serverNodes){
        StringBuffer stringBuffer = new StringBuffer("[");
        serverNodes.forEach(node -> stringBuffer.append(node.getCurrentWeight()+",") );
        return stringBuffer.substring(0, stringBuffer.length() - 1) + "]";
    }

    // 并发测试：两个线程循环获取节点
    public static void main(String[] args) throws InterruptedException {
        // 循环次数
        int loop = 18;

        new Thread(() -> {
            WeightedRoundRobinAvailable weightedRoundRobin1 = new WeightedRoundRobinAvailable();
            for(int i=1;i<=loop;i++){
                ServerNode serverNode = weightedRoundRobin1.selectNode();
                System.out.println(Thread.currentThread().getName()+"==第"+i+"次轮询选中[当前权重最大]的节点：" + serverNode + "\n");
            }
        }).start();
        //
        new Thread(() -> {
            WeightedRoundRobinAvailable weightedRoundRobin2 = new WeightedRoundRobinAvailable();
            for(int i=1;i<=loop;i++){
                ServerNode serverNode = weightedRoundRobin2.selectNode();
                System.out.println(Thread.currentThread().getName()+"==第"+i+"次轮询选中[当前权重最大]的节点：" + serverNode + "\n");
            }
        }).start();

        //main 线程睡了一下，再偷偷把 所有宕机 拉起来：模拟服务器恢复正常
        Thread.sleep(5);
        for (ServerNode serverNode:serverNodes){
            if(!serverNode.isAvailable())
                serverNode.setIsAvailable(true);
        }
    }
}

3.分析结果

执行结果：将执行结果的前中后四次抽出来分析

Thread-0==[正常节点]：{ip='192.168.1.101', weight=1, effectiveWeight=1, currentWeight=1, isAvailable=true}

Thread-0==[宕机节点]：{ip='192.168.1.102', weight=3, effectiveWeight=2, currentWeight=3, isAvailable=false}

Thread-0==[正常节点]：{ip='192.168.1.103', weight=2, effectiveWeight=2, currentWeight=2, isAvailable=true}

Thread-0==加权轮询--[当前权重]值的变化：[1,3,2] -> [1,-3,2] -> [2,-1,4]

Thread-0==第1次轮询选中[当前权重最大]的节点：{ip='192.168.1.102', weight=3, effectiveWeight=2, currentWeight=3, isAvailable=false}

……

Thread-1==[正常节点]：{ip='192.168.1.101', weight=1, effectiveWeight=1, currentWeight=6, isAvailable=true}

Thread-1==[宕机节点]：{ip='192.168.1.102', weight=3, effectiveWeight=-7, currentWeight=-21, isAvailable=false}

Thread-1==[正常节点]：{ip='192.168.1.103', weight=2, effectiveWeight=2, currentWeight=12, isAvailable=true}

Thread-1==加权轮询--[当前权重]值的变化：[6,-21,12] -> [6,-21,15] -> [7,-28,17]

Thread-1==第5次轮询选中[当前权重最大]的节点：{ip='192.168.1.103', weight=2, effectiveWeight=2, currentWeight=12, isAvailable=true}

……

Thread-0==[正常节点]：{ip='192.168.1.101', weight=1, effectiveWeight=1, currentWeight=13, isAvailable=true}

Thread-0==[正常节点]：{ip='192.168.1.102', weight=3, effectiveWeight=3, currentWeight=-19, isAvailable=true}

Thread-0==[正常节点]：{ip='192.168.1.103', weight=2, effectiveWeight=2, currentWeight=12, isAvailable=true}

Thread-0==加权轮询--[当前权重]值的变化：[13,-19,12] -> [7,-19,12] -> [8,-16,14]

Thread-0==第15次轮询选中[当前权重最大]的节点：{ip='192.168.1.101', weight=1, effectiveWeight=1, currentWeight=13, isAvailable=true}

……

Thread-1==[正常节点]：{ip='192.168.1.101', weight=1, effectiveWeight=1, currentWeight=2, isAvailable=true}

Thread-1==[正常节点]：{ip='192.168.1.102', weight=3, effectiveWeight=3, currentWeight=2, isAvailable=true}

Thread-1==[正常节点]：{ip='192.168.1.103', weight=2, effectiveWeight=2, currentWeight=2, isAvailable=true}

Thread-1==加权轮询--[当前权重]值的变化：[2,2,2] -> [2,2,-4] -> [3,5,-2]

Thread-1==第18次轮询选中[当前权重最大]的节点：{ip='192.168.1.103', weight=2, effectiveWeight=2, currentWeight=2, isAvailable=true}

分析

一开始权重最高的节点虽然是宕机了，但是还是会被选中并返回；

“有效权重总和” 和 “当前权重总和”都减少了1，因为设置轮询到失败节点，都会自减1；

到第5次轮询时，当前权重已经变成了[7,-28,17]，可以看出宕机节点越往后当前权重越小，所以后面根本不会再选中宕机节点，虽然没剔除故障节点，但却起到不分配宕机节点；

到第15次轮询时，有效权重已经恢复起始值，当前权重变为[8,-16,14]，当前权重只能慢慢恢复，并不是节点一正常就立即恢复宕机过的节点，起到对故障节点的缓冲恢复（故障过的节点可能还存在问题）；

最后1次轮询时，因为没有宕机节点，所以有效权重不变，当前权重已经恢复[3,5,-2]，如果再轮询一次，那就会访问到一开始故障的节点了。

4.结论

降权起到缓慢“剔除”宕机节点的效果；提权起到缓冲恢复宕机节点的效果。

对比上一篇文章可以看到：

当前权重（currentWeight）：针对的是节点的选择，受有效权重影响，起到缓慢“剔除”宕机节点和缓冲恢复宕机节点的效果，当前权重最高就会被选择；

有效权重（effectiveWeight）：针对的是权重的变化，也即是降权和提权，降权/提权只会直接操作有效权重；

权重（weight）：针对的是存储起始配置，限定有效权重的提权。

Java实现负载均衡算法--轮询和加权轮询

Java往期文章

Java全栈学习路线、学习资源和面试题一条龙

我心里优秀架构师是怎样的？

免费下载经典编程书籍