负载均衡之平滑加权轮询(Smooth Weighted Round Robin)详解与实现

在分布式系统、负载均衡器、RPC 框架中,常常需要在多个服务实例之间进行流量分配。如果实例的能力不同,就需要根据"权重"来调度请求。常见的负载算法有:

  • 普通轮询 (Round Robin, RR):所有实例均匀分配请求;
  • 加权随机 (Weighted Random):按权重比例随机选择实例;
  • 加权轮询 (Weighted Round Robin, WRR):按照权重比例循环选择实例。

但是,传统的 WRR 有个缺点:容易出现高权重实例连续被选中的情况,短时间内分布不均衡。

为了解决这个问题,Nginx 提出了 平滑加权轮询(Smooth Weighted Round Robin, SWRR) 算法。


1. 为什么需要 SWRR?

举个例子:有三个实例,权重分别是 A=5, B=1, C=1。

传统 WRR 的可能调度序列:

复制代码
A, A, A, A, A, B, C

A 连续五次,虽然长期比例正确,但短期内不平滑,可能导致抖动。

SWRR 的调度序列:

复制代码
A, A, B, A, C, A, A

同样保证了比例 5:1:1,但分布更加均匀,用户体验更稳定。


2. SWRR 算法原理

SWRR 的核心思想是:

每个实例维护一个动态权重 currentWeight,每次调用时:

  1. 累加 :对所有实例执行 currentWeight += weight
  2. 选择 :选择 currentWeight 最大的实例作为本次结果;
  3. 扣减 :对选中的实例执行 currentWeight -= totalWeighttotalWeight 是所有权重之和)

这样不断迭代,currentWeight 会在一个"周期"内上下波动,保证:

  • 长期调度比例符合权重;
  • 短期分布尽量均匀;
  • 周期结束时,currentWeight 回到全 0,进入下一轮。

3. 示例推演(A=5, B=1, C=1)

初始状态:(A=0, B=0, C=0),总权重=7。

调用次序 累加后 (A,B,C) 选中 扣减后 (A,B,C)
1 (5,1,1) A (-2,1,1)
2 (3,2,2) A (-4,2,2)
3 (1,3,3) B (1,-4,3)
4 (6,-3,4) A (-1,-3,4)
5 (4,-2,5) C (4,-2,-2)
6 (9,-1,-1) A (2,-1,-1)
7 (7,0,0) A (0,0,0)

初始: (0,0,0) 累加: (5,1,1)
选择 A → 减 7 → (-2,1,1) 累加: (3,2,2)
选择 A → 减 7 → (-4,2,2) 累加: (1,3,3)
选择 B → 减 7 → (1,-4,3) 累加: (6,-3,4)
选择 A → 减 7 → (-1,-3,4) 累加: (4,-2,5)
选择 C → 减 7 → (4,-2,-2) 累加: (9,-1,-1)
选择 A → 减 7 → (2,-1,-1) 累加: (7,0,0)
选择 A → 减 7 → (0,0,0)

可以看到:

  • 7 次调用刚好符合权重比例 (5:1:1);
  • 结束后状态回到 (0,0,0),进入下一个周期;
  • 周期性地产生 A, A, B, A, C, A, A 这样平滑的序列。

4. Java 实现

下面给出一个 Java 版的 SWRR 负载均衡器 ,它能够在多实例(ModelInstance)之间,按照各自权重 长期比例接近、短期分布尽量平滑 地分配请求。

核心处理流程:

  1. 每次选择前:对每个实例的 currentWeight += weight。

  2. 选出 currentWeight 最大的实例 best。

  3. 令 best.currentWeight -= totalWeight(totalWeight = Σ weight)。

  4. 下次选择重复 1~3。

  5. 这样就能得到"比例接近权重、且不突发"的平滑分布(如 5:1:1 会呈现 A,A,B,A,C,A,A 这种节奏)。

所有权重之和 ≤ 0 时,这里退化为 普通轮询(Round Robin)

java 复制代码
import per.mjn.route_rule.domain.entity.ModelInstance;

import java.util.*;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.locks.ReentrantLock;

/**
 * 平滑加权轮询(Smooth Weighted Round Robin)
 * 按模型粒度加锁,线程安全。
 */
public class WeightedRoundRobinLoadBalancer {

    private static class Weighted {
        int weight;             // 固定权重
        int currentWeight = 0;  // 动态权重
        ModelInstance instance;
    }

    /** 每个模型的状态 */
    private final Map<String, List<Weighted>> stateMap = new ConcurrentHashMap<>();
    /** 每个模型一把锁,避免全局阻塞 */
    private final Map<String, ReentrantLock> locks = new ConcurrentHashMap<>();
    /** 当所有权重 <= 0 时,退化为普通轮询 */
    private final Map<String, AtomicInteger> rrCounters = new ConcurrentHashMap<>();

    public ModelInstance choose(List<ModelInstance> instances, String modelKey) {
        if (instances == null || instances.isEmpty()) return null;
        final String key = String.valueOf(modelKey);

        ReentrantLock lock = locks.computeIfAbsent(key, k -> new ReentrantLock());
        lock.lock();
        try {
            instances.sort(Comparator.comparing(ModelInstance::getId));

            // 初始化/刷新状态
            List<Weighted> state = stateMap.compute(key, (k, old) -> {
                if (old == null || old.size() != instances.size() || !equalsByWeight(old, instances)) {
                    return build(instances);
                }
                return old;
            });

            int totalWeight = state.stream().mapToInt(w -> w.weight).sum();
            if (totalWeight <= 0) {
                AtomicInteger c = rrCounters.computeIfAbsent(key, k -> new AtomicInteger(0));
                int idx = Math.floorMod(c.getAndIncrement(), state.size());
                return state.get(idx).instance;
            }

            Weighted best = null;
            for (Weighted w : state) {
                w.currentWeight += w.weight;
                if (best == null || w.currentWeight > best.currentWeight) {
                    best = w;
                }
            }
            best.currentWeight -= totalWeight;
            return best.instance;
        } finally {
            lock.unlock();
        }
    }

    private boolean equalsByWeight(List<Weighted> old, List<ModelInstance> instances) {
        if (old.size() != instances.size()) return false;
        for (int i = 0; i < old.size(); i++) {
            if (!Objects.equals(old.get(i).instance.getId(), instances.get(i).getId())) return false;
            if (old.get(i).weight != safeWeight(instances.get(i).getWeight())) return false;
        }
        return true;
    }

    private List<Weighted> build(List<ModelInstance> instances) {
        List<Weighted> list = new ArrayList<>();
        for (ModelInstance ins : instances) {
            Weighted w = new Weighted();
            w.weight = safeWeight(ins.getWeight());
            w.instance = ins;
            list.add(w);
        }
        return list;
    }

    private int safeWeight(Integer w) {
        return (w == null || w < 0) ? 1 : w;
    }
}

流程图如下:
是 否 否 是 是 否 开始: choose 方法 instances 是否为空? 返回 null 结束 按实例ID排序 保证稳定顺序 stateMap 是否存在 且与实例及权重一致? 重建状态: currentWeight 置零 并载入权重 读取旧状态 计算总权重: 所有权重之和 总权重 小于等于 0 ? 退化为普通轮询 计算 idx: 计数器的值对 N 取模 返回第 idx 个实例 遍历: 每项的 currentWeight 加上自身权重 选出 currentWeight 最大的实例 记为 best 将 best 的 currentWeight 减去 总权重 返回 best 实例


5. 特点与优势

  • 平滑性:避免了高权重实例集中出现,调度更加均匀;
  • 比例准确:长期来看,命中次数严格符合权重;
  • 周期性:每次执行"权重和"次调用后,状态回到初始值,进入下一个周期;
  • 适合高并发:通过按模型粒度加锁,避免全局瓶颈。

6. 总结

  • 普通轮询:公平,但无法体现实例差异。
  • 加权随机:简单,可能有抖动。
  • 加权轮询:比例正确,但不够平滑。
  • 平滑加权轮询 (SWRR):既保证比例,又让分布均匀,是生产环境负载均衡的常用算法(Nginx、Dubbo 等框架都在用)。

SWRR 看似简单,实则优雅。它通过"累加 + 扣总和"的小技巧,把"按比例"与"平滑性"完美结合,是一个非常值得学习的经典调度算法。如果需要一个既能体现实例权重,又能保持稳定性的算法,那么SWRR 不失为一种好的选择