-
Notifications
You must be signed in to change notification settings - Fork 26.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
The current RoundRobinLoadBalance implementation performance is not satisfactory #2578
Comments
I will check it today |
In your case 1:
If we modify the weight array to {100, 100, 200, 200, 300, 300, 400, 400, 500, 5000}; So we must optimize this RR algorithm. |
If we modify a weight to 50000 :
then it need 10minutes to finish select. |
@kimmking find the greatest common divisor first? public class RoundRobinLoadBalance extends AbstractLoadBalance {
public static final String NAME = "roundrobin";
private final ConcurrentMap<String, AtomicPositiveInteger> sequences = new ConcurrentHashMap<String, AtomicPositiveInteger>();
@Override
protected <T> Invoker<T> doSelect(List<Invoker<T>> invokers, URL url, Invocation invocation) {
String key = invokers.get(0).getUrl().getServiceKey() + "." + invocation.getMethodName();
int length = invokers.size(); // Number of invokers
int maxWeight = 0; // The maximum weight
int minWeight = Integer.MAX_VALUE; // The minimum weight
final LinkedHashMap<Invoker<T>, IntegerWrapper> invokerToWeightMap = new LinkedHashMap<Invoker<T>, IntegerWrapper>();
int weightSum = 0;
int gcd = 0;
for (int i = 0; i < length; i++) {
int weight = getWeight(invokers.get(i), invocation);
maxWeight = Math.max(maxWeight, weight); // Choose the maximum weight
minWeight = Math.min(minWeight, weight); // Choose the minimum weight
if (weight > 0) {
invokerToWeightMap.put(invokers.get(i), new IntegerWrapper(weight));
weightSum += weight;
if (gcd != 1) {
gcd = gcd(weight, gcd);
}
}
}
AtomicPositiveInteger sequence = sequences.get(key);
if (sequence == null) {
sequences.putIfAbsent(key, new AtomicPositiveInteger());
sequence = sequences.get(key);
}
int currentSequence = sequence.getAndIncrement();
if (maxWeight > 0 && minWeight < maxWeight) {
int mod = currentSequence % (weightSum / gcd);
for (int i = 0; i < maxWeight / gcd; i++) {
for (Map.Entry<Invoker<T>, IntegerWrapper> each : invokerToWeightMap.entrySet()) {
final Invoker<T> k = each.getKey();
final IntegerWrapper v = each.getValue();
if (mod == 0 && v.getValue() > 0) {
return k;
}
if (v.getValue() > 0) {
v.decrement(gcd);
mod--;
}
}
}
}
// Round robin
return invokers.get(currentSequence % length);
}
private static final class IntegerWrapper {
private int value;
public IntegerWrapper(int value) {
this.value = value;
}
public int getValue() {
return value;
}
public void setValue(int value) {
this.value = value;
}
public void decrement() {
this.value--;
}
public void decrement(int a) {
this.value -= a;
}
}
/**
* greatest common divisor
*/
public static int gcd(int number1, int number2) {
if (number2 == 0) {
return number1;
}
while (true) {
if ((number1 = number1 % number2) == 0) {
return number2;
}
if ((number2 = number2 % number1) == 0) {
return number1;
}
}
}
} |
@kimmking The above implementation is also very bad when the greatest common divisor is 1. This performance will be much better. Help me to see if there are any problems. thx! public class RoundRobinLoadBalance extends AbstractLoadBalance {
public static final String NAME = "roundrobin";
private final ConcurrentMap<String, AtomicPositiveInteger> sequences = new ConcurrentHashMap<String, AtomicPositiveInteger>();
private final ConcurrentMap<String, AtomicPositiveInteger> indexSeqs = new ConcurrentHashMap<String, AtomicPositiveInteger>();
@Override
protected <T> Invoker<T> doSelect(List<Invoker<T>> invokers, URL url, Invocation invocation) {
String key = invokers.get(0).getUrl().getServiceKey() + "." + invocation.getMethodName();
int length = invokers.size(); // Number of invokers
int maxWeight = 0; // The maximum weight
int minWeight = Integer.MAX_VALUE; // The minimum weight
final List<Invoker<T>> invokerToWeightList = new ArrayList<>();
for (int i = 0; i < length; i++) {
int weight = getWeight(invokers.get(i), invocation);
maxWeight = Math.max(maxWeight, weight); // Choose the maximum weight
minWeight = Math.min(minWeight, weight); // Choose the minimum weight
if (weight > 0) {
invokerToWeightList.add(invokers.get(i));
}
}
AtomicPositiveInteger sequence = sequences.get(key);
if (sequence == null) {
sequences.putIfAbsent(key, new AtomicPositiveInteger());
sequence = sequences.get(key);
}
AtomicPositiveInteger indexSeq = indexSeqs.get(key);
if (indexSeq == null) {
indexSeqs.putIfAbsent(key, new AtomicPositiveInteger(-1));
indexSeq = indexSeqs.get(key);
}
if (maxWeight > 0 && minWeight < maxWeight) {
length = invokerToWeightList.size();
while (true) {
int index = indexSeq.incrementAndGet() % length;
int currentWeight = sequence.get() % maxWeight;
if (index == 0) {
currentWeight = sequence.incrementAndGet() % maxWeight;
}
if (getWeight(invokerToWeightList.get(index), invocation) > currentWeight) {
return invokerToWeightList.get(index);
}
}
}
// Round robin
return invokers.get(sequence.incrementAndGet() % length);
}
} |
At present, it seems that only a small number of weights can be configured to avoid this problem. |
The second code snippet works well. |
@kimmking OK |
Hi, |
Hi, all I have another implementation of SMOOTH WRR including concurrent locking. What is And it's more smooth especially the factor number is large as: [90 : 100 : 111] test code is below: package com.yoloho.rocketmq.consumers.demo;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.LinkedHashMap;
import java.util.List;
import java.util.Map;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.ConcurrentMap;
import java.util.concurrent.ThreadLocalRandom;
import java.util.concurrent.atomic.AtomicInteger;
import com.alibaba.dubbo.common.utils.AtomicPositiveInteger;
import com.alibaba.fastjson.JSON;
import com.google.common.collect.Lists;
public final class DemoDubbo {
/**
* 假设该接口有10个可用的Invoker
*/
private static final int[] INVOKER_WEIGHT_ARRAY = new int[]{100, 100, 200, 200, 300, 300, 400, 400, 500, 500};
private static final String SERVICE_KEY = "com.test.Test.testMethod";
private static final ConcurrentMap<String, AtomicPositiveInteger> sequences = new ConcurrentHashMap<String, AtomicPositiveInteger>();
private static final ConcurrentMap<String, AtomicPositiveInteger> weightSequences = new ConcurrentHashMap<String, AtomicPositiveInteger>();
public static void main(String[] args) {
AtomicInteger a;
int times = 1000000;
int[] selectArray = new int[]{0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
long start = System.nanoTime();
sequences.clear();
weightSequences.clear();
while (times-- > 0) {
int select = currentSelect();
selectArray[select]++;
}
System.out.println("最新dubbo的RoundRobinLoadBalance耗时:" + (System.nanoTime() - start) / 1000000);
System.out.println("最新dubbo的RoundRobinLoadBalance流量分布:" + JSON.toJSONString(selectArray));
times = 1000000;
selectArray = new int[]{0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
start = System.nanoTime();
sequences.clear();
weightSequences.clear();
while (times-- > 0) {
int select = oldSelect();
selectArray[select]++;
}
System.out.println("dubbo-2.5.3的RoundRobinLoadBalance耗时:" + (System.nanoTime() - start) / 1000000);
System.out.println("dubbo-2.5.3的RoundRobinLoadBalance流量分布:" + JSON.toJSONString(selectArray));
times = 1000000;
selectArray = new int[]{0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
start = System.nanoTime();
sequences.clear();
weightSequences.clear();
while (times-- > 0) {
int select = oldRandomSelect();
selectArray[select]++;
}
System.out.println("dubbo-2.5.3的RandomLoadBalance耗时:" + (System.nanoTime() - start) / 1000000);
System.out.println("dubbo-2.5.3的RandomLoadBalance流量分布:" + JSON.toJSONString(selectArray));
times = 1000000;
selectArray = new int[]{0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
start = System.nanoTime();
while (times-- > 0) {
int select = wrrSelect();
selectArray[select]++;
}
System.out.println("smooth的wrr耗时:" + (System.nanoTime() - start) / 1000000);
System.out.println("smooth的wrr流量分布:" + JSON.toJSONString(selectArray));
}
/**
* 当前最新版本dubbo master分支中实现方式
*
* @return 选择的invoker的index
*/
private static int currentSelect() {
// 为了测试方便,key默认写死
String key = SERVICE_KEY;
// invoker默认是10个
int length = INVOKER_WEIGHT_ARRAY.length; // Number of invokers
int maxWeight = 0; // The maximum weight
int minWeight = Integer.MAX_VALUE; // The minimum weight
final LinkedHashMap<Integer, IntegerWrapper> invokerToWeightMap = new LinkedHashMap<Integer, IntegerWrapper>();
int weightSum = 0;
for (int i = 0; i < length; i++) {
int weight = getWeight(i);
maxWeight = Math.max(maxWeight, weight); // Choose the maximum weight
minWeight = Math.min(minWeight, weight); // Choose the minimum weight
if (weight > 0) {
invokerToWeightMap.put(i, new IntegerWrapper(weight));
weightSum += weight;
}
}
AtomicPositiveInteger sequence = sequences.get(key);
if (sequence == null) {
sequences.putIfAbsent(key, new AtomicPositiveInteger());
sequence = sequences.get(key);
}
int currentSequence = sequence.getAndIncrement();
if (maxWeight > 0 && minWeight < maxWeight) {
int mod = currentSequence % weightSum;
for (int i = 0; i < maxWeight; i++) {
for (Map.Entry<Integer, IntegerWrapper> each : invokerToWeightMap.entrySet()) {
final Integer k = each.getKey();
final IntegerWrapper v = each.getValue();
if (mod == 0 && v.getValue() > 0) {
return k;
}
if (v.getValue() > 0) {
v.decrement();
mod--;
}
}
}
}
// Round robin
return currentSequence % length;
}
/**
* 2.5.3版本的roundrobin方式
*
* @return
*/
private static int oldSelect() {
// 为了测试方便,key默认写死
String key = SERVICE_KEY;
// invoker默认是10个
int length = INVOKER_WEIGHT_ARRAY.length; // Number of invokers
List<Integer> invokers = Lists.newArrayList();
int maxWeight = 0; // 最大权重
int minWeight = Integer.MAX_VALUE; // 最小权重
for (int i = 0; i < length; i++) {
int weight = getWeight(i);
maxWeight = Math.max(maxWeight, weight); // 累计最大权重
minWeight = Math.min(minWeight, weight); // 累计最小权重
}
if (maxWeight > 0 && minWeight < maxWeight) { // 权重不一样
AtomicPositiveInteger weightSequence = weightSequences.get(key);
if (weightSequence == null) {
weightSequences.putIfAbsent(key, new AtomicPositiveInteger());
weightSequence = weightSequences.get(key);
}
int currentWeight = weightSequence.getAndIncrement() % maxWeight;
List<Integer> weightInvokers = new ArrayList<Integer>();
for (int i = 0; i < length; i++) { // 筛选权重大于当前权重基数的Invoker
if (getWeight(i) > currentWeight) {
weightInvokers.add(i);
}
}
int weightLength = weightInvokers.size();
if (weightLength == 1) {
return weightInvokers.get(0);
} else if (weightLength > 1) {
invokers = weightInvokers;
length = invokers.size();
}
}
AtomicPositiveInteger sequence = sequences.get(key);
if (sequence == null) {
sequences.putIfAbsent(key, new AtomicPositiveInteger());
sequence = sequences.get(key);
}
// 取模轮循
return invokers.get(sequence.getAndIncrement() % length);
}
/**
* 2.5.3版本的random方式
*
* @return
*/
private static int oldRandomSelect() {
// 为了测试方便,key默认写死
String key = SERVICE_KEY;
// invoker默认是10个
int length = INVOKER_WEIGHT_ARRAY.length; // Number of invokers
int totalWeight = 0; // 总权重
boolean sameWeight = true; // 权重是否都一样
for (int i = 0; i < length; i++) {
int weight = getWeight(i);
totalWeight += weight; // 累计总权重
if (sameWeight && i > 0
&& weight != getWeight(i - 1)) {
sameWeight = false; // 计算所有权重是否一样
}
}
if (totalWeight > 0 && !sameWeight) {
// 如果权重不相同且权重大于0则按总权重数随机
int offset = ThreadLocalRandom.current().nextInt(totalWeight);
// 并确定随机值落在哪个片断上
for (int i = 0; i < length; i++) {
offset -= getWeight(i);
if (offset < 0) {
return i;
}
}
}
// 如果权重相同或权重为0则均等随机
return ThreadLocalRandom.current().nextInt(length);
}
private static int getWeight(int invokerIndex) {
return INVOKER_WEIGHT_ARRAY[invokerIndex];
}
private static final class IntegerWrapper {
private int value;
public IntegerWrapper(int value) {
this.value = value;
}
public int getValue() {
return value;
}
public void setValue(int value) {
this.value = value;
}
public void decrement() {
this.value--;
}
}
private static class WeightedRoundRobin {
private int weight;
private int current;
public int getWeight() {
return weight;
}
public void setWeight(int weight) {
this.weight = weight;
}
public void setCurrent(int current) {
this.current = current;
}
public int increaseWeight() {
current += weight;
return current;
}
public void sel(int total) {
current -= total;
}
}
private final static ConcurrentMap<String, Map<String, WeightedRoundRobin>> methodWeightMap = new ConcurrentHashMap<String, Map<String, WeightedRoundRobin>>();
private static Map<String, WeightedRoundRobin> getEntry(String key) {
Map<String, WeightedRoundRobin> map = methodWeightMap.get(key);
if (map == null) {
methodWeightMap.putIfAbsent(key, new HashMap<String, WeightedRoundRobin>());
map = methodWeightMap.get(key);
}
return map;
}
private static int wrrSelect() {
// 为了测试方便,key默认写死
String key = SERVICE_KEY;
Map<String, WeightedRoundRobin> map = getEntry(key);
synchronized (map) {
int totalWeight = 0;
int maxCurrent = Integer.MIN_VALUE;
int selectedInvoker = -1;
WeightedRoundRobin selectedWRR = null;
for (int i = 0; i < INVOKER_WEIGHT_ARRAY.length; i++) {
String addr = "server:" + i;
WeightedRoundRobin weightedRoundRobin = map.get(addr);
int weight = INVOKER_WEIGHT_ARRAY[i];
if (weight < 0) {
weight = 0;
}
if (weightedRoundRobin == null) {
weightedRoundRobin = new WeightedRoundRobin();
weightedRoundRobin.setCurrent(0);
weightedRoundRobin.setWeight(weight);
map.put(addr, weightedRoundRobin);
}
if (weight != weightedRoundRobin.getWeight()) {
//weight changed
weightedRoundRobin.setCurrent(0);
weightedRoundRobin.setWeight(weight);
}
int cur = weightedRoundRobin.increaseWeight();
if (cur > maxCurrent) {
maxCurrent = cur;
selectedInvoker = i;
selectedWRR = weightedRoundRobin;
}
totalWeight += weight;
}
if (selectedInvoker >= 0) {
selectedWRR.sel(totalWeight);
return selectedInvoker;
}
}
return -1;
}
} Result:
Is it acceptable? Time cost: 390 ms |
it's good way. |
Environment
Steps to reproduce this issue
Our company currently uses dubbo-2.5.3, LoadBalance chooses roundrobin. After the weight adjustment is made, the distribution of traffic is not in line with expectations. So I saw the implementation of RoundRobinLoadBalance of 2.5.3, which logically led to the distribution of traffic. Not in line with expectations, so I read the latest dubbo version of the implementation, found that the latest version of RoundRobinLoadBalance implementation traffic distribution is in line with expectations, but the performance is not satisfactory.
Test Code:
In the latest version of the RoundRobinLoadBalance implementation, a large number of for loops are used to do the minus one operation. This is not optimistic when the weight distribution is relatively large (hundreds of thousands of weights).
The text was updated successfully, but these errors were encountered: