[BALANCER] Demonstration of ResourceBalancer & ResourceUsageHint
#1674
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Now support One/Five/Fifteen minute rate. By default use the one minute rate.
garyparrot
commented
Apr 25, 2023
Comment on lines
+25
to
+35
| public interface ResourceUsageHint { | ||
|
|
||
| ResourceUsage evaluateClusterResourceUsage( | ||
| ClusterInfo clusterInfo, ClusterBean clusterBean, Replica target); | ||
|
|
||
| ResourceUsage evaluateReplicaResourceUsage( | ||
| ClusterInfo clusterInfo, ClusterBean clusterBean, Replica target); | ||
|
|
||
| Collection<ResourceCapacity> evaluateClusterResourceCapacity( | ||
| ClusterInfo clusterInfo, ClusterBean clusterBean); | ||
| } |
There was a problem hiding this comment.
這段是 ResourceUsageHint 的實作，裡面就三個函數
evaluateClusterResourceCapacity 回傳這個 ResourceUsageHint 管理哪些 ResourceCapacity，每個 ResourceCapacity 描述的建議值，還有提供檢查解合法的 Predicate 和比較優劣程度的函數 idealness 和對應的 comparator(可能後面會移除)
Comment on lines
+233
to
+269
| Collection<ResourceCapacity> evaluateIngressResourceCapacity( | ||
| ClusterInfo clusterInfo, ClusterBean clusterBean) { | ||
| final var cachedCalculation = | ||
| calculationCache.computeIfAbsent(clusterBean, CachedCalculation::new); | ||
| final var sumIngress = | ||
| cachedCalculation.partitionIngressRate.values().stream().mapToDouble(x -> x).sum(); | ||
| final var avgIngressPerBroker = sumIngress / clusterInfo.brokers().size(); | ||
|
|
||
| Set<ResourceCapacity> a = clusterInfo.brokers().stream() | ||
| .map( | ||
| broker -> | ||
| new NetworkResourceCapacity( | ||
| NETWORK_COST_BROKER_RESOURCE_PREFIX_INGRESS + broker.id(), avgIngressPerBroker)) | ||
| .collect(Collectors.toSet()); | ||
|
|
||
| a.add(new NetworkResourceCapacity(NETWORK_COST_REPLICA_RESOURCE_PREFIX_INGRESS, 0)); | ||
| return a; | ||
| } | ||
|
|
||
| Collection<ResourceCapacity> evaluateEgressResourceCapacity( | ||
| ClusterInfo clusterInfo, ClusterBean clusterBean) { | ||
| final var cachedCalculation = | ||
| calculationCache.computeIfAbsent(clusterBean, CachedCalculation::new); | ||
| final var sumEgress = | ||
| cachedCalculation.partitionEgressRate.values().stream().mapToDouble(x -> x).sum(); | ||
| final var avgEgressPerBroker = sumEgress / clusterInfo.brokers().size(); | ||
|
|
||
| Set<ResourceCapacity> a = clusterInfo.brokers().stream() | ||
| .map( | ||
| broker -> | ||
| new NetworkResourceCapacity( | ||
| NETWORK_COST_BROKER_RESOURCE_PREFIX_EGRESS + broker.id(), avgEgressPerBroker)) | ||
| .collect(Collectors.toSet()); | ||
|
|
||
| a.add(new NetworkResourceCapacity(NETWORK_COST_REPLICA_RESOURCE_PREFIX_EGRESS, 0)); | ||
| return a; | ||
| } |
There was a problem hiding this comment.
這是 NetworkCost 會回傳的 ResourceCapacity，他會希望每個 Broker 的網路流量趨近 整體叢集流量/節點數量。然後有對應的 idealness 和 comparator 實作，Predicator 的部分沒有，因為他不是 constraint。
There was a problem hiding this comment.
這一段在這個部分
Comment on lines
+126
to
+164
| public Collection<ResourceCapacity> evaluateClusterResourceCapacity(ClusterInfo clusterInfo, ClusterBean clusterBean) { | ||
| return List.of(new ResourceCapacity() { | ||
| @Override | ||
| public String resourceName() { | ||
| return "migrated_leaders"; | ||
| } | ||
|
|
||
| @Override | ||
| public double optimalUsage() { | ||
| return config.string(MAX_MIGRATE_LEADER_KEY).map(Long::parseLong).orElse(Long.MAX_VALUE); | ||
| } | ||
|
|
||
| @Override | ||
| public double idealness(ResourceUsage usage) { | ||
| var a = usage.usage().getOrDefault(resourceName(), 0.0) / optimalUsage(); | ||
| return a; | ||
| } | ||
|
|
||
| @Override | ||
| public Comparator<ResourceUsage> usageIdealnessComparator() { | ||
| return Comparator.comparingDouble(this::idealness); | ||
| // return (a,b) -> { | ||
| // var aa = a.usage().getOrDefault(resourceName(), 0.0); | ||
| // var bb = b.usage().getOrDefault(resourceName(), 0.0); | ||
|
|
||
| // if(aa >= optimalUsage() && bb >= optimalUsage()) return 0; | ||
| // if(aa >= optimalUsage()) return 1; | ||
| // if(bb >= optimalUsage()) return -1; | ||
| // return 0; | ||
| // }; | ||
| // return Comparator.comparingDouble(ru -> ru.usage().getOrDefault(resourceName(), 0.0)); | ||
| } | ||
|
|
||
| @Override | ||
| public Predicate<ResourceUsage> usageValidnessPredicate() { | ||
| return (ru) -> ru.usage().getOrDefault(resourceName(), 0.0) < optimalUsage(); | ||
| } | ||
| }); | ||
| } |
There was a problem hiding this comment.
上面是修正後的 ReplicaLeaderCost，他有一個 Predicate 描述輸入的叢集 ResourceUsage 是不是合法的 (不超過特定數目的 leader 被變更)。
| import org.astraea.common.cost.ResourceUsage; | ||
| import org.astraea.common.cost.ResourceUsageHint; | ||
|
|
||
| public class ResourceBalancer implements Balancer { |
There was a problem hiding this comment.
這是目前 POC 版本的 Resource Balancer 的實作
Comment on lines
+360
to
+368
| static Comparator<Replica> usageDominationComparator( | ||
| List<ResourceCapacity> resourceCapacities, Function<Replica, ResourceUsage> usageHints) { | ||
| var cmp = Comparator.<ResourceUsage>comparingDouble(u -> resourceCapacities.stream() | ||
| .mapToDouble(c -> c.idealness(u)) | ||
| .average() | ||
| .orElseThrow()); | ||
|
|
||
| return Comparator.comparing(usageHints, cmp).reversed(); | ||
| } |
There was a problem hiding this comment.
這個 Comparator 被 ResourceBalancer 用來衡量哪個 Replica 是重要的
Comment on lines
+385
to
+407
| static Comparator<ResourceUsage> usageIdealnessDominationComparator(ResourceUsage base, List<ResourceCapacity> resourceCapacities) { | ||
| var comparators = | ||
| resourceCapacities.stream() | ||
| .map(ResourceCapacity::usageIdealnessComparator) | ||
| .collect(Collectors.toUnmodifiableSet()); | ||
|
|
||
| Comparator<ResourceUsage> dominatedCmp = (lhs, rhs) -> { | ||
| var dominatedByL = comparators.stream().filter(e -> e.compare(lhs, rhs) <= 0).count(); | ||
| var dominatedByR = comparators.stream().filter(e -> e.compare(rhs, lhs) <= 0).count(); | ||
|
|
||
| return -Long.compare(dominatedByL, dominatedByR); | ||
| }; | ||
|
|
||
| // return usageIdealnessDominationComparator(resourceCapacities) | ||
| // .thenComparingDouble(usage -> resourceCapacities.stream() | ||
| // .mapToDouble(ca -> ca.idealness(usage)) | ||
| // .average() | ||
| // .orElseThrow()); | ||
| return dominatedCmp.thenComparingDouble(usage -> resourceCapacities.stream() | ||
| .mapToDouble(ca -> ca.idealness(usage)) | ||
| .average() | ||
| .orElseThrow()); | ||
| } |
There was a problem hiding this comment.
這個 Comparator 被 ResourceBalancer 用來衡量每個 tweak 的潛力，裡面的實作是一個很假的 dominant sort，實際上這個 Comparator 應該會讓某些輸入的排序出現 violation，不過目前的實驗情境因為一些巧妙的關係所以沒有觸發，後面也許會改正。
Comment on lines
+188
to
+260
| private void search( | ||
| Consumer<List<Replica>> updateAnswer, | ||
| int next, | ||
| List<Replica> originalReplicas, | ||
| Map<TopicPartition, List<Replica>> currentAllocation, | ||
| ResourceUsage currentResourceUsage) { | ||
| if (System.currentTimeMillis() > deadline) | ||
| return; | ||
| if (originalReplicas.size() == next) { | ||
| // if this is a complete answer, call update function and return | ||
| updateAnswer.accept( | ||
| currentAllocation.entrySet().stream() | ||
| .flatMap(x -> x.getValue().stream()) | ||
| .collect(Collectors.toUnmodifiableList())); | ||
| } else { | ||
| var nextReplica = originalReplicas.get(next); | ||
|
|
||
| List<Map.Entry<ResourceUsage, Tweak>> possibleTweaks = | ||
| tweaks(currentAllocation, nextReplica).stream() | ||
| .map( | ||
| tweaks -> { | ||
| var usageAfterTweaked = new ResourceUsage(currentResourceUsage.usage()); | ||
| tweaks.toRemove.stream() | ||
| .flatMap(this::evaluateReplicaUsage) | ||
| .forEach(usageAfterTweaked::removeUsage); | ||
| tweaks.toReplace.stream() | ||
| .flatMap(this::evaluateReplicaUsage) | ||
| .forEach(usageAfterTweaked::mergeUsage); | ||
|
|
||
| return Map.entry(usageAfterTweaked, tweaks); | ||
| }) | ||
| .filter(e -> feasibleUsage.test(e.getKey())) | ||
| .sorted(Map.Entry.comparingByKey(usageIdealnessDominationComparator(currentResourceUsage, this.resourceCapacities))) | ||
| // TODO: maybe change to probability style | ||
| .limit(trials(next)) | ||
| .collect(Collectors.toUnmodifiableList()); | ||
|
|
||
| for (Map.Entry<ResourceUsage, Tweak> entry : possibleTweaks) { | ||
| // the tweak we are going to use | ||
| var newResourceUsage = entry.getKey(); | ||
| var tweaks = entry.getValue(); | ||
|
|
||
| // replace the replicas | ||
| tweaks.toRemove.stream() | ||
| .filter(replica -> !currentAllocation.get(replica.topicPartition()).remove(replica)) | ||
| .forEach( | ||
| nonexistReplica -> { | ||
| throw new IllegalStateException( | ||
| "Attempt to remove " | ||
| + nonexistReplica.topicPartitionReplica() | ||
| + " but it does not exists"); | ||
| }); | ||
| tweaks.toReplace.forEach( | ||
| replica -> currentAllocation.get(replica.topicPartition()).add(replica)); | ||
|
|
||
| // start next search stage | ||
| search(updateAnswer, next + 1, originalReplicas, currentAllocation, newResourceUsage); | ||
|
|
||
| // undo the tweak, restore the previous state | ||
| tweaks.toReplace.stream() | ||
| .filter(replica -> !currentAllocation.get(replica.topicPartition()).remove(replica)) | ||
| .forEach( | ||
| nonexistReplica -> { | ||
| throw new IllegalStateException( | ||
| "Attempt to remove " | ||
| + nonexistReplica.topicPartitionReplica() | ||
| + " but it does not exists"); | ||
| }); | ||
| tweaks.toRemove.forEach( | ||
| replica -> currentAllocation.get(replica.topicPartition()).add(replica)); | ||
| } | ||
| } | ||
| } |
There was a problem hiding this comment.
搜尋演算法的主體在這裡，每個 replica 會遞迴一次，如果有 10000 個 replica 則這個函數會呼叫 10000 次，在預設的 JVM stack size 下會 StackOverflow，目前的解法是調整 -Xss20m，這個數字足以容納 10000+ 次的 recursion，後面比較正式的版本會再處理這部分。
garyparrot
changed the title
ResourceBalancer & ResourceUsageHintResourceBalancer & ResourceUsageHint
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
這個 PR 描述 POC 版本的
ResourceBalancer實作和ResourceUsageHint框架。這個 PR 純討論用，裡面的程式碼不是以要合併為目標下去撰寫。