From b65890da1de34404d775c0b223e9601010c208d4 Mon Sep 17 00:00:00 2001 From: Nick Dimiduk Date: Fri, 16 Apr 2021 09:34:33 -0700 Subject: [PATCH] Revert "HBASE-25739 TableSkewCostFunction need to use aggregated deviation (#3067)" This reverts commit 533c84d330f7314cca4f2f14b4cf6b8454eafb0a. --- .../master/balancer/BaseLoadBalancer.java | 120 ++++-------------- .../balancer/StochasticLoadBalancer.java | 101 ++++++++++----- .../master/balancer/TestBaseLoadBalancer.java | 4 +- 3 files changed, 94 insertions(+), 131 deletions(-) diff --git a/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/BaseLoadBalancer.java b/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/BaseLoadBalancer.java index 99873dd3176..6c2e86d1cdb 100644 --- a/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/BaseLoadBalancer.java +++ b/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/BaseLoadBalancer.java @@ -163,12 +163,7 @@ public abstract class BaseLoadBalancer implements LoadBalancer { int[] initialRegionIndexToServerIndex; //regionIndex -> serverIndex (initial cluster state) int[] regionIndexToTableIndex; //regionIndex -> tableIndex int[][] numRegionsPerServerPerTable; //serverIndex -> tableIndex -> # regions - int[] numRegionsPerTable; // tableIndex -> region count - double[] meanRegionsPerTable; // mean region count per table - double regionSkewByTable; // skew on RS per by table - double minRegionSkewByTable; // min skew on RS per by table - double maxRegionSkewByTable; // max skew on RS per by table - + int[] numMaxRegionsPerTable; //tableIndex -> max number of regions in a single RS int[] regionIndexToPrimaryIndex; //regionIndex -> regionIndex of the primary boolean hasRegionReplicas = false; //whether there is regions with replicas @@ -376,7 +371,6 @@ public abstract class BaseLoadBalancer implements LoadBalancer { numTables = tables.size(); numRegionsPerServerPerTable = new int[numServers][numTables]; - numRegionsPerTable = new int[numTables]; for (int i = 0; i < numServers; i++) { for (int j = 0; j < numTables; j++) { @@ -384,29 +378,21 @@ public abstract class BaseLoadBalancer implements LoadBalancer { } } - for (int i = 0; i < numTables; i++) { - numRegionsPerTable[i] = 0; - } - for (int i=0; i < regionIndexToServerIndex.length; i++) { if (regionIndexToServerIndex[i] >= 0) { numRegionsPerServerPerTable[regionIndexToServerIndex[i]][regionIndexToTableIndex[i]]++; - numRegionsPerTable[regionIndexToTableIndex[i]]++; } } - // Avoid repeated computation for planning - meanRegionsPerTable = new double[numTables]; - maxRegionSkewByTable = 0; - minRegionSkewByTable = 0; - for (int i = 0; i < numTables; i++) { - meanRegionsPerTable[i] = Double.valueOf(numRegionsPerTable[i]) / numServers; - minRegionSkewByTable += Cluster.getMinSkew(numRegionsPerTable[i], numServers); - maxRegionSkewByTable += Cluster.getMaxSkew(numRegionsPerTable[i], numServers); + numMaxRegionsPerTable = new int[numTables]; + for (int[] aNumRegionsPerServerPerTable : numRegionsPerServerPerTable) { + for (tableIndex = 0; tableIndex < aNumRegionsPerServerPerTable.length; tableIndex++) { + if (aNumRegionsPerServerPerTable[tableIndex] > numMaxRegionsPerTable[tableIndex]) { + numMaxRegionsPerTable[tableIndex] = aNumRegionsPerServerPerTable[tableIndex]; + } + } } - computeRegionSkewPerTable(); - for (int i = 0; i < regions.length; i ++) { RegionInfo info = regions[i]; if (RegionReplicaUtil.isDefaultReplica(info)) { @@ -531,53 +517,6 @@ public abstract class BaseLoadBalancer implements LoadBalancer { return numRegions < minLoad; } - /** - * Return the min skew of distribution - */ - public static double getMinSkew(double total, double numServers) { - double mean = total / numServers; - // It's possible that there aren't enough regions to go around - double min; - if (numServers > total) { - min = ((numServers - total) * mean + (1 - mean) * total) ; - } else { - // Some will have 1 more than everything else. - int numHigh = (int) (total - (Math.floor(mean) * numServers)); - int numLow = (int) (numServers - numHigh); - min = numHigh * (Math.ceil(mean) - mean) + numLow * (mean - Math.floor(mean)); - } - return min; - } - - /** - * Return the max deviation of distribution - * Compute max as if all region servers had 0 and one had the sum of all costs. This must be - * a zero sum cost for this to make sense. - */ - public static double getMaxSkew(double total, double numServers) { - double mean = total / numServers; - return (total - mean) + (numServers - 1) * mean; - } - - /** - * Scale the value between 0 and 1. - * - * @param min Min value - * @param max The Max value - * @param value The value to be scaled. - * @return The scaled value. - */ - public static double scale(double min, double max, double value) { - if (max <= min || value <= min) { - return 0; - } - if ((max - min) == 0) { - return 0; - } - - return Math.max(0d, Math.min(1d, (value - min) / (max - min))); - } - /** * Retrieves and lazily initializes a field storing the locality of * every region/server combination @@ -635,21 +574,6 @@ public abstract class BaseLoadBalancer implements LoadBalancer { return regionLoads[region].getLast().getStorefileSizeMB(); } - /** - * Recompute the region skew during init or plan of moves. - */ - private void computeRegionSkewPerTable() { - // reinitialize for recomputation - regionSkewByTable = 0; - - for (int[] aNumRegionsPerServerPerTable : numRegionsPerServerPerTable) { - for (int tableIndex = 0; tableIndex < aNumRegionsPerServerPerTable.length; tableIndex++) { - regionSkewByTable += Math.abs(aNumRegionsPerServerPerTable[tableIndex] - - meanRegionsPerTable[tableIndex]); - } - } - } - /** * Computes and caches the locality for each region/rack combinations, * as well as storing a mapping of region -> server and region -> rack such that server @@ -905,20 +829,22 @@ public abstract class BaseLoadBalancer implements LoadBalancer { int tableIndex = regionIndexToTableIndex[region]; if (oldServer >= 0) { numRegionsPerServerPerTable[oldServer][tableIndex]--; - // update regionSkewPerTable for the move from old server - regionSkewByTable += - Math.abs(numRegionsPerServerPerTable[oldServer][tableIndex] - - meanRegionsPerTable[tableIndex]) - - Math.abs(numRegionsPerServerPerTable[oldServer][tableIndex] + 1 - - meanRegionsPerTable[tableIndex]); } numRegionsPerServerPerTable[newServer][tableIndex]++; - // update regionSkewPerTable for the move to new server - regionSkewByTable += - Math.abs(numRegionsPerServerPerTable[newServer][tableIndex] - - meanRegionsPerTable[tableIndex]) - - Math.abs(numRegionsPerServerPerTable[newServer][tableIndex] - 1 - - meanRegionsPerTable[tableIndex]); + + //check whether this caused maxRegionsPerTable in the new Server to be updated + if (numRegionsPerServerPerTable[newServer][tableIndex] > numMaxRegionsPerTable[tableIndex]) { + numMaxRegionsPerTable[tableIndex] = numRegionsPerServerPerTable[newServer][tableIndex]; + } else if (oldServer >= 0 && (numRegionsPerServerPerTable[oldServer][tableIndex] + 1) + == numMaxRegionsPerTable[tableIndex]) { + //recompute maxRegionsPerTable since the previous value was coming from the old server + numMaxRegionsPerTable[tableIndex] = 0; + for (int[] aNumRegionsPerServerPerTable : numRegionsPerServerPerTable) { + if (aNumRegionsPerServerPerTable[tableIndex] > numMaxRegionsPerTable[tableIndex]) { + numMaxRegionsPerTable[tableIndex] = aNumRegionsPerServerPerTable[tableIndex]; + } + } + } // update for servers int primary = regionIndexToPrimaryIndex[region]; @@ -1088,7 +1014,7 @@ public abstract class BaseLoadBalancer implements LoadBalancer { .append(Arrays.toString(serverIndicesSortedByRegionCount)) .append(", regionsPerServer=").append(Arrays.deepToString(regionsPerServer)); - desc.append(", regionSkewByTable=").append(regionSkewByTable) + desc.append(", numMaxRegionsPerTable=").append(Arrays.toString(numMaxRegionsPerTable)) .append(", numRegions=").append(numRegions).append(", numServers=").append(numServers) .append(", numTables=").append(numTables).append(", numMovedRegions=") .append(numMovedRegions).append('}'); diff --git a/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/StochasticLoadBalancer.java b/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/StochasticLoadBalancer.java index e8a7788b7f7..b40779ade23 100644 --- a/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/StochasticLoadBalancer.java +++ b/hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/StochasticLoadBalancer.java @@ -762,7 +762,6 @@ public class StochasticLoadBalancer extends BaseLoadBalancer { boolean isNeeded() { return true; } - float getMultiplier() { return multiplier; } @@ -771,39 +770,35 @@ public class StochasticLoadBalancer extends BaseLoadBalancer { this.multiplier = m; } - /** - * Called once per LB invocation to give the cost function + /** Called once per LB invocation to give the cost function * to initialize it's state, and perform any costly calculation. */ void init(Cluster cluster) { this.cluster = cluster; } - /** - * Called once per cluster Action to give the cost function + /** Called once per cluster Action to give the cost function * an opportunity to update it's state. postAction() is always * called at least once before cost() is called with the cluster - * that this action is performed on. - */ + * that this action is performed on. */ void postAction(Action action) { switch (action.type) { - case NULL: - break; - case ASSIGN_REGION: - AssignRegionAction ar = (AssignRegionAction) action; - regionMoved(ar.region, -1, ar.server); - break; - case MOVE_REGION: - MoveRegionAction mra = (MoveRegionAction) action; - regionMoved(mra.region, mra.fromServer, mra.toServer); - break; - case SWAP_REGIONS: - SwapRegionsAction a = (SwapRegionsAction) action; - regionMoved(a.fromRegion, a.fromServer, a.toServer); - regionMoved(a.toRegion, a.toServer, a.fromServer); - break; - default: - throw new RuntimeException("Uknown action:" + action.type); + case NULL: break; + case ASSIGN_REGION: + AssignRegionAction ar = (AssignRegionAction) action; + regionMoved(ar.region, -1, ar.server); + break; + case MOVE_REGION: + MoveRegionAction mra = (MoveRegionAction) action; + regionMoved(mra.region, mra.fromServer, mra.toServer); + break; + case SWAP_REGIONS: + SwapRegionsAction a = (SwapRegionsAction) action; + regionMoved(a.fromRegion, a.fromServer, a.toServer); + regionMoved(a.toRegion, a.toServer, a.fromServer); + break; + default: + throw new RuntimeException("Uknown action:" + action.type); } } @@ -827,25 +822,59 @@ public class StochasticLoadBalancer extends BaseLoadBalancer { double total = getSum(stats); double count = stats.length; - double mean = total / count; + double mean = total/count; - for (int i = 0; i < stats.length; i++) { + // Compute max as if all region servers had 0 and one had the sum of all costs. This must be + // a zero sum cost for this to make sense. + double max = ((count - 1) * mean) + (total - mean); + + // It's possible that there aren't enough regions to go around + double min; + if (count > total) { + min = ((count - total) * mean) + ((1 - mean) * total); + } else { + // Some will have 1 more than everything else. + int numHigh = (int) (total - (Math.floor(mean) * count)); + int numLow = (int) (count - numHigh); + + min = (numHigh * (Math.ceil(mean) - mean)) + (numLow * (mean - Math.floor(mean))); + + } + min = Math.max(0, min); + for (int i=0; i