001/** 002 * Licensed to the Apache Software Foundation (ASF) under one 003 * or more contributor license agreements. See the NOTICE file 004 * distributed with this work for additional information 005 * regarding copyright ownership. The ASF licenses this file 006 * to you under the Apache License, Version 2.0 (the 007 * "License"); you may not use this file except in compliance 008 * with the License. You may obtain a copy of the License at 009 * 010 * http://www.apache.org/licenses/LICENSE-2.0 011 * 012 * Unless required by applicable law or agreed to in writing, software 013 * distributed under the License is distributed on an "AS IS" BASIS, 014 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 015 * See the License for the specific language governing permissions and 016 * limitations under the License. 017 */ 018package org.apache.hadoop.hbase.master.balancer; 019 020import java.util.ArrayDeque; 021import java.util.ArrayList; 022import java.util.Arrays; 023import java.util.Collection; 024import java.util.Collections; 025import java.util.Deque; 026import java.util.HashMap; 027import java.util.LinkedList; 028import java.util.List; 029import java.util.Map; 030import java.util.Random; 031import org.apache.hadoop.conf.Configuration; 032import org.apache.hadoop.hbase.ClusterMetrics; 033import org.apache.hadoop.hbase.HBaseInterfaceAudience; 034import org.apache.hadoop.hbase.HConstants; 035import org.apache.hadoop.hbase.RegionMetrics; 036import org.apache.hadoop.hbase.ServerMetrics; 037import org.apache.hadoop.hbase.ServerName; 038import org.apache.hadoop.hbase.TableName; 039import org.apache.hadoop.hbase.client.RegionInfo; 040import org.apache.hadoop.hbase.master.MasterServices; 041import org.apache.hadoop.hbase.master.RegionPlan; 042import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.Action; 043import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.Action.Type; 044import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.AssignRegionAction; 045import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.LocalityType; 046import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.MoveRegionAction; 047import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.SwapRegionsAction; 048import org.apache.hadoop.hbase.util.EnvironmentEdgeManager; 049import org.apache.yetus.audience.InterfaceAudience; 050import org.slf4j.Logger; 051import org.slf4j.LoggerFactory; 052 053import org.apache.hbase.thirdparty.com.google.common.annotations.VisibleForTesting; 054import org.apache.hbase.thirdparty.com.google.common.base.Optional; 055import org.apache.hbase.thirdparty.com.google.common.collect.Lists; 056 057 058/** 059 * <p>This is a best effort load balancer. Given a Cost function F(C) => x It will 060 * randomly try and mutate the cluster to Cprime. If F(Cprime) < F(C) then the 061 * new cluster state becomes the plan. It includes costs functions to compute the cost of:</p> 062 * <ul> 063 * <li>Region Load</li> 064 * <li>Table Load</li> 065 * <li>Data Locality</li> 066 * <li>Memstore Sizes</li> 067 * <li>Storefile Sizes</li> 068 * </ul> 069 * 070 * 071 * <p>Every cost function returns a number between 0 and 1 inclusive; where 0 is the lowest cost 072 * best solution, and 1 is the highest possible cost and the worst solution. The computed costs are 073 * scaled by their respective multipliers:</p> 074 * 075 * <ul> 076 * <li>hbase.master.balancer.stochastic.regionLoadCost</li> 077 * <li>hbase.master.balancer.stochastic.moveCost</li> 078 * <li>hbase.master.balancer.stochastic.tableLoadCost</li> 079 * <li>hbase.master.balancer.stochastic.localityCost</li> 080 * <li>hbase.master.balancer.stochastic.memstoreSizeCost</li> 081 * <li>hbase.master.balancer.stochastic.storefileSizeCost</li> 082 * </ul> 083 * 084 * <p>In addition to the above configurations, the balancer can be tuned by the following 085 * configuration values:</p> 086 * <ul> 087 * <li>hbase.master.balancer.stochastic.maxMoveRegions which 088 * controls what the max number of regions that can be moved in a single invocation of this 089 * balancer.</li> 090 * <li>hbase.master.balancer.stochastic.stepsPerRegion is the coefficient by which the number of 091 * regions is multiplied to try and get the number of times the balancer will 092 * mutate all servers.</li> 093 * <li>hbase.master.balancer.stochastic.maxSteps which controls the maximum number of times that 094 * the balancer will try and mutate all the servers. The balancer will use the minimum of this 095 * value and the above computation.</li> 096 * </ul> 097 * 098 * <p>This balancer is best used with hbase.master.loadbalance.bytable set to false 099 * so that the balancer gets the full picture of all loads on the cluster.</p> 100 */ 101@InterfaceAudience.LimitedPrivate(HBaseInterfaceAudience.CONFIG) 102@edu.umd.cs.findbugs.annotations.SuppressWarnings(value="IS2_INCONSISTENT_SYNC", 103 justification="Complaint is about costFunctions not being synchronized; not end of the world") 104public class StochasticLoadBalancer extends BaseLoadBalancer { 105 106 protected static final String STEPS_PER_REGION_KEY = 107 "hbase.master.balancer.stochastic.stepsPerRegion"; 108 protected static final String MAX_STEPS_KEY = 109 "hbase.master.balancer.stochastic.maxSteps"; 110 protected static final String RUN_MAX_STEPS_KEY = 111 "hbase.master.balancer.stochastic.runMaxSteps"; 112 protected static final String MAX_RUNNING_TIME_KEY = 113 "hbase.master.balancer.stochastic.maxRunningTime"; 114 protected static final String KEEP_REGION_LOADS = 115 "hbase.master.balancer.stochastic.numRegionLoadsToRemember"; 116 private static final String TABLE_FUNCTION_SEP = "_"; 117 protected static final String MIN_COST_NEED_BALANCE_KEY = 118 "hbase.master.balancer.stochastic.minCostNeedBalance"; 119 120 protected static final Random RANDOM = new Random(System.currentTimeMillis()); 121 private static final Logger LOG = LoggerFactory.getLogger(StochasticLoadBalancer.class); 122 123 Map<String, Deque<BalancerRegionLoad>> loads = new HashMap<>(); 124 125 // values are defaults 126 private int maxSteps = 1000000; 127 private boolean runMaxSteps = false; 128 private int stepsPerRegion = 800; 129 private long maxRunningTime = 30 * 1000 * 1; // 30 seconds. 130 private int numRegionLoadsToRemember = 15; 131 private float minCostNeedBalance = 0.05f; 132 133 private List<CandidateGenerator> candidateGenerators; 134 private CostFromRegionLoadFunction[] regionLoadFunctions; 135 private CostFunction[] costFunctions; // FindBugs: Wants this protected; IS2_INCONSISTENT_SYNC 136 137 // to save and report costs to JMX 138 private Double curOverallCost = 0d; 139 private Double[] tempFunctionCosts; 140 private Double[] curFunctionCosts; 141 142 // Keep locality based picker and cost function to alert them 143 // when new services are offered 144 private LocalityBasedCandidateGenerator localityCandidateGenerator; 145 private ServerLocalityCostFunction localityCost; 146 private RackLocalityCostFunction rackLocalityCost; 147 private RegionReplicaHostCostFunction regionReplicaHostCostFunction; 148 private RegionReplicaRackCostFunction regionReplicaRackCostFunction; 149 private boolean isByTable = false; 150 private TableName tableName = null; 151 152 /** 153 * The constructor that pass a MetricsStochasticBalancer to BaseLoadBalancer to replace its 154 * default MetricsBalancer 155 */ 156 public StochasticLoadBalancer() { 157 super(new MetricsStochasticBalancer()); 158 } 159 160 @Override 161 public void onConfigurationChange(Configuration conf) { 162 setConf(conf); 163 } 164 165 @Override 166 public synchronized void setConf(Configuration conf) { 167 super.setConf(conf); 168 maxSteps = conf.getInt(MAX_STEPS_KEY, maxSteps); 169 stepsPerRegion = conf.getInt(STEPS_PER_REGION_KEY, stepsPerRegion); 170 maxRunningTime = conf.getLong(MAX_RUNNING_TIME_KEY, maxRunningTime); 171 runMaxSteps = conf.getBoolean(RUN_MAX_STEPS_KEY, runMaxSteps); 172 173 numRegionLoadsToRemember = conf.getInt(KEEP_REGION_LOADS, numRegionLoadsToRemember); 174 isByTable = conf.getBoolean(HConstants.HBASE_MASTER_LOADBALANCE_BYTABLE, isByTable); 175 minCostNeedBalance = conf.getFloat(MIN_COST_NEED_BALANCE_KEY, minCostNeedBalance); 176 if (localityCandidateGenerator == null) { 177 localityCandidateGenerator = new LocalityBasedCandidateGenerator(services); 178 } 179 localityCost = new ServerLocalityCostFunction(conf, services); 180 rackLocalityCost = new RackLocalityCostFunction(conf, services); 181 182 if (this.candidateGenerators == null) { 183 candidateGenerators = Lists.newArrayList(); 184 candidateGenerators.add(new RandomCandidateGenerator()); 185 candidateGenerators.add(new LoadCandidateGenerator()); 186 candidateGenerators.add(localityCandidateGenerator); 187 candidateGenerators.add(new RegionReplicaRackCandidateGenerator()); 188 } 189 regionLoadFunctions = new CostFromRegionLoadFunction[] { 190 new ReadRequestCostFunction(conf), 191 new WriteRequestCostFunction(conf), 192 new MemStoreSizeCostFunction(conf), 193 new StoreFileCostFunction(conf) 194 }; 195 regionReplicaHostCostFunction = new RegionReplicaHostCostFunction(conf); 196 regionReplicaRackCostFunction = new RegionReplicaRackCostFunction(conf); 197 costFunctions = new CostFunction[]{ 198 new RegionCountSkewCostFunction(conf), 199 new PrimaryRegionCountSkewCostFunction(conf), 200 new MoveCostFunction(conf), 201 localityCost, 202 rackLocalityCost, 203 new TableSkewCostFunction(conf), 204 regionReplicaHostCostFunction, 205 regionReplicaRackCostFunction, 206 regionLoadFunctions[0], 207 regionLoadFunctions[1], 208 regionLoadFunctions[2], 209 regionLoadFunctions[3], 210 }; 211 curFunctionCosts= new Double[costFunctions.length]; 212 tempFunctionCosts= new Double[costFunctions.length]; 213 LOG.info("Loaded config; maxSteps=" + maxSteps + ", stepsPerRegion=" + stepsPerRegion + 214 ", maxRunningTime=" + maxRunningTime + ", isByTable=" + isByTable + ", etc."); 215 } 216 217 protected void setCandidateGenerators(List<CandidateGenerator> customCandidateGenerators) { 218 this.candidateGenerators = customCandidateGenerators; 219 } 220 221 @Override 222 protected void setSlop(Configuration conf) { 223 this.slop = conf.getFloat("hbase.regions.slop", 0.001F); 224 } 225 226 @Override 227 public synchronized void setClusterMetrics(ClusterMetrics st) { 228 super.setClusterMetrics(st); 229 updateRegionLoad(); 230 for(CostFromRegionLoadFunction cost : regionLoadFunctions) { 231 cost.setClusterMetrics(st); 232 } 233 234 // update metrics size 235 try { 236 // by-table or ensemble mode 237 int tablesCount = isByTable ? services.getTableDescriptors().getAll().size() : 1; 238 int functionsCount = getCostFunctionNames().length; 239 240 updateMetricsSize(tablesCount * (functionsCount + 1)); // +1 for overall 241 } catch (Exception e) { 242 LOG.error("failed to get the size of all tables", e); 243 } 244 } 245 246 /** 247 * Update the number of metrics that are reported to JMX 248 */ 249 public void updateMetricsSize(int size) { 250 if (metricsBalancer instanceof MetricsStochasticBalancer) { 251 ((MetricsStochasticBalancer) metricsBalancer).updateMetricsSize(size); 252 } 253 } 254 255 @Override 256 public synchronized void setMasterServices(MasterServices masterServices) { 257 super.setMasterServices(masterServices); 258 this.localityCost.setServices(masterServices); 259 this.rackLocalityCost.setServices(masterServices); 260 this.localityCandidateGenerator.setServices(masterServices); 261 } 262 263 @Override 264 protected synchronized boolean areSomeRegionReplicasColocated(Cluster c) { 265 regionReplicaHostCostFunction.init(c); 266 if (regionReplicaHostCostFunction.cost() > 0) return true; 267 regionReplicaRackCostFunction.init(c); 268 if (regionReplicaRackCostFunction.cost() > 0) return true; 269 return false; 270 } 271 272 @Override 273 protected boolean needsBalance(Cluster cluster) { 274 ClusterLoadState cs = new ClusterLoadState(cluster.clusterState); 275 if (cs.getNumServers() < MIN_SERVER_BALANCE) { 276 if (LOG.isDebugEnabled()) { 277 LOG.debug("Not running balancer because only " + cs.getNumServers() 278 + " active regionserver(s)"); 279 } 280 return false; 281 } 282 if (areSomeRegionReplicasColocated(cluster)) { 283 return true; 284 } 285 286 double total = 0.0; 287 float sumMultiplier = 0.0f; 288 for (CostFunction c : costFunctions) { 289 float multiplier = c.getMultiplier(); 290 if (multiplier <= 0) { 291 continue; 292 } 293 if (!c.isNeeded()) { 294 LOG.debug("{} not needed", c.getClass().getSimpleName()); 295 continue; 296 } 297 sumMultiplier += multiplier; 298 total += c.cost() * multiplier; 299 } 300 301 if (total <= 0 || sumMultiplier <= 0 302 || (sumMultiplier > 0 && (total / sumMultiplier) < minCostNeedBalance)) { 303 if (LOG.isTraceEnabled()) { 304 LOG.trace("Skipping load balancing because balanced cluster; " + "total cost is " + total 305 + ", sum multiplier is " + sumMultiplier + " min cost which need balance is " 306 + minCostNeedBalance); 307 } 308 return false; 309 } 310 return true; 311 } 312 313 @Override 314 public synchronized List<RegionPlan> balanceCluster(TableName tableName, Map<ServerName, 315 List<RegionInfo>> clusterState) { 316 this.tableName = tableName; 317 return balanceCluster(clusterState); 318 } 319 320 @VisibleForTesting 321 Cluster.Action nextAction(Cluster cluster) { 322 return candidateGenerators.get(RANDOM.nextInt(candidateGenerators.size())) 323 .generate(cluster); 324 } 325 326 /** 327 * Given the cluster state this will try and approach an optimal balance. This 328 * should always approach the optimal state given enough steps. 329 */ 330 @Override 331 public synchronized List<RegionPlan> balanceCluster(Map<ServerName, 332 List<RegionInfo>> clusterState) { 333 List<RegionPlan> plans = balanceMasterRegions(clusterState); 334 if (plans != null || clusterState == null || clusterState.size() <= 1) { 335 return plans; 336 } 337 338 if (masterServerName != null && clusterState.containsKey(masterServerName)) { 339 if (clusterState.size() <= 2) { 340 return null; 341 } 342 clusterState = new HashMap<>(clusterState); 343 clusterState.remove(masterServerName); 344 } 345 346 // On clusters with lots of HFileLinks or lots of reference files, 347 // instantiating the storefile infos can be quite expensive. 348 // Allow turning this feature off if the locality cost is not going to 349 // be used in any computations. 350 RegionLocationFinder finder = null; 351 if ((this.localityCost != null && this.localityCost.getMultiplier() > 0) 352 || (this.rackLocalityCost != null && this.rackLocalityCost.getMultiplier() > 0)) { 353 finder = this.regionFinder; 354 } 355 356 //The clusterState that is given to this method contains the state 357 //of all the regions in the table(s) (that's true today) 358 // Keep track of servers to iterate through them. 359 Cluster cluster = new Cluster(clusterState, loads, finder, rackManager); 360 361 long startTime = EnvironmentEdgeManager.currentTime(); 362 363 initCosts(cluster); 364 365 if (!needsBalance(cluster)) { 366 return null; 367 } 368 369 double currentCost = computeCost(cluster, Double.MAX_VALUE); 370 curOverallCost = currentCost; 371 for (int i = 0; i < this.curFunctionCosts.length; i++) { 372 curFunctionCosts[i] = tempFunctionCosts[i]; 373 } 374 double initCost = currentCost; 375 double newCost = currentCost; 376 377 long computedMaxSteps; 378 if (runMaxSteps) { 379 computedMaxSteps = Math.max(this.maxSteps, 380 ((long)cluster.numRegions * (long)this.stepsPerRegion * (long)cluster.numServers)); 381 } else { 382 long calculatedMaxSteps = (long)cluster.numRegions * (long)this.stepsPerRegion * 383 (long)cluster.numServers; 384 computedMaxSteps = Math.min(this.maxSteps, calculatedMaxSteps); 385 if (calculatedMaxSteps > maxSteps) { 386 LOG.warn("calculatedMaxSteps:{} for loadbalancer's stochastic walk is larger than " 387 + "maxSteps:{}. Hence load balancing may not work well. Setting parameter " 388 + "\"hbase.master.balancer.stochastic.runMaxSteps\" to true can overcome this issue." 389 + "(This config change does not require service restart)", calculatedMaxSteps, 390 maxRunningTime); 391 } 392 } 393 LOG.info("start StochasticLoadBalancer.balancer, initCost=" + currentCost + ", functionCost=" 394 + functionCost() + " computedMaxSteps: " + computedMaxSteps); 395 396 // Perform a stochastic walk to see if we can get a good fit. 397 long step; 398 399 for (step = 0; step < computedMaxSteps; step++) { 400 Cluster.Action action = nextAction(cluster); 401 402 if (action.type == Type.NULL) { 403 continue; 404 } 405 406 cluster.doAction(action); 407 updateCostsWithAction(cluster, action); 408 409 newCost = computeCost(cluster, currentCost); 410 411 // Should this be kept? 412 if (newCost < currentCost) { 413 currentCost = newCost; 414 415 // save for JMX 416 curOverallCost = currentCost; 417 for (int i = 0; i < this.curFunctionCosts.length; i++) { 418 curFunctionCosts[i] = tempFunctionCosts[i]; 419 } 420 } else { 421 // Put things back the way they were before. 422 // TODO: undo by remembering old values 423 Action undoAction = action.undoAction(); 424 cluster.doAction(undoAction); 425 updateCostsWithAction(cluster, undoAction); 426 } 427 428 if (EnvironmentEdgeManager.currentTime() - startTime > 429 maxRunningTime) { 430 break; 431 } 432 } 433 long endTime = EnvironmentEdgeManager.currentTime(); 434 435 metricsBalancer.balanceCluster(endTime - startTime); 436 437 // update costs metrics 438 updateStochasticCosts(tableName, curOverallCost, curFunctionCosts); 439 if (initCost > currentCost) { 440 plans = createRegionPlans(cluster); 441 LOG.info("Finished computing new load balance plan. Computation took {}" + 442 " to try {} different iterations. Found a solution that moves " + 443 "{} regions; Going from a computed cost of {}" + 444 " to a new cost of {}", java.time.Duration.ofMillis(endTime - startTime), 445 step, plans.size(), initCost, currentCost); 446 return plans; 447 } 448 LOG.info("Could not find a better load balance plan. Tried {} different configurations in " + 449 "{}, and did not find anything with a computed cost less than {}", step, 450 java.time.Duration.ofMillis(endTime - startTime), initCost); 451 return null; 452 } 453 454 /** 455 * update costs to JMX 456 */ 457 private void updateStochasticCosts(TableName tableName, Double overall, Double[] subCosts) { 458 if (tableName == null) return; 459 460 // check if the metricsBalancer is MetricsStochasticBalancer before casting 461 if (metricsBalancer instanceof MetricsStochasticBalancer) { 462 MetricsStochasticBalancer balancer = (MetricsStochasticBalancer) metricsBalancer; 463 // overall cost 464 balancer.updateStochasticCost(tableName.getNameAsString(), 465 "Overall", "Overall cost", overall); 466 467 // each cost function 468 for (int i = 0; i < costFunctions.length; i++) { 469 CostFunction costFunction = costFunctions[i]; 470 String costFunctionName = costFunction.getClass().getSimpleName(); 471 Double costPercent = (overall == 0) ? 0 : (subCosts[i] / overall); 472 // TODO: cost function may need a specific description 473 balancer.updateStochasticCost(tableName.getNameAsString(), costFunctionName, 474 "The percent of " + costFunctionName, costPercent); 475 } 476 } 477 } 478 479 private String functionCost() { 480 StringBuilder builder = new StringBuilder(); 481 for (CostFunction c:costFunctions) { 482 builder.append(c.getClass().getSimpleName()); 483 builder.append(" : ("); 484 builder.append(c.getMultiplier()); 485 builder.append(", "); 486 builder.append(c.cost()); 487 builder.append("); "); 488 } 489 return builder.toString(); 490 } 491 492 /** 493 * Create all of the RegionPlan's needed to move from the initial cluster state to the desired 494 * state. 495 * 496 * @param cluster The state of the cluster 497 * @return List of RegionPlan's that represent the moves needed to get to desired final state. 498 */ 499 private List<RegionPlan> createRegionPlans(Cluster cluster) { 500 List<RegionPlan> plans = new LinkedList<>(); 501 for (int regionIndex = 0; 502 regionIndex < cluster.regionIndexToServerIndex.length; regionIndex++) { 503 int initialServerIndex = cluster.initialRegionIndexToServerIndex[regionIndex]; 504 int newServerIndex = cluster.regionIndexToServerIndex[regionIndex]; 505 506 if (initialServerIndex != newServerIndex) { 507 RegionInfo region = cluster.regions[regionIndex]; 508 ServerName initialServer = cluster.servers[initialServerIndex]; 509 ServerName newServer = cluster.servers[newServerIndex]; 510 511 if (LOG.isTraceEnabled()) { 512 LOG.trace("Moving Region " + region.getEncodedName() + " from server " 513 + initialServer.getHostname() + " to " + newServer.getHostname()); 514 } 515 RegionPlan rp = new RegionPlan(region, initialServer, newServer); 516 plans.add(rp); 517 } 518 } 519 return plans; 520 } 521 522 /** 523 * Store the current region loads. 524 */ 525 private synchronized void updateRegionLoad() { 526 // We create a new hashmap so that regions that are no longer there are removed. 527 // However we temporarily need the old loads so we can use them to keep the rolling average. 528 Map<String, Deque<BalancerRegionLoad>> oldLoads = loads; 529 loads = new HashMap<>(); 530 531 clusterStatus.getLiveServerMetrics().forEach((ServerName sn, ServerMetrics sm) -> { 532 sm.getRegionMetrics().forEach((byte[] regionName, RegionMetrics rm) -> { 533 String regionNameAsString = RegionInfo.getRegionNameAsString(regionName); 534 Deque<BalancerRegionLoad> rLoads = oldLoads.get(regionNameAsString); 535 if (rLoads == null) { 536 rLoads = new ArrayDeque<>(numRegionLoadsToRemember + 1); 537 } else if (rLoads.size() >= numRegionLoadsToRemember) { 538 rLoads.remove(); 539 } 540 rLoads.add(new BalancerRegionLoad(rm)); 541 loads.put(regionNameAsString, rLoads); 542 }); 543 }); 544 545 for(CostFromRegionLoadFunction cost : regionLoadFunctions) { 546 cost.setLoads(loads); 547 } 548 } 549 550 protected void initCosts(Cluster cluster) { 551 for (CostFunction c:costFunctions) { 552 c.init(cluster); 553 } 554 } 555 556 protected void updateCostsWithAction(Cluster cluster, Action action) { 557 for (CostFunction c : costFunctions) { 558 c.postAction(action); 559 } 560 } 561 562 /** 563 * Get the names of the cost functions 564 */ 565 public String[] getCostFunctionNames() { 566 if (costFunctions == null) return null; 567 String[] ret = new String[costFunctions.length]; 568 for (int i = 0; i < costFunctions.length; i++) { 569 CostFunction c = costFunctions[i]; 570 ret[i] = c.getClass().getSimpleName(); 571 } 572 573 return ret; 574 } 575 576 /** 577 * This is the main cost function. It will compute a cost associated with a proposed cluster 578 * state. All different costs will be combined with their multipliers to produce a double cost. 579 * 580 * @param cluster The state of the cluster 581 * @param previousCost the previous cost. This is used as an early out. 582 * @return a double of a cost associated with the proposed cluster state. This cost is an 583 * aggregate of all individual cost functions. 584 */ 585 protected double computeCost(Cluster cluster, double previousCost) { 586 double total = 0; 587 588 for (int i = 0; i < costFunctions.length; i++) { 589 CostFunction c = costFunctions[i]; 590 this.tempFunctionCosts[i] = 0.0; 591 592 if (c.getMultiplier() <= 0) { 593 continue; 594 } 595 596 Float multiplier = c.getMultiplier(); 597 Double cost = c.cost(); 598 599 this.tempFunctionCosts[i] = multiplier*cost; 600 total += this.tempFunctionCosts[i]; 601 602 if (total > previousCost) { 603 break; 604 } 605 } 606 607 return total; 608 } 609 610 /** Generates a candidate action to be applied to the cluster for cost function search */ 611 abstract static class CandidateGenerator { 612 abstract Cluster.Action generate(Cluster cluster); 613 614 /** 615 * From a list of regions pick a random one. Null can be returned which 616 * {@link StochasticLoadBalancer#balanceCluster(Map)} recognize as signal to try a region move 617 * rather than swap. 618 * 619 * @param cluster The state of the cluster 620 * @param server index of the server 621 * @param chanceOfNoSwap Chance that this will decide to try a move rather 622 * than a swap. 623 * @return a random {@link RegionInfo} or null if an asymmetrical move is 624 * suggested. 625 */ 626 protected int pickRandomRegion(Cluster cluster, int server, double chanceOfNoSwap) { 627 // Check to see if this is just a move. 628 if (cluster.regionsPerServer[server].length == 0 || RANDOM.nextFloat() < chanceOfNoSwap) { 629 // signal a move only. 630 return -1; 631 } 632 int rand = RANDOM.nextInt(cluster.regionsPerServer[server].length); 633 return cluster.regionsPerServer[server][rand]; 634 635 } 636 protected int pickRandomServer(Cluster cluster) { 637 if (cluster.numServers < 1) { 638 return -1; 639 } 640 641 return RANDOM.nextInt(cluster.numServers); 642 } 643 644 protected int pickRandomRack(Cluster cluster) { 645 if (cluster.numRacks < 1) { 646 return -1; 647 } 648 649 return RANDOM.nextInt(cluster.numRacks); 650 } 651 652 protected int pickOtherRandomServer(Cluster cluster, int serverIndex) { 653 if (cluster.numServers < 2) { 654 return -1; 655 } 656 while (true) { 657 int otherServerIndex = pickRandomServer(cluster); 658 if (otherServerIndex != serverIndex) { 659 return otherServerIndex; 660 } 661 } 662 } 663 664 protected int pickOtherRandomRack(Cluster cluster, int rackIndex) { 665 if (cluster.numRacks < 2) { 666 return -1; 667 } 668 while (true) { 669 int otherRackIndex = pickRandomRack(cluster); 670 if (otherRackIndex != rackIndex) { 671 return otherRackIndex; 672 } 673 } 674 } 675 676 protected Cluster.Action pickRandomRegions(Cluster cluster, 677 int thisServer, 678 int otherServer) { 679 if (thisServer < 0 || otherServer < 0) { 680 return Cluster.NullAction; 681 } 682 683 // Decide who is most likely to need another region 684 int thisRegionCount = cluster.getNumRegions(thisServer); 685 int otherRegionCount = cluster.getNumRegions(otherServer); 686 687 // Assign the chance based upon the above 688 double thisChance = (thisRegionCount > otherRegionCount) ? 0 : 0.5; 689 double otherChance = (thisRegionCount <= otherRegionCount) ? 0 : 0.5; 690 691 int thisRegion = pickRandomRegion(cluster, thisServer, thisChance); 692 int otherRegion = pickRandomRegion(cluster, otherServer, otherChance); 693 694 return getAction(thisServer, thisRegion, otherServer, otherRegion); 695 } 696 697 protected Cluster.Action getAction(int fromServer, int fromRegion, 698 int toServer, int toRegion) { 699 if (fromServer < 0 || toServer < 0) { 700 return Cluster.NullAction; 701 } 702 if (fromRegion > 0 && toRegion > 0) { 703 return new Cluster.SwapRegionsAction(fromServer, fromRegion, 704 toServer, toRegion); 705 } else if (fromRegion > 0) { 706 return new Cluster.MoveRegionAction(fromRegion, fromServer, toServer); 707 } else if (toRegion > 0) { 708 return new Cluster.MoveRegionAction(toRegion, toServer, fromServer); 709 } else { 710 return Cluster.NullAction; 711 } 712 } 713 714 /** 715 * Returns a random iteration order of indexes of an array with size length 716 */ 717 protected List<Integer> getRandomIterationOrder(int length) { 718 ArrayList<Integer> order = new ArrayList<>(length); 719 for (int i = 0; i < length; i++) { 720 order.add(i); 721 } 722 Collections.shuffle(order); 723 return order; 724 } 725 } 726 727 static class RandomCandidateGenerator extends CandidateGenerator { 728 729 @Override 730 Cluster.Action generate(Cluster cluster) { 731 732 int thisServer = pickRandomServer(cluster); 733 734 // Pick the other server 735 int otherServer = pickOtherRandomServer(cluster, thisServer); 736 737 return pickRandomRegions(cluster, thisServer, otherServer); 738 } 739 } 740 741 static class LoadCandidateGenerator extends CandidateGenerator { 742 743 @Override 744 Cluster.Action generate(Cluster cluster) { 745 cluster.sortServersByRegionCount(); 746 int thisServer = pickMostLoadedServer(cluster, -1); 747 int otherServer = pickLeastLoadedServer(cluster, thisServer); 748 749 return pickRandomRegions(cluster, thisServer, otherServer); 750 } 751 752 private int pickLeastLoadedServer(final Cluster cluster, int thisServer) { 753 Integer[] servers = cluster.serverIndicesSortedByRegionCount; 754 755 int index = 0; 756 while (servers[index] == null || servers[index] == thisServer) { 757 index++; 758 if (index == servers.length) { 759 return -1; 760 } 761 } 762 return servers[index]; 763 } 764 765 private int pickMostLoadedServer(final Cluster cluster, int thisServer) { 766 Integer[] servers = cluster.serverIndicesSortedByRegionCount; 767 768 int index = servers.length - 1; 769 while (servers[index] == null || servers[index] == thisServer) { 770 index--; 771 if (index < 0) { 772 return -1; 773 } 774 } 775 return servers[index]; 776 } 777 } 778 779 static class LocalityBasedCandidateGenerator extends CandidateGenerator { 780 781 private MasterServices masterServices; 782 783 LocalityBasedCandidateGenerator(MasterServices masterServices) { 784 this.masterServices = masterServices; 785 } 786 787 @Override 788 Cluster.Action generate(Cluster cluster) { 789 if (this.masterServices == null) { 790 int thisServer = pickRandomServer(cluster); 791 // Pick the other server 792 int otherServer = pickOtherRandomServer(cluster, thisServer); 793 return pickRandomRegions(cluster, thisServer, otherServer); 794 } 795 796 // Randomly iterate through regions until you find one that is not on ideal host 797 for (int region : getRandomIterationOrder(cluster.numRegions)) { 798 int currentServer = cluster.regionIndexToServerIndex[region]; 799 if (currentServer != cluster.getOrComputeRegionsToMostLocalEntities(LocalityType.SERVER)[region]) { 800 Optional<Action> potential = tryMoveOrSwap( 801 cluster, 802 currentServer, 803 region, 804 cluster.getOrComputeRegionsToMostLocalEntities(LocalityType.SERVER)[region] 805 ); 806 if (potential.isPresent()) { 807 return potential.get(); 808 } 809 } 810 } 811 return Cluster.NullAction; 812 } 813 814 /** 815 * Try to generate a move/swap fromRegion between fromServer and toServer such that locality is improved. 816 * Returns empty optional if no move can be found 817 */ 818 private Optional<Action> tryMoveOrSwap(Cluster cluster, 819 int fromServer, 820 int fromRegion, 821 int toServer) { 822 // Try move first. We know apriori fromRegion has the highest locality on toServer 823 if (cluster.serverHasTooFewRegions(toServer)) { 824 return Optional.of(getAction(fromServer, fromRegion, toServer, -1)); 825 } 826 827 // Compare locality gain/loss from swapping fromRegion with regions on toServer 828 double fromRegionLocalityDelta = 829 getWeightedLocality(cluster, fromRegion, toServer) - getWeightedLocality(cluster, fromRegion, fromServer); 830 for (int toRegionIndex : getRandomIterationOrder(cluster.regionsPerServer[toServer].length)) { 831 int toRegion = cluster.regionsPerServer[toServer][toRegionIndex]; 832 double toRegionLocalityDelta = 833 getWeightedLocality(cluster, toRegion, fromServer) - getWeightedLocality(cluster, toRegion, toServer); 834 // If locality would remain neutral or improve, attempt the swap 835 if (fromRegionLocalityDelta + toRegionLocalityDelta >= 0) { 836 return Optional.of(getAction(fromServer, fromRegion, toServer, toRegion)); 837 } 838 } 839 840 return Optional.absent(); 841 } 842 843 private double getWeightedLocality(Cluster cluster, int region, int server) { 844 return cluster.getOrComputeWeightedLocality(region, server, LocalityType.SERVER); 845 } 846 847 void setServices(MasterServices services) { 848 this.masterServices = services; 849 } 850 } 851 852 /** 853 * Generates candidates which moves the replicas out of the region server for 854 * co-hosted region replicas 855 */ 856 static class RegionReplicaCandidateGenerator extends CandidateGenerator { 857 858 RandomCandidateGenerator randomGenerator = new RandomCandidateGenerator(); 859 860 /** 861 * Randomly select one regionIndex out of all region replicas co-hosted in the same group 862 * (a group is a server, host or rack) 863 * @param primariesOfRegionsPerGroup either Cluster.primariesOfRegionsPerServer, 864 * primariesOfRegionsPerHost or primariesOfRegionsPerRack 865 * @param regionsPerGroup either Cluster.regionsPerServer, regionsPerHost or regionsPerRack 866 * @param regionIndexToPrimaryIndex Cluster.regionsIndexToPrimaryIndex 867 * @return a regionIndex for the selected primary or -1 if there is no co-locating 868 */ 869 int selectCoHostedRegionPerGroup(int[] primariesOfRegionsPerGroup, int[] regionsPerGroup 870 , int[] regionIndexToPrimaryIndex) { 871 int currentPrimary = -1; 872 int currentPrimaryIndex = -1; 873 int selectedPrimaryIndex = -1; 874 double currentLargestRandom = -1; 875 // primariesOfRegionsPerGroup is a sorted array. Since it contains the primary region 876 // ids for the regions hosted in server, a consecutive repetition means that replicas 877 // are co-hosted 878 for (int j = 0; j <= primariesOfRegionsPerGroup.length; j++) { 879 int primary = j < primariesOfRegionsPerGroup.length 880 ? primariesOfRegionsPerGroup[j] : -1; 881 if (primary != currentPrimary) { // check for whether we see a new primary 882 int numReplicas = j - currentPrimaryIndex; 883 if (numReplicas > 1) { // means consecutive primaries, indicating co-location 884 // decide to select this primary region id or not 885 double currentRandom = RANDOM.nextDouble(); 886 // we don't know how many region replicas are co-hosted, we will randomly select one 887 // using reservoir sampling (http://gregable.com/2007/10/reservoir-sampling.html) 888 if (currentRandom > currentLargestRandom) { 889 selectedPrimaryIndex = currentPrimary; 890 currentLargestRandom = currentRandom; 891 } 892 } 893 currentPrimary = primary; 894 currentPrimaryIndex = j; 895 } 896 } 897 898 // we have found the primary id for the region to move. Now find the actual regionIndex 899 // with the given primary, prefer to move the secondary region. 900 for (int j = 0; j < regionsPerGroup.length; j++) { 901 int regionIndex = regionsPerGroup[j]; 902 if (selectedPrimaryIndex == regionIndexToPrimaryIndex[regionIndex]) { 903 // always move the secondary, not the primary 904 if (selectedPrimaryIndex != regionIndex) { 905 return regionIndex; 906 } 907 } 908 } 909 return -1; 910 } 911 912 @Override 913 Cluster.Action generate(Cluster cluster) { 914 int serverIndex = pickRandomServer(cluster); 915 if (cluster.numServers <= 1 || serverIndex == -1) { 916 return Cluster.NullAction; 917 } 918 919 int regionIndex = selectCoHostedRegionPerGroup( 920 cluster.primariesOfRegionsPerServer[serverIndex], 921 cluster.regionsPerServer[serverIndex], 922 cluster.regionIndexToPrimaryIndex); 923 924 // if there are no pairs of region replicas co-hosted, default to random generator 925 if (regionIndex == -1) { 926 // default to randompicker 927 return randomGenerator.generate(cluster); 928 } 929 930 int toServerIndex = pickOtherRandomServer(cluster, serverIndex); 931 int toRegionIndex = pickRandomRegion(cluster, toServerIndex, 0.9f); 932 return getAction(serverIndex, regionIndex, toServerIndex, toRegionIndex); 933 } 934 } 935 936 /** 937 * Generates candidates which moves the replicas out of the rack for 938 * co-hosted region replicas in the same rack 939 */ 940 static class RegionReplicaRackCandidateGenerator extends RegionReplicaCandidateGenerator { 941 @Override 942 Cluster.Action generate(Cluster cluster) { 943 int rackIndex = pickRandomRack(cluster); 944 if (cluster.numRacks <= 1 || rackIndex == -1) { 945 return super.generate(cluster); 946 } 947 948 int regionIndex = selectCoHostedRegionPerGroup( 949 cluster.primariesOfRegionsPerRack[rackIndex], 950 cluster.regionsPerRack[rackIndex], 951 cluster.regionIndexToPrimaryIndex); 952 953 // if there are no pairs of region replicas co-hosted, default to random generator 954 if (regionIndex == -1) { 955 // default to randompicker 956 return randomGenerator.generate(cluster); 957 } 958 959 int serverIndex = cluster.regionIndexToServerIndex[regionIndex]; 960 int toRackIndex = pickOtherRandomRack(cluster, rackIndex); 961 962 int rand = RANDOM.nextInt(cluster.serversPerRack[toRackIndex].length); 963 int toServerIndex = cluster.serversPerRack[toRackIndex][rand]; 964 int toRegionIndex = pickRandomRegion(cluster, toServerIndex, 0.9f); 965 return getAction(serverIndex, regionIndex, toServerIndex, toRegionIndex); 966 } 967 } 968 969 /** 970 * Base class of StochasticLoadBalancer's Cost Functions. 971 */ 972 abstract static class CostFunction { 973 974 private float multiplier = 0; 975 976 protected Cluster cluster; 977 978 CostFunction(Configuration c) { 979 } 980 981 boolean isNeeded() { 982 return true; 983 } 984 float getMultiplier() { 985 return multiplier; 986 } 987 988 void setMultiplier(float m) { 989 this.multiplier = m; 990 } 991 992 /** Called once per LB invocation to give the cost function 993 * to initialize it's state, and perform any costly calculation. 994 */ 995 void init(Cluster cluster) { 996 this.cluster = cluster; 997 } 998 999 /** Called once per cluster Action to give the cost function 1000 * an opportunity to update it's state. postAction() is always 1001 * called at least once before cost() is called with the cluster 1002 * that this action is performed on. */ 1003 void postAction(Action action) { 1004 switch (action.type) { 1005 case NULL: break; 1006 case ASSIGN_REGION: 1007 AssignRegionAction ar = (AssignRegionAction) action; 1008 regionMoved(ar.region, -1, ar.server); 1009 break; 1010 case MOVE_REGION: 1011 MoveRegionAction mra = (MoveRegionAction) action; 1012 regionMoved(mra.region, mra.fromServer, mra.toServer); 1013 break; 1014 case SWAP_REGIONS: 1015 SwapRegionsAction a = (SwapRegionsAction) action; 1016 regionMoved(a.fromRegion, a.fromServer, a.toServer); 1017 regionMoved(a.toRegion, a.toServer, a.fromServer); 1018 break; 1019 default: 1020 throw new RuntimeException("Uknown action:" + action.type); 1021 } 1022 } 1023 1024 protected void regionMoved(int region, int oldServer, int newServer) { 1025 } 1026 1027 abstract double cost(); 1028 1029 /** 1030 * Function to compute a scaled cost using {@link org.apache.commons.math3.stat.descriptive.DescriptiveStatistics}. 1031 * It assumes that this is a zero sum set of costs. It assumes that the worst case 1032 * possible is all of the elements in one region server and the rest having 0. 1033 * 1034 * @param stats the costs 1035 * @return a scaled set of costs. 1036 */ 1037 protected double costFromArray(double[] stats) { 1038 double totalCost = 0; 1039 double total = getSum(stats); 1040 1041 double count = stats.length; 1042 double mean = total/count; 1043 1044 // Compute max as if all region servers had 0 and one had the sum of all costs. This must be 1045 // a zero sum cost for this to make sense. 1046 double max = ((count - 1) * mean) + (total - mean); 1047 1048 // It's possible that there aren't enough regions to go around 1049 double min; 1050 if (count > total) { 1051 min = ((count - total) * mean) + ((1 - mean) * total); 1052 } else { 1053 // Some will have 1 more than everything else. 1054 int numHigh = (int) (total - (Math.floor(mean) * count)); 1055 int numLow = (int) (count - numHigh); 1056 1057 min = (numHigh * (Math.ceil(mean) - mean)) + (numLow * (mean - Math.floor(mean))); 1058 1059 } 1060 min = Math.max(0, min); 1061 for (int i=0; i<stats.length; i++) { 1062 double n = stats[i]; 1063 double diff = Math.abs(mean - n); 1064 totalCost += diff; 1065 } 1066 1067 double scaled = scale(min, max, totalCost); 1068 return scaled; 1069 } 1070 1071 private double getSum(double[] stats) { 1072 double total = 0; 1073 for(double s:stats) { 1074 total += s; 1075 } 1076 return total; 1077 } 1078 1079 /** 1080 * Scale the value between 0 and 1. 1081 * 1082 * @param min Min value 1083 * @param max The Max value 1084 * @param value The value to be scaled. 1085 * @return The scaled value. 1086 */ 1087 protected double scale(double min, double max, double value) { 1088 if (max <= min || value <= min) { 1089 return 0; 1090 } 1091 if ((max - min) == 0) return 0; 1092 1093 return Math.max(0d, Math.min(1d, (value - min) / (max - min))); 1094 } 1095 } 1096 1097 /** 1098 * Given the starting state of the regions and a potential ending state 1099 * compute cost based upon the number of regions that have moved. 1100 */ 1101 static class MoveCostFunction extends CostFunction { 1102 private static final String MOVE_COST_KEY = "hbase.master.balancer.stochastic.moveCost"; 1103 private static final String MAX_MOVES_PERCENT_KEY = 1104 "hbase.master.balancer.stochastic.maxMovePercent"; 1105 private static final float DEFAULT_MOVE_COST = 7; 1106 private static final int DEFAULT_MAX_MOVES = 600; 1107 private static final float DEFAULT_MAX_MOVE_PERCENT = 0.25f; 1108 1109 private final float maxMovesPercent; 1110 1111 MoveCostFunction(Configuration conf) { 1112 super(conf); 1113 1114 // Move cost multiplier should be the same cost or higher than the rest of the costs to ensure 1115 // that large benefits are need to overcome the cost of a move. 1116 this.setMultiplier(conf.getFloat(MOVE_COST_KEY, DEFAULT_MOVE_COST)); 1117 // What percent of the number of regions a single run of the balancer can move. 1118 maxMovesPercent = conf.getFloat(MAX_MOVES_PERCENT_KEY, DEFAULT_MAX_MOVE_PERCENT); 1119 } 1120 1121 @Override 1122 double cost() { 1123 // Try and size the max number of Moves, but always be prepared to move some. 1124 int maxMoves = Math.max((int) (cluster.numRegions * maxMovesPercent), 1125 DEFAULT_MAX_MOVES); 1126 1127 double moveCost = cluster.numMovedRegions; 1128 1129 // Don't let this single balance move more than the max moves. 1130 // This allows better scaling to accurately represent the actual cost of a move. 1131 if (moveCost > maxMoves) { 1132 return 1000000; // return a number much greater than any of the other cost 1133 } 1134 1135 return scale(0, Math.min(cluster.numRegions, maxMoves), moveCost); 1136 } 1137 } 1138 1139 /** 1140 * Compute the cost of a potential cluster state from skew in number of 1141 * regions on a cluster. 1142 */ 1143 static class RegionCountSkewCostFunction extends CostFunction { 1144 private static final String REGION_COUNT_SKEW_COST_KEY = 1145 "hbase.master.balancer.stochastic.regionCountCost"; 1146 private static final float DEFAULT_REGION_COUNT_SKEW_COST = 500; 1147 1148 private double[] stats = null; 1149 1150 RegionCountSkewCostFunction(Configuration conf) { 1151 super(conf); 1152 // Load multiplier should be the greatest as it is the most general way to balance data. 1153 this.setMultiplier(conf.getFloat(REGION_COUNT_SKEW_COST_KEY, DEFAULT_REGION_COUNT_SKEW_COST)); 1154 } 1155 1156 @Override 1157 double cost() { 1158 if (stats == null || stats.length != cluster.numServers) { 1159 stats = new double[cluster.numServers]; 1160 } 1161 1162 for (int i =0; i < cluster.numServers; i++) { 1163 stats[i] = cluster.regionsPerServer[i].length; 1164 } 1165 1166 return costFromArray(stats); 1167 } 1168 } 1169 1170 /** 1171 * Compute the cost of a potential cluster state from skew in number of 1172 * primary regions on a cluster. 1173 */ 1174 static class PrimaryRegionCountSkewCostFunction extends CostFunction { 1175 private static final String PRIMARY_REGION_COUNT_SKEW_COST_KEY = 1176 "hbase.master.balancer.stochastic.primaryRegionCountCost"; 1177 private static final float DEFAULT_PRIMARY_REGION_COUNT_SKEW_COST = 500; 1178 1179 private double[] stats = null; 1180 1181 PrimaryRegionCountSkewCostFunction(Configuration conf) { 1182 super(conf); 1183 // Load multiplier should be the greatest as primary regions serve majority of reads/writes. 1184 this.setMultiplier(conf.getFloat(PRIMARY_REGION_COUNT_SKEW_COST_KEY, 1185 DEFAULT_PRIMARY_REGION_COUNT_SKEW_COST)); 1186 } 1187 1188 @Override 1189 double cost() { 1190 if (!cluster.hasRegionReplicas) { 1191 return 0; 1192 } 1193 if (stats == null || stats.length != cluster.numServers) { 1194 stats = new double[cluster.numServers]; 1195 } 1196 1197 for (int i = 0; i < cluster.numServers; i++) { 1198 stats[i] = 0; 1199 for (int regionIdx : cluster.regionsPerServer[i]) { 1200 if (regionIdx == cluster.regionIndexToPrimaryIndex[regionIdx]) { 1201 stats[i]++; 1202 } 1203 } 1204 } 1205 1206 return costFromArray(stats); 1207 } 1208 } 1209 1210 /** 1211 * Compute the cost of a potential cluster configuration based upon how evenly 1212 * distributed tables are. 1213 */ 1214 static class TableSkewCostFunction extends CostFunction { 1215 1216 private static final String TABLE_SKEW_COST_KEY = 1217 "hbase.master.balancer.stochastic.tableSkewCost"; 1218 private static final float DEFAULT_TABLE_SKEW_COST = 35; 1219 1220 TableSkewCostFunction(Configuration conf) { 1221 super(conf); 1222 this.setMultiplier(conf.getFloat(TABLE_SKEW_COST_KEY, DEFAULT_TABLE_SKEW_COST)); 1223 } 1224 1225 @Override 1226 double cost() { 1227 double max = cluster.numRegions; 1228 double min = ((double) cluster.numRegions) / cluster.numServers; 1229 double value = 0; 1230 1231 for (int i = 0; i < cluster.numMaxRegionsPerTable.length; i++) { 1232 value += cluster.numMaxRegionsPerTable[i]; 1233 } 1234 1235 return scale(min, max, value); 1236 } 1237 } 1238 1239 /** 1240 * Compute a cost of a potential cluster configuration based upon where 1241 * {@link org.apache.hadoop.hbase.regionserver.HStoreFile}s are located. 1242 */ 1243 static abstract class LocalityBasedCostFunction extends CostFunction { 1244 1245 private final LocalityType type; 1246 1247 private double bestLocality; // best case locality across cluster weighted by local data size 1248 private double locality; // current locality across cluster weighted by local data size 1249 1250 private MasterServices services; 1251 1252 LocalityBasedCostFunction(Configuration conf, 1253 MasterServices srv, 1254 LocalityType type, 1255 String localityCostKey, 1256 float defaultLocalityCost) { 1257 super(conf); 1258 this.type = type; 1259 this.setMultiplier(conf.getFloat(localityCostKey, defaultLocalityCost)); 1260 this.services = srv; 1261 this.locality = 0.0; 1262 this.bestLocality = 0.0; 1263 } 1264 1265 /** 1266 * Maps region to the current entity (server or rack) on which it is stored 1267 */ 1268 abstract int regionIndexToEntityIndex(int region); 1269 1270 public void setServices(MasterServices srvc) { 1271 this.services = srvc; 1272 } 1273 1274 @Override 1275 void init(Cluster cluster) { 1276 super.init(cluster); 1277 locality = 0.0; 1278 bestLocality = 0.0; 1279 1280 // If no master, no computation will work, so assume 0 cost 1281 if (this.services == null) { 1282 return; 1283 } 1284 1285 for (int region = 0; region < cluster.numRegions; region++) { 1286 locality += getWeightedLocality(region, regionIndexToEntityIndex(region)); 1287 bestLocality += getWeightedLocality(region, getMostLocalEntityForRegion(region)); 1288 } 1289 1290 // We normalize locality to be a score between 0 and 1.0 representing how good it 1291 // is compared to how good it could be. If bestLocality is 0, assume locality is 100 1292 // (and the cost is 0) 1293 locality = bestLocality == 0 ? 1.0 : locality / bestLocality; 1294 } 1295 1296 @Override 1297 protected void regionMoved(int region, int oldServer, int newServer) { 1298 int oldEntity = type == LocalityType.SERVER ? oldServer : cluster.serverIndexToRackIndex[oldServer]; 1299 int newEntity = type == LocalityType.SERVER ? newServer : cluster.serverIndexToRackIndex[newServer]; 1300 if (this.services == null) { 1301 return; 1302 } 1303 double localityDelta = getWeightedLocality(region, newEntity) - getWeightedLocality(region, oldEntity); 1304 double normalizedDelta = bestLocality == 0 ? 0.0 : localityDelta / bestLocality; 1305 locality += normalizedDelta; 1306 } 1307 1308 @Override 1309 double cost() { 1310 return 1 - locality; 1311 } 1312 1313 private int getMostLocalEntityForRegion(int region) { 1314 return cluster.getOrComputeRegionsToMostLocalEntities(type)[region]; 1315 } 1316 1317 private double getWeightedLocality(int region, int entity) { 1318 return cluster.getOrComputeWeightedLocality(region, entity, type); 1319 } 1320 1321 } 1322 1323 static class ServerLocalityCostFunction extends LocalityBasedCostFunction { 1324 1325 private static final String LOCALITY_COST_KEY = "hbase.master.balancer.stochastic.localityCost"; 1326 private static final float DEFAULT_LOCALITY_COST = 25; 1327 1328 ServerLocalityCostFunction(Configuration conf, MasterServices srv) { 1329 super( 1330 conf, 1331 srv, 1332 LocalityType.SERVER, 1333 LOCALITY_COST_KEY, 1334 DEFAULT_LOCALITY_COST 1335 ); 1336 } 1337 1338 @Override 1339 int regionIndexToEntityIndex(int region) { 1340 return cluster.regionIndexToServerIndex[region]; 1341 } 1342 } 1343 1344 static class RackLocalityCostFunction extends LocalityBasedCostFunction { 1345 1346 private static final String RACK_LOCALITY_COST_KEY = "hbase.master.balancer.stochastic.rackLocalityCost"; 1347 private static final float DEFAULT_RACK_LOCALITY_COST = 15; 1348 1349 public RackLocalityCostFunction(Configuration conf, MasterServices services) { 1350 super( 1351 conf, 1352 services, 1353 LocalityType.RACK, 1354 RACK_LOCALITY_COST_KEY, 1355 DEFAULT_RACK_LOCALITY_COST 1356 ); 1357 } 1358 1359 @Override 1360 int regionIndexToEntityIndex(int region) { 1361 return cluster.getRackForRegion(region); 1362 } 1363 } 1364 1365 /** 1366 * Base class the allows writing costs functions from rolling average of some 1367 * number from RegionLoad. 1368 */ 1369 abstract static class CostFromRegionLoadFunction extends CostFunction { 1370 1371 private ClusterMetrics clusterStatus = null; 1372 private Map<String, Deque<BalancerRegionLoad>> loads = null; 1373 private double[] stats = null; 1374 CostFromRegionLoadFunction(Configuration conf) { 1375 super(conf); 1376 } 1377 1378 void setClusterMetrics(ClusterMetrics status) { 1379 this.clusterStatus = status; 1380 } 1381 1382 void setLoads(Map<String, Deque<BalancerRegionLoad>> l) { 1383 this.loads = l; 1384 } 1385 1386 @Override 1387 double cost() { 1388 if (clusterStatus == null || loads == null) { 1389 return 0; 1390 } 1391 1392 if (stats == null || stats.length != cluster.numServers) { 1393 stats = new double[cluster.numServers]; 1394 } 1395 1396 for (int i =0; i < stats.length; i++) { 1397 //Cost this server has from RegionLoad 1398 long cost = 0; 1399 1400 // for every region on this server get the rl 1401 for(int regionIndex:cluster.regionsPerServer[i]) { 1402 Collection<BalancerRegionLoad> regionLoadList = cluster.regionLoads[regionIndex]; 1403 1404 // Now if we found a region load get the type of cost that was requested. 1405 if (regionLoadList != null) { 1406 cost = (long) (cost + getRegionLoadCost(regionLoadList)); 1407 } 1408 } 1409 1410 // Add the total cost to the stats. 1411 stats[i] = cost; 1412 } 1413 1414 // Now return the scaled cost from data held in the stats object. 1415 return costFromArray(stats); 1416 } 1417 1418 protected double getRegionLoadCost(Collection<BalancerRegionLoad> regionLoadList) { 1419 double cost = 0; 1420 for (BalancerRegionLoad rl : regionLoadList) { 1421 cost += getCostFromRl(rl); 1422 } 1423 return cost / regionLoadList.size(); 1424 } 1425 1426 protected abstract double getCostFromRl(BalancerRegionLoad rl); 1427 } 1428 1429 /** 1430 * Class to be used for the subset of RegionLoad costs that should be treated as rates. 1431 * We do not compare about the actual rate in requests per second but rather the rate relative 1432 * to the rest of the regions. 1433 */ 1434 abstract static class CostFromRegionLoadAsRateFunction extends CostFromRegionLoadFunction { 1435 1436 CostFromRegionLoadAsRateFunction(Configuration conf) { 1437 super(conf); 1438 } 1439 1440 @Override 1441 protected double getRegionLoadCost(Collection<BalancerRegionLoad> regionLoadList) { 1442 double cost = 0; 1443 double previous = 0; 1444 boolean isFirst = true; 1445 for (BalancerRegionLoad rl : regionLoadList) { 1446 double current = getCostFromRl(rl); 1447 if (isFirst) { 1448 isFirst = false; 1449 } else { 1450 cost += current - previous; 1451 } 1452 previous = current; 1453 } 1454 return Math.max(0, cost / (regionLoadList.size() - 1)); 1455 } 1456 } 1457 1458 /** 1459 * Compute the cost of total number of read requests The more unbalanced the higher the 1460 * computed cost will be. This uses a rolling average of regionload. 1461 */ 1462 1463 static class ReadRequestCostFunction extends CostFromRegionLoadAsRateFunction { 1464 1465 private static final String READ_REQUEST_COST_KEY = 1466 "hbase.master.balancer.stochastic.readRequestCost"; 1467 private static final float DEFAULT_READ_REQUEST_COST = 5; 1468 1469 ReadRequestCostFunction(Configuration conf) { 1470 super(conf); 1471 this.setMultiplier(conf.getFloat(READ_REQUEST_COST_KEY, DEFAULT_READ_REQUEST_COST)); 1472 } 1473 1474 @Override 1475 protected double getCostFromRl(BalancerRegionLoad rl) { 1476 return rl.getReadRequestsCount(); 1477 } 1478 } 1479 1480 /** 1481 * Compute the cost of total number of write requests. The more unbalanced the higher the 1482 * computed cost will be. This uses a rolling average of regionload. 1483 */ 1484 static class WriteRequestCostFunction extends CostFromRegionLoadAsRateFunction { 1485 1486 private static final String WRITE_REQUEST_COST_KEY = 1487 "hbase.master.balancer.stochastic.writeRequestCost"; 1488 private static final float DEFAULT_WRITE_REQUEST_COST = 5; 1489 1490 WriteRequestCostFunction(Configuration conf) { 1491 super(conf); 1492 this.setMultiplier(conf.getFloat(WRITE_REQUEST_COST_KEY, DEFAULT_WRITE_REQUEST_COST)); 1493 } 1494 1495 @Override 1496 protected double getCostFromRl(BalancerRegionLoad rl) { 1497 return rl.getWriteRequestsCount(); 1498 } 1499 } 1500 1501 /** 1502 * A cost function for region replicas. We give a very high cost to hosting 1503 * replicas of the same region in the same host. We do not prevent the case 1504 * though, since if numReplicas > numRegionServers, we still want to keep the 1505 * replica open. 1506 */ 1507 static class RegionReplicaHostCostFunction extends CostFunction { 1508 private static final String REGION_REPLICA_HOST_COST_KEY = 1509 "hbase.master.balancer.stochastic.regionReplicaHostCostKey"; 1510 private static final float DEFAULT_REGION_REPLICA_HOST_COST_KEY = 100000; 1511 1512 long maxCost = 0; 1513 long[] costsPerGroup; // group is either server, host or rack 1514 int[][] primariesOfRegionsPerGroup; 1515 1516 public RegionReplicaHostCostFunction(Configuration conf) { 1517 super(conf); 1518 this.setMultiplier(conf.getFloat(REGION_REPLICA_HOST_COST_KEY, 1519 DEFAULT_REGION_REPLICA_HOST_COST_KEY)); 1520 } 1521 1522 @Override 1523 void init(Cluster cluster) { 1524 super.init(cluster); 1525 // max cost is the case where every region replica is hosted together regardless of host 1526 maxCost = cluster.numHosts > 1 ? getMaxCost(cluster) : 0; 1527 costsPerGroup = new long[cluster.numHosts]; 1528 primariesOfRegionsPerGroup = cluster.multiServersPerHost // either server based or host based 1529 ? cluster.primariesOfRegionsPerHost 1530 : cluster.primariesOfRegionsPerServer; 1531 for (int i = 0 ; i < primariesOfRegionsPerGroup.length; i++) { 1532 costsPerGroup[i] = costPerGroup(primariesOfRegionsPerGroup[i]); 1533 } 1534 } 1535 1536 long getMaxCost(Cluster cluster) { 1537 if (!cluster.hasRegionReplicas) { 1538 return 0; // short circuit 1539 } 1540 // max cost is the case where every region replica is hosted together regardless of host 1541 int[] primariesOfRegions = new int[cluster.numRegions]; 1542 System.arraycopy(cluster.regionIndexToPrimaryIndex, 0, primariesOfRegions, 0, 1543 cluster.regions.length); 1544 1545 Arrays.sort(primariesOfRegions); 1546 1547 // compute numReplicas from the sorted array 1548 return costPerGroup(primariesOfRegions); 1549 } 1550 1551 @Override 1552 boolean isNeeded() { 1553 return cluster.hasRegionReplicas; 1554 } 1555 1556 @Override 1557 double cost() { 1558 if (maxCost <= 0) { 1559 return 0; 1560 } 1561 1562 long totalCost = 0; 1563 for (int i = 0 ; i < costsPerGroup.length; i++) { 1564 totalCost += costsPerGroup[i]; 1565 } 1566 return scale(0, maxCost, totalCost); 1567 } 1568 1569 /** 1570 * For each primary region, it computes the total number of replicas in the array (numReplicas) 1571 * and returns a sum of numReplicas-1 squared. For example, if the server hosts 1572 * regions a, b, c, d, e, f where a and b are same replicas, and c,d,e are same replicas, it 1573 * returns (2-1) * (2-1) + (3-1) * (3-1) + (1-1) * (1-1). 1574 * @param primariesOfRegions a sorted array of primary regions ids for the regions hosted 1575 * @return a sum of numReplicas-1 squared for each primary region in the group. 1576 */ 1577 protected long costPerGroup(int[] primariesOfRegions) { 1578 long cost = 0; 1579 int currentPrimary = -1; 1580 int currentPrimaryIndex = -1; 1581 // primariesOfRegions is a sorted array of primary ids of regions. Replicas of regions 1582 // sharing the same primary will have consecutive numbers in the array. 1583 for (int j = 0 ; j <= primariesOfRegions.length; j++) { 1584 int primary = j < primariesOfRegions.length ? primariesOfRegions[j] : -1; 1585 if (primary != currentPrimary) { // we see a new primary 1586 int numReplicas = j - currentPrimaryIndex; 1587 // square the cost 1588 if (numReplicas > 1) { // means consecutive primaries, indicating co-location 1589 cost += (numReplicas - 1) * (numReplicas - 1); 1590 } 1591 currentPrimary = primary; 1592 currentPrimaryIndex = j; 1593 } 1594 } 1595 1596 return cost; 1597 } 1598 1599 @Override 1600 protected void regionMoved(int region, int oldServer, int newServer) { 1601 if (maxCost <= 0) { 1602 return; // no need to compute 1603 } 1604 if (cluster.multiServersPerHost) { 1605 int oldHost = cluster.serverIndexToHostIndex[oldServer]; 1606 int newHost = cluster.serverIndexToHostIndex[newServer]; 1607 if (newHost != oldHost) { 1608 costsPerGroup[oldHost] = costPerGroup(cluster.primariesOfRegionsPerHost[oldHost]); 1609 costsPerGroup[newHost] = costPerGroup(cluster.primariesOfRegionsPerHost[newHost]); 1610 } 1611 } else { 1612 costsPerGroup[oldServer] = costPerGroup(cluster.primariesOfRegionsPerServer[oldServer]); 1613 costsPerGroup[newServer] = costPerGroup(cluster.primariesOfRegionsPerServer[newServer]); 1614 } 1615 } 1616 } 1617 1618 /** 1619 * A cost function for region replicas for the rack distribution. We give a relatively high 1620 * cost to hosting replicas of the same region in the same rack. We do not prevent the case 1621 * though. 1622 */ 1623 static class RegionReplicaRackCostFunction extends RegionReplicaHostCostFunction { 1624 private static final String REGION_REPLICA_RACK_COST_KEY = 1625 "hbase.master.balancer.stochastic.regionReplicaRackCostKey"; 1626 private static final float DEFAULT_REGION_REPLICA_RACK_COST_KEY = 10000; 1627 1628 public RegionReplicaRackCostFunction(Configuration conf) { 1629 super(conf); 1630 this.setMultiplier(conf.getFloat(REGION_REPLICA_RACK_COST_KEY, 1631 DEFAULT_REGION_REPLICA_RACK_COST_KEY)); 1632 } 1633 1634 @Override 1635 void init(Cluster cluster) { 1636 this.cluster = cluster; 1637 if (cluster.numRacks <= 1) { 1638 maxCost = 0; 1639 return; // disabled for 1 rack 1640 } 1641 // max cost is the case where every region replica is hosted together regardless of rack 1642 maxCost = getMaxCost(cluster); 1643 costsPerGroup = new long[cluster.numRacks]; 1644 for (int i = 0 ; i < cluster.primariesOfRegionsPerRack.length; i++) { 1645 costsPerGroup[i] = costPerGroup(cluster.primariesOfRegionsPerRack[i]); 1646 } 1647 } 1648 1649 @Override 1650 protected void regionMoved(int region, int oldServer, int newServer) { 1651 if (maxCost <= 0) { 1652 return; // no need to compute 1653 } 1654 int oldRack = cluster.serverIndexToRackIndex[oldServer]; 1655 int newRack = cluster.serverIndexToRackIndex[newServer]; 1656 if (newRack != oldRack) { 1657 costsPerGroup[oldRack] = costPerGroup(cluster.primariesOfRegionsPerRack[oldRack]); 1658 costsPerGroup[newRack] = costPerGroup(cluster.primariesOfRegionsPerRack[newRack]); 1659 } 1660 } 1661 } 1662 1663 /** 1664 * Compute the cost of total memstore size. The more unbalanced the higher the 1665 * computed cost will be. This uses a rolling average of regionload. 1666 */ 1667 static class MemStoreSizeCostFunction extends CostFromRegionLoadAsRateFunction { 1668 1669 private static final String MEMSTORE_SIZE_COST_KEY = 1670 "hbase.master.balancer.stochastic.memstoreSizeCost"; 1671 private static final float DEFAULT_MEMSTORE_SIZE_COST = 5; 1672 1673 MemStoreSizeCostFunction(Configuration conf) { 1674 super(conf); 1675 this.setMultiplier(conf.getFloat(MEMSTORE_SIZE_COST_KEY, DEFAULT_MEMSTORE_SIZE_COST)); 1676 } 1677 1678 @Override 1679 protected double getCostFromRl(BalancerRegionLoad rl) { 1680 return rl.getMemStoreSizeMB(); 1681 } 1682 } 1683 /** 1684 * Compute the cost of total open storefiles size. The more unbalanced the higher the 1685 * computed cost will be. This uses a rolling average of regionload. 1686 */ 1687 static class StoreFileCostFunction extends CostFromRegionLoadFunction { 1688 1689 private static final String STOREFILE_SIZE_COST_KEY = 1690 "hbase.master.balancer.stochastic.storefileSizeCost"; 1691 private static final float DEFAULT_STOREFILE_SIZE_COST = 5; 1692 1693 StoreFileCostFunction(Configuration conf) { 1694 super(conf); 1695 this.setMultiplier(conf.getFloat(STOREFILE_SIZE_COST_KEY, DEFAULT_STOREFILE_SIZE_COST)); 1696 } 1697 1698 @Override 1699 protected double getCostFromRl(BalancerRegionLoad rl) { 1700 return rl.getStorefileSizeMB(); 1701 } 1702 } 1703 1704 /** 1705 * A helper function to compose the attribute name from tablename and costfunction name 1706 */ 1707 public static String composeAttributeName(String tableName, String costFunctionName) { 1708 return tableName + TABLE_FUNCTION_SEP + costFunctionName; 1709 } 1710}