/**
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.hadoop.hbase.master.balancer;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collections;
import java.util.HashMap;
import java.util.List;
import java.util.Map;
import java.util.NavigableMap;
import java.util.Random;
import java.util.TreeMap;
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.classification.InterfaceAudience;
import org.apache.hadoop.hbase.HRegionInfo;
import org.apache.hadoop.hbase.ServerName;
import org.apache.hadoop.hbase.master.AssignmentManager;
import org.apache.hadoop.hbase.master.RegionPlan;
import com.google.common.collect.MinMaxPriorityQueue;
/**
* Makes decisions about the placement and movement of Regions across
* RegionServers.
*
* <p>Cluster-wide load balancing will occur only when there are no regions in
* transition and according to a fixed period of a time using {@link #balanceCluster(Map)}.
*
* <p>Inline region placement with {@link #immediateAssignment} can be used when
* the Master needs to handle closed regions that it currently does not have
* a destination set for. This can happen during master failover.
*
* <p>On cluster startup, bulk assignment can be used to determine
* locations for all Regions in a cluster.
*
* <p>This classes produces plans for the {@link AssignmentManager} to execute.
*/
@InterfaceAudience.Private
public class DefaultLoadBalancer extends BaseLoadBalancer {
private static final Log LOG = LogFactory.getLog(DefaultLoadBalancer.class);
private static final Random RANDOM = new Random(System.currentTimeMillis());
private RegionInfoComparator riComparator = new RegionInfoComparator();
private RegionPlan.RegionPlanComparator rpComparator = new RegionPlan.RegionPlanComparator();
/**
* Stores additional per-server information about the regions added/removed
* during the run of the balancing algorithm.
*
* For servers that shed regions, we need to track which regions we have already
* shed. <b>nextRegionForUnload</b> contains the index in the list of regions on
* the server that is the next to be shed.
*/
static class BalanceInfo {
private final int nextRegionForUnload;
private int numRegionsAdded;
public BalanceInfo(int nextRegionForUnload, int numRegionsAdded) {
this.nextRegionForUnload = nextRegionForUnload;
this.numRegionsAdded = numRegionsAdded;
}
int getNextRegionForUnload() {
return nextRegionForUnload;
}
int getNumRegionsAdded() {
return numRegionsAdded;
}
void setNumRegionsAdded(int numAdded) {
this.numRegionsAdded = numAdded;
}
}
/**
* Generate a global load balancing plan according to the specified map of
* server information to the most loaded regions of each server.
*
* The load balancing invariant is that all servers are within 1 region of the
* average number of regions per server. If the average is an integer number,
* all servers will be balanced to the average. Otherwise, all servers will
* have either floor(average) or ceiling(average) regions.
*
* HBASE-3609 Modeled regionsToMove using Guava's MinMaxPriorityQueue so that
* we can fetch from both ends of the queue.
* At the beginning, we check whether there was empty region server
* just discovered by Master. If so, we alternately choose new / old
* regions from head / tail of regionsToMove, respectively. This alternation
* avoids clustering young regions on the newly discovered region server.
* Otherwise, we choose new regions from head of regionsToMove.
*
* Another improvement from HBASE-3609 is that we assign regions from
* regionsToMove to underloaded servers in round-robin fashion.
* Previously one underloaded server would be filled before we move onto
* the next underloaded server, leading to clustering of young regions.
*
* Finally, we randomly shuffle underloaded servers so that they receive
* offloaded regions relatively evenly across calls to balanceCluster().
*
* The algorithm is currently implemented as such:
*
* <ol>
* <li>Determine the two valid numbers of regions each server should have,
* <b>MIN</b>=floor(average) and <b>MAX</b>=ceiling(average).
*
* <li>Iterate down the most loaded servers, shedding regions from each so
* each server hosts exactly <b>MAX</b> regions. Stop once you reach a
* server that already has <= <b>MAX</b> regions.
* <p>
* Order the regions to move from most recent to least.
*
* <li>Iterate down the least loaded servers, assigning regions so each server
* has exactly </b>MIN</b> regions. Stop once you reach a server that
* already has >= <b>MIN</b> regions.
*
* Regions being assigned to underloaded servers are those that were shed
* in the previous step. It is possible that there were not enough
* regions shed to fill each underloaded server to <b>MIN</b>. If so we
* end up with a number of regions required to do so, <b>neededRegions</b>.
*
* It is also possible that we were able to fill each underloaded but ended
* up with regions that were unassigned from overloaded servers but that
* still do not have assignment.
*
* If neither of these conditions hold (no regions needed to fill the
* underloaded servers, no regions leftover from overloaded servers),
* we are done and return. Otherwise we handle these cases below.
*
* <li>If <b>neededRegions</b> is non-zero (still have underloaded servers),
* we iterate the most loaded servers again, shedding a single server from
* each (this brings them from having <b>MAX</b> regions to having
* <b>MIN</b> regions).
*
* <li>We now definitely have more regions that need assignment, either from
* the previous step or from the original shedding from overloaded servers.
* Iterate the least loaded servers filling each to <b>MIN</b>.
*
* <li>If we still have more regions that need assignment, again iterate the
* least loaded servers, this time giving each one (filling them to
* </b>MAX</b>) until we run out.
*
* <li>All servers will now either host <b>MIN</b> or <b>MAX</b> regions.
*
* In addition, any server hosting >= <b>MAX</b> regions is guaranteed
* to end up with <b>MAX</b> regions at the end of the balancing. This
* ensures the minimal number of regions possible are moved.
* </ol>
*
* TODO: We can at-most reassign the number of regions away from a particular
* server to be how many they report as most loaded.
* Should we just keep all assignment in memory? Any objections?
* Does this mean we need HeapSize on HMaster? Or just careful monitor?
* (current thinking is we will hold all assignments in memory)
*
* @param clusterState Map of regionservers and their load/region information to
* a list of their most loaded regions
* @return a list of regions to be moved, including source and destination,
* or null if cluster is already balanced
*/
public List<RegionPlan> balanceCluster(
Map<ServerName, List<HRegionInfo>> clusterMap) {
boolean emptyRegionServerPresent = false;
long startTime = System.currentTimeMillis();
ClusterLoadState cs = new ClusterLoadState(clusterMap);
int numServers = cs.getNumServers();
if (numServers == 0) {
LOG.debug("numServers=0 so skipping load balancing");
return null;
}
NavigableMap<ServerAndLoad, List<HRegionInfo>> serversByLoad = cs.getServersByLoad();
int numRegions = cs.getNumRegions();
if (!this.needsBalance(cs)) {
// Skipped because no server outside (min,max) range
float average = cs.getLoadAverage(); // for logging
LOG.info("Skipping load balancing because balanced cluster; " +
"servers=" + numServers + " " +
"regions=" + numRegions + " average=" + average + " " +
"mostloaded=" + serversByLoad.lastKey().getLoad() +
" leastloaded=" + serversByLoad.firstKey().getLoad());
return null;
}
int min = numRegions / numServers;
int max = numRegions % numServers == 0 ? min : min + 1;
// Using to check balance result.
StringBuilder strBalanceParam = new StringBuilder();
strBalanceParam.append("Balance parameter: numRegions=").append(numRegions)
.append(", numServers=").append(numServers).append(", max=").append(max)
.append(", min=").append(min);
LOG.debug(strBalanceParam.toString());
// Balance the cluster
// TODO: Look at data block locality or a more complex load to do this
MinMaxPriorityQueue<RegionPlan> regionsToMove =
MinMaxPriorityQueue.orderedBy(rpComparator).create();
List<RegionPlan> regionsToReturn = new ArrayList<RegionPlan>();
// Walk down most loaded, pruning each to the max
int serversOverloaded = 0;
// flag used to fetch regions from head and tail of list, alternately
boolean fetchFromTail = false;
Map<ServerName, BalanceInfo> serverBalanceInfo =
new TreeMap<ServerName, BalanceInfo>();
for (Map.Entry<ServerAndLoad, List<HRegionInfo>> server:
serversByLoad.descendingMap().entrySet()) {
ServerAndLoad sal = server.getKey();
int regionCount = sal.getLoad();
if (regionCount <= max) {
serverBalanceInfo.put(sal.getServerName(), new BalanceInfo(0, 0));
break;
}
serversOverloaded++;
List<HRegionInfo> regions = server.getValue();
int numToOffload = Math.min(regionCount - max, regions.size());
// account for the out-of-band regions which were assigned to this server
// after some other region server crashed
Collections.sort(regions, riComparator);
int numTaken = 0;
for (int i = 0; i <= numToOffload; ) {
HRegionInfo hri = regions.get(i); // fetch from head
if (fetchFromTail) {
hri = regions.get(regions.size() - 1 - i);
}
i++;
// Don't rebalance meta regions.
if (hri.isMetaRegion()) continue;
regionsToMove.add(new RegionPlan(hri, sal.getServerName(), null));
numTaken++;
if (numTaken >= numToOffload) break;
// fetch in alternate order if there is new region server
if (emptyRegionServerPresent) {
fetchFromTail = !fetchFromTail;
}
}
serverBalanceInfo.put(sal.getServerName(),
new BalanceInfo(numToOffload, (-1)*numTaken));
}
int totalNumMoved = regionsToMove.size();
// Walk down least loaded, filling each to the min
int neededRegions = 0; // number of regions needed to bring all up to min
fetchFromTail = false;
Map<ServerName, Integer> underloadedServers = new HashMap<ServerName, Integer>();
float average = (float)numRegions / numServers; // for logging
int maxToTake = numRegions - (int)average;
for (Map.Entry<ServerAndLoad, List<HRegionInfo>> server:
serversByLoad.entrySet()) {
if (maxToTake == 0) break; // no more to take
int regionCount = server.getKey().getLoad();
if (regionCount >= min && regionCount > 0) {
continue; // look for other servers which haven't reached min
}
int regionsToPut = min - regionCount;
if (regionsToPut == 0)
{
regionsToPut = 1;
maxToTake--;
}
underloadedServers.put(server.getKey().getServerName(), regionsToPut);
}
// number of servers that get new regions
int serversUnderloaded = underloadedServers.size();
int incr = 1;
List<ServerName> sns =
Arrays.asList(underloadedServers.keySet().toArray(new ServerName[serversUnderloaded]));
Collections.shuffle(sns, RANDOM);
while (regionsToMove.size() > 0) {
int cnt = 0;
int i = incr > 0 ? 0 : underloadedServers.size()-1;
for (; i >= 0 && i < underloadedServers.size(); i += incr) {
if (regionsToMove.isEmpty()) break;
ServerName si = sns.get(i);
int numToTake = underloadedServers.get(si);
if (numToTake == 0) continue;
addRegionPlan(regionsToMove, fetchFromTail, si, regionsToReturn);
if (emptyRegionServerPresent) {
fetchFromTail = !fetchFromTail;
}
underloadedServers.put(si, numToTake-1);
cnt++;
BalanceInfo bi = serverBalanceInfo.get(si);
if (bi == null) {
bi = new BalanceInfo(0, 0);
serverBalanceInfo.put(si, bi);
}
bi.setNumRegionsAdded(bi.getNumRegionsAdded()+1);
}
if (cnt == 0) break;
// iterates underloadedServers in the other direction
incr = -incr;
}
for (Integer i : underloadedServers.values()) {
// If we still want to take some, increment needed
neededRegions += i;
}
// If none needed to fill all to min and none left to drain all to max,
// we are done
if (neededRegions == 0 && regionsToMove.isEmpty()) {
long endTime = System.currentTimeMillis();
LOG.info("Calculated a load balance in " + (endTime-startTime) + "ms. " +
"Moving " + totalNumMoved + " regions off of " +
serversOverloaded + " overloaded servers onto " +
serversUnderloaded + " less loaded servers");
return regionsToReturn;
}
// Need to do a second pass.
// Either more regions to assign out or servers that are still underloaded
// If we need more to fill min, grab one from each most loaded until enough
if (neededRegions != 0) {
// Walk down most loaded, grabbing one from each until we get enough
for (Map.Entry<ServerAndLoad, List<HRegionInfo>> server :
serversByLoad.descendingMap().entrySet()) {
BalanceInfo balanceInfo =
serverBalanceInfo.get(server.getKey().getServerName());
int idx =
balanceInfo == null ? 0 : balanceInfo.getNextRegionForUnload();
if (idx >= server.getValue().size()) break;
HRegionInfo region = server.getValue().get(idx);
if (region.isMetaRegion()) continue; // Don't move meta regions.
regionsToMove.add(new RegionPlan(region, server.getKey().getServerName(), null));
totalNumMoved++;
if (--neededRegions == 0) {
// No more regions needed, done shedding
break;
}
}
}
// Now we have a set of regions that must be all assigned out
// Assign each underloaded up to the min, then if leftovers, assign to max
// Walk down least loaded, assigning to each to fill up to min
for (Map.Entry<ServerAndLoad, List<HRegionInfo>> server :
serversByLoad.entrySet()) {
int regionCount = server.getKey().getLoad();
if (regionCount >= min) break;
BalanceInfo balanceInfo = serverBalanceInfo.get(server.getKey().getServerName());
if(balanceInfo != null) {
regionCount += balanceInfo.getNumRegionsAdded();
}
if(regionCount >= min) {
continue;
}
int numToTake = min - regionCount;
int numTaken = 0;
while(numTaken < numToTake && 0 < regionsToMove.size()) {
addRegionPlan(regionsToMove, fetchFromTail,
server.getKey().getServerName(), regionsToReturn);
numTaken++;
if (emptyRegionServerPresent) {
fetchFromTail = !fetchFromTail;
}
}
}
// If we still have regions to dish out, assign underloaded to max
if (0 < regionsToMove.size()) {
for (Map.Entry<ServerAndLoad, List<HRegionInfo>> server :
serversByLoad.entrySet()) {
int regionCount = server.getKey().getLoad();
if(regionCount >= max) {
break;
}
addRegionPlan(regionsToMove, fetchFromTail,
server.getKey().getServerName(), regionsToReturn);
if (emptyRegionServerPresent) {
fetchFromTail = !fetchFromTail;
}
if (regionsToMove.isEmpty()) {
break;
}
}
}
long endTime = System.currentTimeMillis();
if (!regionsToMove.isEmpty() || neededRegions != 0) {
// Emit data so can diagnose how balancer went astray.
LOG.warn("regionsToMove=" + totalNumMoved +
", numServers=" + numServers + ", serversOverloaded=" + serversOverloaded +
", serversUnderloaded=" + serversUnderloaded);
StringBuilder sb = new StringBuilder();
for (Map.Entry<ServerName, List<HRegionInfo>> e: clusterMap.entrySet()) {
if (sb.length() > 0) sb.append(", ");
sb.append(e.getKey().toString());
sb.append(" ");
sb.append(e.getValue().size());
}
LOG.warn("Input " + sb.toString());
}
// All done!
LOG.info("Done. Calculated a load balance in " + (endTime-startTime) + "ms. " +
"Moving " + totalNumMoved + " regions off of " +
serversOverloaded + " overloaded servers onto " +
serversUnderloaded + " less loaded servers");
return regionsToReturn;
}
/**
* Add a region from the head or tail to the List of regions to return.
*/
private void addRegionPlan(final MinMaxPriorityQueue<RegionPlan> regionsToMove,
final boolean fetchFromTail, final ServerName sn, List<RegionPlan> regionsToReturn) {
RegionPlan rp = null;
if (!fetchFromTail) rp = regionsToMove.remove();
else rp = regionsToMove.removeLast();
rp.setDestination(sn);
regionsToReturn.add(rp);
}
}