/** Invariant is that all servers have between floor(avg) and ceiling(avg) number of regions. */
public void assertClusterAsBalanced(List<ServerAndLoad> servers) {
int numServers = servers.size();
int numRegions = 0;
int maxRegions = 0;
int minRegions = Integer.MAX_VALUE;
for (ServerAndLoad server : servers) {
int nr = server.getLoad();
if (nr > maxRegions) {
maxRegions = nr;
}
if (nr < minRegions) {
minRegions = nr;
}
numRegions += nr;
}
if (maxRegions - minRegions < 2) {
// less than 2 between max and min, can't balance
return;
}
int min = numRegions / numServers;
int max = numRegions % numServers == 0 ? min : min + 1;
for (ServerAndLoad server : servers) {
assertTrue(server.getLoad() >= 0);
assertTrue(server.getLoad() <= max);
assertTrue(server.getLoad() >= min);
}
}
/** Invariant is that all servers have between acceptable range number of regions. */
public boolean assertClusterOverallAsBalanced(List<ServerAndLoad> servers, int tablenum) {
int numServers = servers.size();
int numRegions = 0;
int maxRegions = 0;
int minRegions = Integer.MAX_VALUE;
for (ServerAndLoad server : servers) {
int nr = server.getLoad();
if (nr > maxRegions) {
maxRegions = nr;
}
if (nr < minRegions) {
minRegions = nr;
}
numRegions += nr;
}
if (maxRegions - minRegions < 2) {
// less than 2 between max and min, can't balance
return true;
}
int min = numRegions / numServers;
int max = numRegions % numServers == 0 ? min : min + 1;
for (ServerAndLoad server : servers) {
if (server.getLoad() < 0
|| server.getLoad() > max + tablenum / 2 + 1
|| server.getLoad() < min - tablenum / 2 - 1) return false;
}
return true;
}
protected void updateLoad(
final Map<ServerName, ServerAndLoad> map, final ServerName sn, final int diff) {
ServerAndLoad sal = map.get(sn);
if (sal == null) sal = new ServerAndLoad(sn, 0);
sal = new ServerAndLoad(sn, sal.getLoad() + diff);
map.put(sn, sal);
}
protected String printStats(List<ServerAndLoad> servers) {
int numServers = servers.size();
int totalRegions = 0;
for (ServerAndLoad server : servers) {
totalRegions += server.getLoad();
}
float average = (float) totalRegions / numServers;
int max = (int) Math.ceil(average);
int min = (int) Math.floor(average);
return "[srvr="
+ numServers
+ " rgns="
+ totalRegions
+ " avg="
+ average
+ " max="
+ max
+ " min="
+ min
+ "]";
}
/**
* Generate a global load balancing plan according to the specified map of server information to
* the most loaded regions of each server.
*
* <p>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.
*
* <p>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.
*
* <p>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.
*
* <p>Finally, we randomly shuffle underloaded servers so that they receive offloaded regions
* relatively evenly across calls to balanceCluster().
*
* <p>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.
* <p>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>.
* <p>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.
* <p>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.
* <p>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 clusterMap 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
*/
@Override
public List<RegionPlan> balanceCluster(Map<ServerName, List<HRegionInfo>> clusterMap) {
List<RegionPlan> regionsToReturn = balanceMasterRegions(clusterMap);
if (regionsToReturn != null) {
return regionsToReturn;
}
filterExcludedServers(clusterMap);
boolean emptyRegionServerPresent = false;
long startTime = System.currentTimeMillis();
Collection<ServerName> backupMasters = getBackupMasters();
ClusterLoadState cs =
new ClusterLoadState(masterServerName, backupMasters, backupMasterWeight, clusterMap);
// construct a Cluster object with clusterMap and rest of the
// argument as defaults
Cluster c =
new Cluster(
masterServerName,
clusterMap,
null,
this.regionFinder,
getBackupMasters(),
tablesOnMaster,
this.rackManager);
if (!this.needsBalance(c)) return null;
int numServers = cs.getNumServers();
NavigableMap<ServerAndLoad, List<HRegionInfo>> serversByLoad = cs.getServersByLoad();
int numRegions = cs.getNumRegions();
float average = cs.getLoadAverage();
int max = (int) Math.ceil(average);
int min = (int) average;
// Using to check balance result.
StringBuilder strBalanceParam = new StringBuilder();
strBalanceParam
.append("Balance parameter: numRegions=")
.append(numRegions)
.append(", numServers=")
.append(numServers)
.append(", numBackupMasters=")
.append(cs.getNumBackupMasters())
.append(", backupMasterWeight=")
.append(backupMasterWeight)
.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();
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 load = sal.getLoad();
if (load <= max) {
serverBalanceInfo.put(sal.getServerName(), new BalanceInfo(0, 0));
break;
}
serversOverloaded++;
List<HRegionInfo> regions = server.getValue();
int w = 1; // Normal region server has weight 1
if (backupMasters != null && backupMasters.contains(sal.getServerName())) {
w = backupMasterWeight; // Backup master has heavier weight
}
int numToOffload = Math.min((load - max) / w, 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 special regions.
if (shouldBeOnMaster(hri) && masterServerName.equals(sal.getServerName())) 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>();
int maxToTake = numRegions - min;
for (Map.Entry<ServerAndLoad, List<HRegionInfo>> server : serversByLoad.entrySet()) {
if (maxToTake == 0) break; // no more to take
int load = server.getKey().getLoad();
if (load >= min && load > 0) {
continue; // look for other servers which haven't reached min
}
int w = 1; // Normal region server has weight 1
if (backupMasters != null && backupMasters.contains(server.getKey().getServerName())) {
w = backupMasterWeight; // Backup master has heavier weight
}
int regionsToPut = (min - load) / w;
if (regionsToPut == 0) {
regionsToPut = 1;
}
maxToTake -= regionsToPut;
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();
BalanceInfo balanceInfo = serverBalanceInfo.get(server.getKey().getServerName());
if (balanceInfo != null) {
regionCount += balanceInfo.getNumRegionsAdded();
}
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;
}