/**
* Obtain the UUID of some relatively underutilized data service.
*
* <p>FIXME The LBS should interpret the excludedServiceUUID as the source service UUID and then
* provide a list of those services having an LBS computed service score which is significantly
* lower than the score for this service. Changing this will break some unit tests (for the LBS
* behavior).
*/
private UUID getMoveTarget(
final UUID sourceServiceUUID, final ILoadBalancerService loadBalancerService) {
try {
// request under utilized data service UUIDs (RMI).
final UUID[] uuids =
loadBalancerService.getUnderUtilizedDataServices( //
0, // minCount - no lower bound.
1, // maxCount - no upper bound.
sourceServiceUUID // exclude this data service.
);
if (uuids != null && uuids.length > 0) {
// Found a move target.
return uuids[0];
}
// No move target.
return null;
} catch (TimeoutException t) {
log.warn(t.getMessage());
return null;
} catch (InterruptedException t) {
log.warn(t.getMessage());
return null;
} catch (Throwable t) {
log.error("Could not obtain target service UUIDs: ", t);
return null;
}
}
/**
* Figure out if this data service is considered to be highly utilized, in which case the DS
* should shed some index partitions.
*
* <p>Note: We consult the load balancer service on this since it is able to put the load of this
* service into perspective by also considering the load on the other services in the federation.
*
* @param loadBalancerService The load balancer.
*/
protected boolean shouldMove(final ILoadBalancerService loadBalancerService) {
if (loadBalancerService == null) throw new IllegalArgumentException();
// inquire if this service is highly utilized.
final boolean highlyUtilizedService;
try {
final UUID serviceUUID = resourceManager.getDataServiceUUID();
highlyUtilizedService = loadBalancerService.isHighlyUtilizedDataService(serviceUUID);
} catch (Exception ex) {
log.warn("Could not determine if this data service is highly utilized");
return false;
}
if (!highlyUtilizedService) {
if (log.isInfoEnabled()) log.info("Service is not highly utilized.");
return false;
}
/*
* At this point we know that the LBS considers this host and service to
* be highly utilized (relative to the other hosts and services). If
* there is evidence of resource exhaustion for critical resources (CPU,
* RAM, or DIKS) then we will MOVE index partitions in order to shed
* some load. Otherwise, we will SPLIT hot index partitions in order to
* increase the potential concurrency of the workload for this service.
*
* Note: CPU is the only fungable resource since things will just slow
* down if a host has 100% CPU while it can die if it runs out of DISK
* or RAM (including if it begins to swap heavily).
*
* @todo config options for these triggers.
*/
final ResourceScores resourceScores = resourceManager.getResourceScores();
final boolean shouldMove = //
// heavy CPU utilization.
(resourceScores.percentCPUTime >= resourceManager.movePercentCpuTimeThreshold)
||
// swapping heavily.
(resourceScores.majorPageFaultsPerSec > 20)
||
// running out of disk (data dir).
(resourceScores.dataDirBytesFree < Bytes.gigabyte * 5)
||
// running out of disk (tmp dir).
(resourceScores.dataDirBytesFree < Bytes.gigabyte * .5);
return shouldMove;
// if (shouldMove) {
//
// return chooseMoves(loadBalancerService);
//
// }
// return chooseHotSplits();
}
