The different AllocationDeciders are more or less expensive processing wise. They are also more or less likely to return a Decision.NO or Decision.THROTTLE. For large clusters this can result in a 10-15% speedup in recalculating the cluster state (tested with https://github.com/geidies/elasticsearch/blob/optimize_RoutingNodes/src/test/java/org/elasticsearch/cluster/routing/allocation/MassiveClusterRebalanceRoutingTests.java - 3.7 seconds on my test machine compared to 4.2 seconds without the optimization).

ConcurrentRebalanceAllocationDecider - loops over all ShardRoutings, O(n), optimization to O(1) for that in separate PR
DisableAllocationDecider - is O(1)
ClusterRebalanceAllocationDecider - is O( #shards_unassigned ) + O( #shards ) or O( 1 ) + O( #shards ) - optimization for that in separate PR, making it O( 2 )
DiskThresholdDecider - O( 2 )
SnapshotInProgressAllocationDecider - O( 1 )
ReplicaAfterPrimaryActiveAllocationDecider and RebalanceOnlyWhenActiveAllocationDecider - O( #shards ) with current implementation of RoutingNodes.getShardsRoutingFor( MutableShardRouting ) - optimization in separate PR, making it O( #shards in replica set )
ShardsLimitAllocationDecider is O( # shards on node ) + O( 1 ).
AwarenessAllocationDecider is O( # shards in cluster ) * # awareness attributes, making it the most expensive, but least likely to be turned on.
SameShardAllocationDecider is O( # shards on node ) * # nodes on host
ThrottlingAllocationDecider, which is O( #shards_per_node ) + O( #shards_per_node )

In addition to the re-ordering, instead of applying all AllocationDeciders, skip the rest of one return a Decision.NO. This logic is ported from the Decision.Multi class.