I just upgraded the memory of my 3 nodes cluster from 16GB each to 64GB and notice that performance has declined by 30%. Furthermore, at some random points some requests take 100 times more to finish than it usually would.
At first I thought I had to wait until the OS caches Lucene information, but it's been 48hs and performance its still poor. Memory usage seems fine and I have no warning in the logs, but in those times that requests take longer than usual, the cluster seems to be not unresponsive but cluttered.
I'm running ES 5.3.2 on Windows Server 2012R with 64GB on each node (3 of them), 16GB are allocated to ES. Previously I had only 6GB assigned to ES and it was blazing fast.
Did upgrade because I'm planning to run complex aggregations in the future.