Elasticsearch 1.5 cluster going to RED state due to some nodes constantly exiting and rejoining the cluster

We have been noticing our elasticsearch cluster is often going to RED state and queries are timing out .
We have noticed that some nodes have high JVM memory utilization and soon exiting the cluster ( from the logs we see out of memory errors on that node and it looks like the node cannot be reached by other cluster members / Master) .
The node eventually recovers after several hours by itself but this is causing an application outage due to the timeout .
We have set the property
indices.fielddata.cache.size: 40%

as per elasticsearch recommendation which we thought would avoid this problem . However it doesn't seem to work .
We also set
on our machines which have 30 GB RAM .

Any help appreciated .

Do you have Marvel installed? If not I would start there, it'll tell you what is happening with a bit more certainty.

Yes , we have marvel installed , but what should we specifically look for this particular issue .

Heap use and GC, that's what it sounds like.