Did some tests, but was not able to generate the latencies.
So probably it's a complete different issue.
However, I was able to verify that the throughput increases notably if I prevent network raid sync by shutting down one node.
It goes from 35 MB/sec to 54 MB/sec.
Which is still poor.
Test: Robocopy a 20 GB folder with mix of data.
Phy. adapter: HP SmartArray P420i:
naa...074
naa...69a
naa....106
naa...b52
Datastore LUNS:
naa...0070
naa...0062
