The node that I’m investigating is indeed a VM, but the storage system should be well fast enough with <1ms latency.
I hear that there is possibly a bug with bloom filter cleanup on large nodes (>14M pieces). This node has 31M pieces for the us satellite and may be affected. The bloom filter filling up for large nodes (and thus becoming ineffective by allowing all pieces to remain) makes sense.
Lets just hang on and more info will come.