load data from parquet, poor performance for big datasets

[aceforeverd]

parquet file is 3.5G, and possibly 35G in memory. load data from parquet into online storage, spark bootstrap the job, starting 30min and fails of OOM.

[aceforeverd]

OOM happens when

spark.master=local[*]

set in taskmanager, where the machine has 22 logic cores.

[vagetablechicken]

local spark should set driver memory, I'll add a doc for it. If OOM is reported by JVM, e.g. GC overhead, heap space ..., that means driver memory is small.
And I top about java process memory, RES is much bigger than jvm memory limit, that may cause phyiscal OOM, may be terminated by OOM killer? (I'm not ensure that)