Optimizing GC on EMR cluster
I am running a Spark Job written in Scala on EMR and the stdout of each executor is filled with GC allocation failures.
2016-12-07T23:42:20.614+0000: [GC (Allocation Failure) 2016-12-07T23:42:20.614+0000: [ParNew: 909549K->432K(1022400K), 0.0089234 secs] 2279433K->1370373K(3294336K), 0.0090530 secs] [Times: user=0.11 sys=0.00, real=0.00 secs]
2016-12-07T23:42:21.572+0000: [GC (Allocation Failure) 2016-12-07T23:42:21.572+0000: [ParNew: 909296K->435K(1022400K), 0.0089298 secs] 2279237K->1370376K(3294336K), 0.0091147 secs] [Times: user=0.11 sys=0.01, real=0.00 secs]
2016-12-07T23:42:22.525+0000: [GC (Allocation Failure) 2016-12-07T23:42:22.525+0000: [ParNew: 909299K->485K(1022400K), 0.0080858 secs] 2279240K->1370427K(3294336K), 0.0082357 secs] [Times: user=0.12 sys=0.00, real=0.01 secs]
2016-12-07T23:42:23.474+0000: [GC (Allocation Failure) 2016-12-07T23:42:23.474+0000: [ParNew: 909349K->547K(1022400K), 0.0090641 secs] 2279291K->1370489K(3294336K), 0.0091965 secs] [Times: user=0.12 sys=0.00, real=0.00 secs]
I am reading few TB's of data, (mostly string) so I am worried that the constant GC will slow down processing time.
I would appreciate any pointers on how to understand this message and how to optimize GC so that it consumes minimum CPU time.
Answer
Allocation Failure is the normal and the most common reason for initiating GC cycle.
Logs tell that GC happens once a second and takes about 10ms, that is, 1% time. IMO, there is nothing to optimize here.