running a throughput-oriented GC instead of latency/realtime one is a common error of those people (also, throughput oriented ones are usually the default)