If the CMS collector can't 'keep up', meaning it doesn't think it can avoid running out of heap space based on its duty cycle and the current memory state, it will do a full on complete, stop the world GC. Based on the article I'm guessing this is what is kicking in.

They never mentioned using CMS, and CMS isn't the default GC, so I assume they weren't using it.

CMS actually isn't the highest throughput GC, it's there for low-pause systems where there are extra cores available to perform GC in the background.

My point was that none of the GCs completely avoid full on stop everything, sweep and compact all the things GC cycles. They will all do this if they are in a position where they think the heap will be exhausted (specifically when a concurrent mode failure occurs). There really isn't anything else a GC can do in this situation other than throw an OOM exception. Thus any time your pushing the heap to its limits, regardless of what GC you use, you will see complete GC sweeps of the old generation.

If you want stricter failure requirements use GCTimeLimit and GCHeapFreeLimit. By default the JVM with throw an OOM error if it spends 98% of execution time in the GC and frees only 2% of the memory, GCTimeLimit and GCHeapFreeLimit switches allow this these values to be changed. Also only the 'stop the world' portions of a collection apply to the execution time limit, concurrent phases do not. So if you managed to keep the collector in concurrent mode (by cranking up the duty cycle for example) then only the 2 mini-pauses will count even if the concurrent portion were now pegging a CPU core at 100%.