2 ms·
That's partly what the article is about. In a pure mark/sweep GC, as the heap size goes up, the work needed to complete a GC cycle and free some memory goes up
by mike_hearn 10y ago
That's partly what the article is about.
In a pure mark/sweep GC, as the heap size goes up, the work needed to complete a GC cycle and free some memory goes up too. Thus the more memory you use the more cores you need.
In more complicated collectors that isn't necessarily the case: collection is incremental and done only as much as needed to keep up with the creation of garbage.
As I say in the article, there are reports of people running JVM/G1 on terabyte sized heaps. The problem they have is that if they experience a concurrent mode failure (which G1 is not designed to optimise because in a properly tuned system it's meant to never happen) you're looking at, like, a 5-10 minute pause time whilst the entire heap is mark/sweep and cleaned out as much as possible. So with such huge heaps you'd better make sure the collector doesn't fall behind!