Idea about garbage collector

Question

Hi,
I'm working in finance area and there are a lot of talks about latency last years though I never saw system that had guaranteed GC STW spikes < 2 ms
so my question why its not possible to implement smart pointer GC (like in C++) so every reference assignment would increment counter in object
and every reference removing decrement - this would add additional cost but 99 % of objects can be safely collected concurrently - sure it does not deal with circular dependencies but I suppose on some applications it would be just great so it can be used with CMS GC

Answer 1

It's certainly possible to use "smart pointers" to implement reference counting garbage collection. It's just that no one has figured out how to make it performant. Yet. All the counter updates have to be atomic, which hurts performance and scalability on multi-processor boxes. There are schemes to avoid the atomic counter updates, for example for pointer stores to thread stacks, but they start to get really complicated and hard to debug. How much of a performance hit are you willing to take to eliminate the stop-the-world pauses?
As you say, you still need something global to clean up cycles.
<2ms is a fairly stringent target for stop-the-world pauses, but not unthinkable. "Guaranteeing" those pauses is unthinkable, without some limitiations on what you do with your heap. You should work with your Oracle support engineers to tune your collector settings. Or post your GC logs and likely someone can come up with ideas to reduce your pause times.