Implement Julia GC Heuristics for v1.9.2

[qinsoon]

No description provided.

[udesou]

LGTM.

[udesou]

Could we use the count_live_bytes_in_gc feature and have a fairer comparison with stock Julia? Not that we want to do that necessarily, just wondering if it's as simple as that. I'm also wondering how much their heuristics are dependent on the live bytes instead of live pages.

[qinsoon]

Basically the heuristics use these stats: live bytes/pages, allocated bytes/pages, and freed bytes/pages. We can replace live pages with live bytes in the implementation, but we do not have allocated bytes and freed bytes. I don't think it is a good idea to use live bytes with other page-granularity stats.

[qinsoon]

Also I listed the difference between this implementation and the stock Julia's heuristics here:

mmtk-julia/mmtk/src/gc_trigger.rs

Lines 18 to 30 in e5e1db6

	/// This tries to implement Julia-style GC triggering heuristics. However, it is still siginificantly different
	/// from the Julia's GC heuristics.
	/// 1. Julia counts allocation per thread and compares with a per-thread interval, while this impl counts global
	/// allocation and compares that with an estimated global interval. For Julia, the first thread that allocates
	/// the amount of bytes that exceeds the interval will trigger a GC. For us, as MMTk does not count allocation
	/// per thread, we calculate an estiamted global interval (using thread interval * n_mutator / 2), and compare
	/// global allocation with it.
	/// 2. Julia makes the decision of full heap GC after marking and before sweeping (they call it full sweep), while
	/// MMTk makes such decisions before a GC. So for us, we use Julia's decision, but force a full heap GC in the next
	/// next GC (not the current one).
	/// 3. Julia counts the pointers in the remembered set, and if the remembered set is too large (large_frontier),
	/// they will do full heap GC. MMTk does not collect such information about remembered set, so we do not have the heuristic
	/// based on the remembered set.

. They are different in many ways.