Adding stock gc trigger and set it as default

[udesou]

This PR implements Julia's Stock GC heuristics for triggering GC and resizing the heap. It also updates mmtk-core to the latest version.

[udesou]

julia-version
JULIA_REPO=udesou/Julia
JULIA_REF=9f8df7b5d1c6760c0488525dea0a7efa0f609f53

[qinsoon]

The computation looks like mem balancer. If this is Julia's current heuristics, I assume Julia is using a variant of mem balancer (e.g. the thrashing detection).

[udesou]

That's right. I think it does, and that's what I wanted to say in the MMTk meeting.
However, looking at TUNING_FACTOR from Julia, it looks much different than what mmtk-core implements? But again, the equation they use seems different...

[qinsoon]

We did not 'tune' for mmtk's mem balancer. They multiply with the TUNING_FACTOR, we divide by the TUNING_FACTOR. But set this aside, it looks like that they are using a variant of mem balancer.

[gbaraldi]

Yes, our heuristics started with stock membalancer and then we changed the quite a bit, notably by allowing it to scale a lot faster when small

[udesou]

FYI, here is some rough data (in seconds) about time spent doing some of the steps for building Julia and running make test-gc and how it compares when using Membalancer, MMTK with Stock GC's trigger and Stock.

                                                 membal       mmtk + stock's trigger         stock
Base.Compiler ────                               199.87                 191.10               193.89

Sysimage built. Summary:
Base ────────                                    52.54                  42.17                41.35                                                   
Stdlibs ─────                                    20.97                  19.58                19.46                                           
Total ───────                                    73.51                  61.75                60.82                                                

Precompilation complete. Summary:
Total ───────                                    97.21                  70.55                68.22
Outputting sysimage file...
Output ──────                                    157.13                 120.83               128.78

Precompiling packages finished.
  106 dependency precompiled in                  281                    154                  138
  
make test-gc                                     1168.33                233.25               68.98 

Not sure why there's a huge discrepancy in the gc tests (maybe because we only support full heap collections?), but that seems much more reasonable than what we get from MemBalancer.