Diogo Netto

(EDIT: see scaling plots below)

No. Will merge and update that.

yeah it did, hopefully fixed in the latest commit

Latest commits implement @vchuravy and @chflood's idea of batching big arrays (bugs are mine). Seems to be working fine on my machine with the `many_refs` benchmark recently added to `GCBenchmarks`....

@nanosoldier `runbenchmarks(!"scalar", vs=":master")`

If your work-list is a stack, still DFS. But yes, this could increase memory consumption for the GC queues in the case of very large arrays (and a single threaded...

`Test whether prefetching on the mark-loop (c.f. Cher, 2004) would improve GC times.` Issue is that you need to read the tag whenever you enqueue an object, so you incur...

TODO list should be complete now. Updated link at the top, which contains benchmark results for one of the latest commits. Interestingly, `list.jl` had opposite trends on AMD/Intel machines, though...

> I haven't seen https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_hash/8dae1bb_vs_d6c5cb6/PlantHydraulics.primary.log before Lot of build/test failures went away after merging master. Perhaps nanosoldier it again?

@yuyichao, it seems that the `remset` is flushed at the beginning of every collection and reconstructed as the graph is traversed. Do you know what is the motivation behind this,...