support concurrent immix - #300
Conversation
|
May need to change here to add concurrent Immix: Lines 28 to 30 in a519810 |
|
julia-version |
Further CI runs in this PR will use your Julia branch. |
|
I pushed a change (6a85a86) to use this mmtk-core PR: mmtk/mmtk-core#1454 (I will get it merged soon), also a minor fix for scheduling |
|
nice! I've just pushed a bunch of changes to the julia side which might let it get further (will test once I get stock gc compiling again) |
|
@oscardssmith I looked at some time outputs from build steps to roughly check the performance of concurrent immix vs immix. At a glance, it looks reasonable. Concurrent Immix has some expected slowdown, compared to Immix. Immix ...
Compiling the compiler. This may take several minutes ...
Base.Compiler ──── 312.22207379341125 seconds
...
Base ────────── 35.277991 seconds
FileWatching ── 4.001158 seconds
Libdl ───────── 0.001727 seconds
Artifacts ───── 0.290458 seconds
SHA ─────────── 0.251190 seconds
Sockets ─────── 0.766023 seconds
LinearAlgebra ─ 7.174226 seconds
Random ──────── 1.013626 seconds
Stdlibs total ─ 13.503446 seconds
Sysimage built. Summary:
Base ──────── 35.277991 seconds 72.3185%
Stdlibs ───── 13.503446 seconds 27.6815%
Total ─────── 48.781451 seconds
...
Precompilation complete. Summary:
Total ─────── 6.189272 seconds
Outputting sysimage file...
Output ────── 57.502925 secondsConcurrentImmix ...
Compiling the compiler. This may take several minutes ...
Base.Compiler ──── 322.34418296813965 seconds
...
Base ────────── 37.418032 seconds
FileWatching ── 4.162385 seconds
Libdl ───────── 0.001794 seconds
Artifacts ───── 0.321067 seconds
SHA ─────────── 0.258250 seconds
Sockets ─────── 0.878940 seconds
LinearAlgebra ─ 7.397704 seconds
Random ──────── 1.015542 seconds
Stdlibs total ─ 14.040672 seconds
Sysimage built. Summary:
Base ──────── 37.418032 seconds 72.7147%
Stdlibs ───── 14.040672 seconds 27.2853%
Total ─────── 51.458716 seconds
...
Precompilation complete. Summary:
Total ─────── 6.500720 seconds
Outputting sysimage file...
Output ────── 60.748564 seconds |
|
Looks like consistent ~5% slowdown which is on the high side of reasonable, but still reasonable. |
|
Can you get this rebased and merged? I think it's ready. |
I rebased the PRs (this one and the one in the Julia repo). The tests are failing because of the issue in JuliaLang/julia#61769 (JuliaLang/julia#61769 does not resolve the issue, I will work further on it). |
SweepVMSpecific to process_weak_refs.
Co-authored-by: mmtkgc-bot <mmtkgc.bot@gmail.com>
Currently `MMTK_SIDE_LOG_BIT_BASE_ADDRESS` is defined in the binding side. When we embed this symbol in the JIT code, Julia needs to be able to resolve to the variable. This PR moves it to the Julia side. This PR needs JuliaLang/julia#61769 to work.
In preparation for moving flush() out of mmtk-core's ScanMutatorRoots, have the binding flush at the end of its own per-mutator scan. This is currently a double-flush (safe; second one drains an empty buffer) but sets up the contract change in mmtk-core. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
9a116bc to
5d3bb0a
Compare
task and snapshotting a task in the resume hook.
f5244a7 to
9747ebe
Compare
|
Since the binding code has been moved to Julia, the binding changes in this branch is moved to JuliaLang/julia#61215. |
|
yeah. I need to rebase this pr to include the binding changes |
|
never mind. I see you've handled that. thanks! |
Very WIP. Requires mmtk/mmtk-julia#300 --------- Co-authored-by: Yi Lin <qinsoon@gmail.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Cody Tapscott <topolarity@tapscott.me> Co-authored-by: Cody Tapscott <84105208+topolarity@users.noreply.github.com>
2 participants
Requires JuliaLang/julia#61215 to function