Gabriel Scherer

I believe that this is a different performance difference than the one I am reporting here, precisely because there is no major-GC work involved in the slowdown. The slowdown you...

We studied the OCAMLRUNPARAM impact rather carefully with @OlivierNicole on Monday. We believe that the slowdown is not related by OCAMLRUNPARAM, in the following sense: - with the OCaml-default GC...

If marking is the bottleneck, this could be a matter of memory bandwidth?

@ejgallego ran more Coq benchmarks, including a benchmark with my patch applied (the one-liner mentioned above which disables work-accounting for the unmarshaller). The results could suffer from measurement biases, but...

In case it can help, here is my subjective perception of the current state of this PR: I stopped looking at the 4.14 behavior (I understand that it suffers from...

@NickBarnes : @damiendoligez mentioned that you had ideas for an API to declare "ramp up" phases in memory allocation, which may have involved a higher-order wrapper to delineate ramp-up phases....

An obvious question: why do we want to store the information on the side, instead of including the information in the bytecode instructions? I suppose that you don't want to...

Note: I am myself unsatisfied with the strong backseat-driving vibe in this issue. I'm asking critical questions on code written by people that know much more than I do about...

(cc also @xavierleroy, @kayceesrk, @ctk21 and @Engil who know more about this and might want to comment.)

#12579 has been merged so the busy-waiting logic in the runtime has been optimized and tuned, but I believe that some issues remain: 0. There still are various examples of...