Stream: t-compiler/wg-parallel-rustc

Topic: sync 2019.05.24

lwshang (May 24 2019 at 13:45, on Zulip):

Hi all, are we going to have a meeting in 15 minutes?

mw (May 24 2019 at 14:28, on Zulip):

No, there was no meeting scheduled for today (as far as I know)

lwshang (May 24 2019 at 14:39, on Zulip):

That's fine. Shall we make it an async follow-up on current progress? @Zoxc @simulacrum Would you like to give some key points from the data you gathered during past week? Are they positive or negative in general?

Beyond gathering data, I believe we can always discuss about how to accomplish the experiment in following phases. From my recent experience of trying parallel-rustc, the only convincing way to have two kinds of rustc and compare their performance is to have two copy of rust repo locally and compiled them with parallel-compiler = true and parallel-compiler = false respectively. This workflow is too inconvenient to have other users opt-in and do the comparison. Do you have any idea on how to make the experiment user-friendly?

simulacrum (May 24 2019 at 17:12, on Zulip):

I think for a limited set of platforms (e.g., the ones we currently publish -alt builds for) we could enable parallel compiler builds on just those

simulacrum (May 24 2019 at 17:14, on Zulip):

That's linux, macOS, and Windows (64bit).

simulacrum (May 24 2019 at 17:14, on Zulip):

Unfortunately we currently enable LLVM asserts for those builds so they're slower but potentially we could disable those for the time being? cc @Alex Crichton

simulacrum (May 24 2019 at 17:15, on Zulip):

We can also publish one-off try builds with the parallel compiler enabled and suggest installation via rustup-toolchain-install-master

Alex Crichton (May 24 2019 at 17:23, on Zulip):

I'm not really sure what the context of this is, but the *-alt builders are currently the builders with llvm assertions enabled

simulacrum (May 24 2019 at 18:06, on Zulip):

@Alex Crichton ah, I thought that we had some others that ran tests -- I guess we consider bootstrapping "the test" though with -alt

simulacrum (May 24 2019 at 18:06, on Zulip):

we basically want to ask for community feedback on a parallel-enabled compiler and to do so would wants somewhat regular updates of a parallel compiler build

lwshang (May 24 2019 at 18:13, on Zulip):

Our roadmap is to have two experiment phases, opt-in and opt-out. In any of them, the users should be able to switch parallel-rustc on or off. So how can we make it? In detail , during the first phase, what is the default behavior of the rustc? Is it a compiler with parallel-compiler = false? Or parallel-compiler = true & -Zthreads=1? As far as I know, the later one is what we can provide technically. While the first one is what we should have for comparison, so that we can measure the overhead added by parallel-rustc. This is the contradiction that troubles me .

Last update: Mar 30 2020 at 23:55UTC