[subinterpreters] Add --experimental-isolated-subinterpreters build option

[vstinner]

BPO	40514
Nosy	@vstinner, @ericsnowcurrently, @corona10
PRs	bpo-40514: Add --with-experimental-isolated-subinterpreters #19926 [3.9] bpo-40514: Remove --with-experimental-isolated-subinterpreters in 3.9 #20228

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2020-10-31.22:59:17.920>
created_at = <Date 2020-05-05.13:13:51.880>
labels = ['expert-subinterpreters', 'build', '3.10']
title = '[subinterpreters] Add --experimental-isolated-subinterpreters build option'
updated_at = <Date 2020-10-31.22:59:17.920>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2020-10-31.22:59:17.920>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2020-10-31.22:59:17.920>
closer = 'vstinner'
components = ['Build', 'Subinterpreters']
creation = <Date 2020-05-05.13:13:51.880>
creator = 'vstinner'
dependencies = []
files = []
hgrepos = []
issue_num = 40514
keywords = ['patch']
message_count = 6.0
messages = ['368141', '368156', '368159', '368172', '369401', '370745']
nosy_count = 3.0
nosy_names = ['vstinner', 'eric.snow', 'corona10']
pr_nums = ['19926', '20228']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue40514'
versions = ['Python 3.10']

[vstinner]

Modifying Python internals to have one GIL per interpreter (bpo-40512) is a large project which requires to modify many small things (again, see bpo-40512).

I propose to add a temporary build --experimental-isolated-subinterpreters build option to configure for developers who want to hack on isolated subinterpreters. The intent is to speedup bpo-40512 development by making some practice compromises, to have more time to properly design the real fixes.

For example, tuple, dict and frame have free lists which are shared by subinterpreters. A practical solution is to simply disable them at build time to avoid the need to have per-interpreter free lists.

Another example is pymalloc which is shared by all subinterpreters and rely on the unique global interpreter lock (GIL) to protect its internal states. A practical solution is to disable it and force the usage of libc malloc() function instead.

Some compromosises cannot be the default since they have a significant negative impact on performances. So I propose to add a temporary build option until all these small issues will be fixed.

[vstinner]

New changeset c5fa364 by Victor Stinner in branch 'master':
bpo-40514: Add --with-experimental-isolated-subinterpreters (GH-19926)
c5fa364

[vstinner]

--with-experimental-isolated-subinterpreters option name is very long on purpose: to advertize that you must not use it, unless you fully understand its purpose :-)

I didn't document the build option in Misc/SpecialBuilds.txt or Doc/whatsnew/3.9.rst on purpose: you should not use this special build :-) Basically, this option is only for Eric Snow and me :-D

[ericsnowcurrently]

It would probably make sense to remove the build option in the 3.9 release. We can leave it in master, but remove it in the 3.9 branch once it has been created.

[vstinner]

New changeset 9512ad7 by Victor Stinner in branch '3.9':
[3.9] bpo-40514: Remove --with-experimental-isolated-subinterpreters in 3.9 (GH-20228)
9512ad7

[vstinner]

I started to *remove* a few"#ifdef EXPERIMENTAL_ISOLATED_SUBINTERPRETERS":

• commit 88ec919: Make list free list per-interpreter
• commit 3744ed2: Make frame free list per-interpreter
• PR 20645 (under review): Make dict free lists per-interpreter

[kumaraditya303]

Reopened for #93185

[erlend-aasland]

Reopened for #93185

Why do you need to reopen it? The title says "add --experimental-isolated-subinterpreters build option". If you are to remove it, it can be done in the new issue, no?

[kumaraditya303]

I reopened it as #93185 uses this issue in title, if that PR will use a new issue, this can be closed.

[erlend-aasland]

Use #93185 for the removal; there is no need for this to be reopened.