PEP 684 – A Per-Interpreter GIL
- Eric Snow <ericsnowcurrently at gmail.com>
- Discourse thread
- Standards Track
- 08-Mar-2022, 29-Sep-2022, 28-Oct-2022
Table of Contents
- High-Level Summary
- How to Teach This
- Reference Implementation
- Open Issues
- Deferred Functionality
- Rejected Ideas
- Extra Context
Since Python 1.5 (1997), CPython users can run multiple interpreters in the same process. However, interpreters in the same process have always shared a significant amount of global state. This is a source of bugs, with a growing impact as more and more people use the feature. Furthermore, sufficient isolation would facilitate true multi-core parallelism, where interpreters no longer share the GIL. The changes outlined in this proposal will result in that level of interpreter isolation.
At a high level, this proposal changes CPython in the following ways:
- stops sharing the GIL between interpreters, given sufficient isolation
- adds several new interpreter config options for isolation settings
- keeps incompatible extensions from causing problems
The GIL protects concurrent access to most of CPython’s runtime state. So all that GIL-protected global state must move to each interpreter before the GIL can.
(In a handful of cases, other mechanisms can be used to ensure thread-safe sharing instead, such as locks or “immortal” objects.)
CPython Runtime State
Properly isolating interpreters requires that most of CPython’s
runtime state be stored in the
PyInterpreterState struct. Currently,
only a portion of it is; the rest is found either in C global variables
_PyRuntimeState. Most of that will have to be moved.
This directly coincides with an ongoing effort (of many years) to greatly
reduce internal use of global variables and consolidate the runtime
(See Consolidating Runtime Global State below.) That project has
significant merit on its own
and has faced little controversy. So, while a per-interpreter GIL
relies on the completion of that effort, that project should not be
considered a part of this proposal–only a dependency.
Other Isolation Considerations
CPython’s interpreters must be strictly isolated from each other, with few exceptions. To a large extent they already are. Each interpreter has its own copy of all modules, classes, functions, and variables. The CPython C-API docs explain further.
However, aside from what has already been mentioned (e.g. the GIL), there are a couple of ways in which interpreters still share some state.
First of all, some process-global resources (e.g. memory, file descriptors, environment variables) are shared. There are no plans to change this.
Second, some isolation is faulty due to bugs or implementations that did not take multiple interpreters into account. This includes CPython’s runtime and the stdlib, as well as extension modules that rely on global variables. Bugs should be opened in these cases, as some already have been.
Depending on Immortal Objects
PEP 683 introduces immortal objects as a CPython-internal feature. With immortal objects, we can share any otherwise immutable global objects between all interpreters. Consequently, this PEP does not need to address how to deal with the various objects exposed in the public C-API. It also simplifies the question of what to do about the builtin static types. (See Global Objects below.)
Both issues have alternate solutions, but everything is simpler with immortal objects. If PEP 683 is not accepted then this one will be updated with the alternatives. This lets us reduce noise in this proposal.
The fundamental problem we’re solving here is a lack of true multi-core parallelism (for Python code) in the CPython runtime. The GIL is the cause. While it usually isn’t a problem in practice, at the very least it makes Python’s multi-core story murky, which makes the GIL a consistent distraction.
Isolated interpreters are also an effective mechanism to support certain concurrency models. PEP 554 discusses this in more detail.
Most of the effort needed for a per-interpreter GIL has benefits that make those tasks worth doing anyway:
- makes multiple-interpreter behavior more reliable
- has led to fixes for long-standing runtime bugs that otherwise hadn’t been prioritized
- has been exposing (and inspiring fixes for) previously unknown runtime bugs
- has driven cleaner runtime initialization (PEP 432, PEP 587)
- has driven cleaner and more complete runtime finalization
- led to structural layering of the C-API (e.g.
- also see Benefits to Consolidation below
Furthermore, much of that work benefits other CPython-related projects:
- performance improvements (”faster-cpython”)
- pre-fork application deployment (e.g. Instagram server)
- extension module isolation (see PEP 630, etc.)
- embedding CPython
Existing Use of Multiple Interpreters
The C-API for multiple interpreters has been used for many years. However, until relatively recently the feature wasn’t widely known, nor extensively used (with the exception of mod_wsgi).
In the last few years use of multiple interpreters has been increasing. Here are some of the public projects using the feature currently:
Note that, with PEP 554, multiple interpreter usage would likely grow significantly (via Python code rather than the C-API).
PEP 554 (Multiple Interpreters in the Stdlib)
PEP 554 is strictly about providing a minimal stdlib module to give users access to multiple interpreters from Python code. In fact, it specifically avoids proposing any changes related to the GIL. Consider, however, that users of that module would benefit from a per-interpreter GIL, which makes PEP 554 more appealing.
During initial investigations in 2014, a variety of possible solutions for multi-core Python were explored, but each had its drawbacks without simple solutions:
- the existing practice of releasing the GIL in extension modules
- doesn’t help with Python code
- other Python implementations (e.g. Jython, IronPython)
- CPython dominates the community
- remove the GIL (e.g. gilectomy, “no-gil”)
- too much technical risk (at the time)
- Trent Nelson’s “PyParallel” project
- incomplete; Windows-only at the time
- too much work to make it effective enough; high penalties in some situations (at large scale, Windows)
- other parallelism tools (e.g. dask, ray, MPI)
- not a fit for the runtime/stdlib
- give up on multi-core (e.g. async, do nothing)
- this can only end in tears
Even in 2014, it was fairly clear that a solution using isolated interpreters did not have a high level of technical risk and that most of the work was worth doing anyway. (The downside was the volume of work to be done.)
As summarized above, this proposal involves the following changes, in the order they must happen:
- consolidate global runtime state
(including objects) into
- move nearly all of the state down into
- finally, move the GIL down into
- everything else
- update the C-API
- implement extension module restrictions
- work with popular extension maintainers to help with multi-interpreter support
The following runtime state will be moved to
- all global objects that are not safely shareable (fully immutable)
- the GIL
- most mutable data that’s currently protected by the GIL
- mutable data that’s currently protected by some other per-interpreter lock
- mutable data that may be used independently in different interpreters (also applies to extension modules, including those with multi-phase init)
- all other mutable data not otherwise excluded below
Furthermore, a portion of the full global state has already been moved to the interpreter, including GC, warnings, and atexit hooks.
The following runtime state will not be moved:
- global objects that are safely shareable, if any
- immutable data, often
- effectively immutable data (treated as immutable), for example:
- some state is initialized early and never modified again
- hashes for strings (
PyUnicodeObject) are idempotently calculated when first needed and then cached
- all data that is guaranteed to be modified exclusively in the main thread,
- state used only in CPython’s
- the REPL’s state
- data modified only during runtime init (effectively immutable afterward)
- state used only in CPython’s
- mutable data that’s protected by some global lock (other than the GIL)
- global state in atomic variables
- mutable global state that can be changed (sensibly) to atomic variables
This is one of the most sensitive parts of the work to isolate interpreters.
The simplest solution is to move the global state of the internal
“small block” allocator to
PyInterpreterState, as we are doing with
nearly all other runtime state. The following elaborates on the details
CPython provides a memory management C-API, with three allocator domains:
“raw”, “mem”, and “object”. Each provides the equivalent of
free(). A custom allocator for each
domain can be set during runtime initialization and the current allocator
can be wrapped with a hook using the same API (for example, the stdlib
tracemalloc module). The allocators are currently runtime-global,
shared by all interpreters.
The “raw” allocator is expected to be thread-safe and defaults to glibc’s
malloc(), etc.). However, the “mem” and “object” allocators
are not expected to be thread-safe and currently may rely on the GIL for
thread-safety. This is partly because the default allocator for both,
AKA “pyobject”, is not thread-safe. This is due to how all state for
that allocator is stored in C global variables.
Thus we come back to the question of isolating runtime state. In order for interpreters to stop sharing the GIL, allocator thread-safety must be addressed. If interpreters continue sharing the allocators then we need some other way to get thread-safety. Otherwise interpreters must stop sharing the allocators. In both cases there are a number of possible solutions, each with potential downsides.
To keep sharing the allocators, the simplest solution is to use
a granular runtime-global lock around the calls to the “mem” and “object”
PyObject_Malloc(), etc. This would
impact performance, but there are some ways to mitigate that (e.g. only
start locking once the first subinterpreter is created).
Another way to keep sharing the allocators is to require that the “mem” and “object” allocators be thread-safe. This would mean we’d have to make the pyobject allocator implementation thread-safe. That could even involve re-implementing it using an extensible allocator like mimalloc. The potential downside is in the cost to re-implement the allocator and the risk of defects inherent to such an endeavor.
Regardless, a switch to requiring thread-safe allocators would impact anyone that embeds CPython and currently sets a thread-unsafe allocator. We’d need to consider who might be affected and how we reduce any negative impact (e.g. add a basic C-API to help make an allocator thread-safe).
If we did stop sharing the allocators between interpreters, we’d have to do so only for the “mem” and “object” allocators. We might also need to keep a full set of global allocators for certain runtime-level usage. There would be some performance penalty due to looking up the current interpreter and then pointer indirection to get the allocators. Embedders would also likely have to provide a new allocator context for each interpreter. On the plus side, allocator hooks (e.g. tracemalloc) would not be affected.
Ultimately, we will go with the simplest option:
- keep the allocators in the global runtime state
- require that they be thread-safe
- move the state of the default object allocator (AKA “small block”
We experimented with a rough implementation and found it was fairly straightforward, and the performance penalty was essentially zero.
Internally, the interpreter state will now track how the import system should handle extension modules which do not support use with multiple interpreters. See Restricting Extension Modules below. We’ll refer to that setting here as “PyInterpreterState.strict_extension_compat”.
The following API will be made public, if they haven’t been already:
PyThreadState * Py_NewInterpreterFromConfig(PyInterpreterConfig *)
We will add two new fields to
We may add other fields over time, as needed (e.g. “own_initial_thread”).
Regarding the initializer macros,
be used to get an isolated interpreter that also avoids
subinterpreter-unfriendly features. It would be the default for
interpreters created through PEP 554. The unrestricted (status quo)
will continue to be available through
which is already used for the main interpreter and
This will not change.
A note about the “main” interpreter:
Below, we mention the “main” interpreter several times. This refers
to the interpreter created during runtime initialization, for which
PyThreadState corresponds to the process’s main thread.
It is has a number of unique responsibilities (e.g. handling signals),
as well as a special role during runtime initialization/finalization.
It is also usually (for now) the only interpreter.
(Also see https://docs.python.org/3/c-api/init.html#sub-interpreter-support.)
1) then the new interpreter will have its own “global”
interpreter lock. This means the new interpreter can run without
getting interrupted by other interpreters. This effectively unblocks
full use of multiple cores. That is the fundamental goal of this PEP.
0) then the new interpreter will use the main
interpreter’s lock. This is the legacy (pre-3.12) behavior in CPython,
where all interpreters share a single GIL. Sharing the GIL like this
may be desirable when using extension modules that still depend
on the GIL for thread safety.
PyInterpreterConfig_INIT, this will be
PyInterpreterConfig_LEGACY_INIT, this will be
Also, to play it safe, for now we will not allow
own_gil to be true
if a custom allocator was set during runtime init. Wrapping the allocator,
a la tracemalloc, will still be fine.
PyInterpreterConfig.strict_extension_compat is basically the initial
value used for “PyInterpreterState.strict_extension_compat”.
Restricting Extension Modules
Extension modules have many of the same problems as the runtime when state is stored in global variables. PEP 630 covers all the details of what extensions must do to support isolation, and thus safely run in multiple interpreters at once. This includes dealing with their globals.
If an extension implements multi-phase init (see PEP 489) it is considered compatible with multiple interpreters. All other extensions are considered incompatible. (See Extension Module Thread Safety for more details about how a per-interpreter GIL may affect that classification.)
If an incompatible extension is imported and the current
“PyInterpreterState.strict_extension_compat” value is
true then the import
system will raise
false it simply doesn’t check.)
This will be done through
importlib._bootstrap_external.ExtensionFileLoader (really, through
Such imports will never fail in the main interpreter (or in interpreters
“PyInterpreterState.strict_extension_compat” initializes to
false in both
cases. Thus the legacy (pre-3.12) behavior is preserved.
We will work with popular extensions to help them support use in multiple interpreters. This may involve adding to CPython’s public C-API, which we will address on a case-by-case basis.
Extension Module Compatibility
As noted in Extension Modules, many extensions work fine in multiple interpreters (and under a per-interpreter GIL) without needing any changes. The import system will still fail if such a module doesn’t explicitly indicate support. At first, not many extension modules will, so this is a potential source of frustration.
We will address this by adding a context manager to temporarily disable
the check on multiple interpreter support:
importlib.util.allow_all_extensions(). More or less, it will modify
the current “PyInterpreterState.strict_extension_compat” value (e.g. through
Extension Module Thread Safety
If a module supports use with multiple interpreters, that mostly implies it will work even if those interpreters do not share the GIL. The one caveat is where a module links against a library with internal global state that isn’t thread-safe. (Even something as innocuous as a static local variable as a temporary buffer can be a problem.) With a shared GIL, that state is protected. Without one, such modules must wrap any use of that state (e.g. through calls) with a lock.
Currently, it isn’t clear whether or not supports-multiple-interpreters is sufficiently equivalent to supports-per-interpreter-gil, such that we can avoid any special accommodations. This is still a point of meaningful discussion and investigation. The practical distinction between the two (in the Python community, e.g. PyPI) is not yet understood well enough to settle the matter. Likewise, it isn’t clear what we might be able to do to help extension maintainers mitigate the problem (assuming it is one).
In the meantime, we must proceed as though the difference would be large enough to cause problems for enough extension modules out there. The solution we would apply is:
- add a
PyModuleDefslot that indicates an extension can be imported under a per-interpreter GIL (i.e. opt in)
- add that slot as part of the definition of a “compatible” extension, as discussed earlier
The downside is that not a single extension module will be able to take advantage of the per-interpreter GIL without extra effort by the module maintainer, regardless of how minor that effort. This compounds the problem described in Extension Module Compatibility and the same workaround applies. Ideally, we would determine that there isn’t enough difference to matter.
If we do end up requiring an opt-in for imports under a per-interpreter
GIL, and later determine it isn’t necessary, then we can switch the
default at that point, make the old opt-in slot a noop, and add a new
PyModuleDef slot for explicitly opting out. In fact, it makes
sense to add that opt-out slot from the beginning.
- C-API: the “Sub-interpreter support” section of
Doc/c-api/init.rstwill detail the updated API
- C-API: that section will explain about the consequences of a per-interpreter GIL
- importlib: the
ExtensionFileLoaderentry will note import may fail in subinterpreters
- importlib: there will be a new entry about
No behavior or APIs are intended to change due to this proposal, with two exceptions:
- some extensions will fail to import in some subinterpreters (see the next section)
- “mem” and “object” allocators that are currently not thread-safe may now be susceptible to data races when used in combination with multiple interpreters
The existing C-API for managing interpreters will preserve its current behavior, with new behavior exposed through new API. No other API or runtime behavior is meant to change, including compatibility with the stable ABI.
See Objects Exposed in the C-API below for related discussion.
Currently the most common usage of Python, by far, is with the main
interpreter running by itself. This proposal has zero impact on
extension modules in that scenario. Likewise, for better or worse,
there is no change in behavior under multiple interpreters created
using the existing
Keep in mind that some extensions already break when used in multiple interpreters, due to keeping module state in global variables (or due to the internal state of linked libraries). They may crash or, worse, experience inconsistent behavior. That was part of the motivation for PEP 630 and friends, so this is not a new situation nor a consequence of this proposal.
In contrast, when the proposed API is used to create multiple interpreters, with the appropriate settings, the behavior will change for incompatible extensions. In that case, importing such an extension will fail (outside the main interpreter), as explained in Restricting Extension Modules. For extensions that already break in multiple interpreters, this will be an improvement.
Additionally, some extension modules link against libraries with thread-unsafe internal global state. (See Extension Module Thread Safety.) Such modules will have to start wrapping any direct or indirect use of that state in a lock. This is the key difference from other modules that also implement multi-phase init and thus indicate support for multiple interpreters (i.e. isolation).
Now we get to the break in compatibility mentioned above. Some extensions are safe under multiple interpreters (and a per-interpreter GIL), even though they haven’t indicated that. Unfortunately, there is no reliable way for the import system to infer that such an extension is safe, so importing them will still fail. This case is addressed in Extension Module Compatibility above.
Extension Module Maintainers
One related consideration is that a per-interpreter GIL will likely drive increased use of multiple interpreters, particularly if PEP 554 is accepted. Some maintainers of large extension modules have expressed concern about the increased burden they anticipate due to increased use of multiple interpreters.
Specifically, enabling support for multiple interpreters will require substantial work for some extension modules (albeit likely not many). To add that support, the maintainer(s) of such a module (often volunteers) would have to set aside their normal priorities and interests to focus on compatibility (see PEP 630).
Of course, extension maintainers are free to not add support for use in multiple interpreters. However, users will increasingly demand such support, especially if the feature grows in popularity.
Either way, the situation can be stressful for maintainers of such extensions, particularly when they are doing the work in their spare time. The concerns they have expressed are understandable, and we address the partial solution in the Restricting Extension Modules and Extension Module Compatibility sections.
Alternate Python Implementations
Other Python implementation are not required to provide support for multiple interpreters in the same process (though some do already).
There is no known impact to security with this proposal.
On the one hand, this proposal has already motivated a number of improvements that make CPython more maintainable. That is expected to continue. On the other hand, the underlying work has already exposed various pre-existing defects in the runtime that have had to be fixed. That is also expected to continue as multiple interpreters receive more use. Otherwise, there shouldn’t be a significant impact on maintainability, so the net effect should be positive.
The work to consolidate globals has already provided a number of improvements to CPython’s performance, both speeding it up and using less memory, and this should continue. The performance benefits of a per-interpreter GIL specifically have not been explored. At the very least, it is not expected to make CPython slower (as long as interpreters are sufficiently isolated). And, obviously, it enable a variety of multi-core parallelism in Python code.
How to Teach This
Unlike PEP 554, this is an advanced feature meant for a narrow set of users of the C-API. There is no expectation that the specifics of the API nor its direct application will be taught.
That said, if it were taught then it would boil down to the following:
In addition to Py_NewInterpreter(), you can use Py_NewInterpreterFromConfig() to create an interpreter. The config you pass it indicates how you want that interpreter to behave.
Furthermore, the maintainers of any extension modules that create isolated interpreters will likely need to explain the consequences of a per-interpreter GIL to their users. The first thing to explain is what PEP 554 teaches about the concurrency model that isolated interpreters enables. That leads into the point that Python software written using that concurrency model can then take advantage of multi-core parallelism, which is currently prevented by the GIL.
- Are we okay to require “mem” and “object” allcoators to be thread-safe?
- How would a per-interpreter tracemalloc module relate to global allocators?
- Would the faulthandler module be limited to the main interpreter (like the signal module) or would we leak that global state between interpreters (protected by a granular lock)?
- Split out an informational PEP with all the relevant info, based on the “Consolidating Runtime Global State” section?
- How likely is it that a module works under multiple interpreters (isolation) but doesn’t work under a per-interpreter GIL? (See Extension Module Thread Safety.)
- If it is likely enough, what can we do to help extension maintainers mitigate the problem and enjoy use under a per-intepreter GIL?
- What would be a better (scarier-sounding) name
PyInterpreterConfigoption to always run the interpreter in a new thread
PyInterpreterConfigoption to assign a “main” thread to the interpreter and only run in that thread
Consolidating Runtime Global State
As noted in CPython Runtime State above, there is an active effort
(separate from this PEP) to consolidate CPython’s global state into the
_PyRuntimeState struct. Nearly all the work involves moving that
state from global variables. The project is particularly relevant to
this proposal, so below is some extra detail.
Benefits to Consolidation
Consolidating the globals has a variety of benefits:
- greatly reduces the number of C globals (best practice for C code)
- the move draws attention to runtime state that is unstable or broken
- encourages more consistency in how runtime state is used
- makes it easier to discover/identify CPython’s runtime state
- makes it easier to statically allocate runtime state in a consistent way
- better memory locality for runtime state
Furthermore all the benefits listed in Indirect Benefits above also apply here, and the same projects listed there benefit.
Scale of Work
The number of global variables to be moved is large enough to matter,
but most are Python objects that can be dealt with in large groups
Py_IDENTIFIER). In nearly all cases, moving these globals
to the interpreter is highly mechanical. That doesn’t require
cleverness but instead requires someone to put in the time.
State To Be Moved
The remaining global variables can be categorized as follows:
- global objects
- static types (incl. exception types)
- non-static types (incl. heap types, structseq types)
- singletons (static)
- singletons (initialized once)
- cached objects
- will not (or unlikely to) change after init
- only used in the main thread
- initialized lazily
- pre-allocated buffers
Those globals are spread between the core runtime, the builtin modules, and the stdlib extension modules.
For a breakdown of the remaining globals, run:
./python Tools/c-analyzer/table-file.py Tools/c-analyzer/cpython/globals-to-fix.tsv
Already Completed Work
As mentioned, this work has been going on for many years. Here are some of the things that have already been done:
- cleanup of runtime initialization (see PEP 432 / PEP 587)
- extension module isolation machinery (see PEP 384 / PEP 3121 / PEP 489)
- isolation for many builtin modules
- isolation for many stdlib extension modules
- addition of
- no more
- statically allocated:
- empty string
- string literals
- latin-1 strings
- length-1 bytes
- empty tuple
As already indicated, there are several tools to help identify the globals and reason about them.
Tools/c-analyzer/cpython/globals-to-fix.tsv- the list of remaining globals
analyze- identify all the globals
check- fail if there are any unsupported globals that aren’t ignored
Tools/c-analyzer/table-file.py- summarize the known globals
Also, the check for unsupported globals is incorporated into CI so that no new globals are accidentally added.
Global objects that are safe to be shared (without a GIL) between
interpreters can stay on
_PyRuntimeState. Not only must the object
be effectively immutable (e.g. singletons, strings), but not even the
refcount can change for it to be safe. Immortality (PEP 683)
provides that. (The alternative is that no objects are shared, which
adds significant complexity to the solution, particularly for the
objects exposed in the public C-API.)
Builtin static types are a special case of global objects that will be
shared. They are effectively immutable except for one part:
tp_subclasses). We expect that nothing
else on a builtin type will change, even the content
__subclasses__ for the builtin types will be dealt with by making
it a getter that does a lookup on the current
for that type.
- PEP 384 “Defining a Stable ABI”
- PEP 432 “Restructuring the CPython startup sequence”
- PEP 489 “Multi-phase extension module initialization”
- PEP 554 “Multiple Interpreters in the Stdlib”
- PEP 573 “Module State Access from C Extension Methods”
- PEP 587 “Python Initialization Configuration”
- PEP 630 “Isolating Extension Modules”
- PEP 683 “Immortal Objects, Using a Fixed Refcount”
- PEP 3121 “Extension Module Initialization and Finalization”
This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive.
Last modified: 2023-02-06 22:49:00 GMT