Cinder is Meta's internal performance-oriented production version of CPython 3.10. It contains a number of performance optimizations, including bytecode inline caching, eager evaluation of coroutines, a method-at-a-time JIT, and an experimental bytecode compiler that uses type annotations to emit type-specialized bytecode that performs better in the JIT.
Cinder is powering Instagram, where it started, and is increasingly used across more and more Python applications in Meta.
For more information on CPython, see README.cpython.rst.
Short answer: no.
We've made Cinder publicly available in order to facilitate conversation about potentially upstreaming some of this work to CPython and to reduce duplication of effort among people working on CPython performance.
Cinder is not polished or documented for anyone else's use. We don't have the desire for it to become an alternative to CPython. Our goal in making this code available is a unified faster CPython. So while we do run Cinder in production, if you choose to do so you are on your own. We can't commit to fixing external bug reports or reviewing pull requests. We make sure Cinder is sufficiently stable and fast for our production workloads, but we make no assurances about its stability or correctness or performance for any external workloads or use-cases.
That said, if you have experience in dynamic language runtimes and have ideas to make Cinder faster; or if you work on CPython and want to use Cinder as inspiration for improvements in CPython (or help upstream parts of Cinder to CPython), please reach out; we'd love to chat!
Cinder should build just like CPython; configure and make -j. However
as most development and usage of Cinder occurs in the highly specific context of
Meta we do not exercise it much in other environments. As such, the most
reliable way to build and run Cinder is to re-use the Docker-based setup from
our GitHub CI workflow.
If you just want to get a working Cinder without building it yourself, our Runtime Docker Image is going to be the easiest (no repo clone needed!):
- Install and setup Docker.
- Fetch and run our cinder-runtime image:
docker run -it --rm ghcr.io/facebookincubator/cinder-runtime:cinder-3.10
If you want to build it yourself:
- Install and setup Docker.
- Clone the Cinder repo:
git clone https://github.com/facebookincubator/cinder
- Run a shell in the Docker environment used by the CI:
docker run -v "$PWD/cinder:/vol" -w /vol -it --rm ghcr.io/facebookincubator/cinder/python-build-env:latest bash- The above command does the following:
- Downloads (if not already cached) a pre-built Docker image used by the CI from https://ghcr.io/facebookincubator/cinder/python-build-env.
- Makes the Cinder checkout above ($PWD/cinder) available to the Docker environment at the mount point /vol.
- Interactively (-it) runs bash in the /vol directory.
- Cleanup the local image after it's finished (--rm) to avoid disk bloat.
- Build Cinder from the shell started the Docker environment:
./configure && make
Please be aware that Cinder is only built or tested on Linux x64; anything else
(including macOS) probably won't work. The Docker image above is Fedora
Linux-based and built from a Docker spec file in the Cinder repo:
.github/workflows/python-build-env/Dockerfile.
There are some new test targets that might be interesting. make
testcinder is pretty much the same as make test except that it skips a
few tests that are problematic in our dev environment. make
testcinder_jit runs the test suite with the JIT fully enabled, so all
functions are JIT'ed. make testruntime runs a suite of C++ gtest unit
tests for the JIT. And make test_strict_module runs a test suite for
strict modules (see below).
Note that these steps produce a Cinder Python binary without PGO/LTO optimizations enabled, so don't expect to use these instructions to get any speedup on any Python workload.
Cinder Explorer is a live playground, where you can see how Cinder compiles Python code from source to assembly -- you're welcome to try it out! Feel free to file feature requests and bug reports. Keep in mind that the Cinder Explorer, like the rest of this, "supported" on a best-effort basis.
Instagram uses a multi-process webserver architecture; the parent process starts, performs initialization work (e.g. loading code), and forks tens of worker processes to handle client requests. Worker processes are restarted periodically for a number of reasons (e.g. memory leaks, code deployments) and have a relatively short lifetime. In this model, the OS must copy the entire page containing an object that was allocated in the parent process when the object's reference count is modified. In practice, the objects allocated in the parent process outlive workers; all the work related to reference counting them is unnecessary.
Instagram has a very large Python codebase and the overhead due to copy-on-write from reference counting long-lived objects turned out to be significant. We developed a solution called "immortal instances" to provide a way to opt-out objects from reference counting. See Include/object.h for details. This feature is controlled by defining Py_IMMORTAL_INSTANCES and is enabled by default in Cinder. This was a large win for us in production (~5%), but it makes straight-line code slower. Reference counting operations occur frequently and must check whether or not an object participates in reference counting when this feature is enabled.
"Shadowcode" or "shadow bytecode" is our implementation of a specializing
interpreter. It observes particular optimizable cases in the execution of
generic Python opcodes and (for hot functions) dynamically replaces those
opcodes with specialized versions. The core of shadowcode lives in
Shadowcode/shadowcode.c, though the implementations for the specialized
bytecodes are in Python/ceval.c with the rest of the eval loop.
Shadowcode-specific tests are in Lib/test/test_shadowcode.py.
It is similar in spirit to the specializing adaptive interpreter (PEP-659) that will be built into CPython 3.11.
The Instagram Server is an async-heavy workload, where each web request may trigger hundreds of thousands of async tasks, many of which can be completed without suspension (e.g. thanks to memoized values).
We extended the vectorcall protocol to pass a new flag,
Ci_Py_AWAITED_CALL_MARKER, indicating the caller is immediately awaiting
this call.
When used with async function calls that are immediately awaited, we can immediately (eagerly) evaluate the called function, up to completion, or up to its first suspension. If the function completes without suspending, we are able to return the value immediately, with no extra heap allocations.
When used with async gather, we can immediately (eagerly) evaluate the set of passed awaitables, potentially avoiding the cost of creation and scheduling of multiple tasks for coroutines that could be completed synchronously, completed futures, memoized values, etc.
These optimizations resulted in a significant (~5%) CPU efficiency improvement.
This is mostly implemented in Python/ceval.c, via a new vectorcall flag
Ci_Py_AWAITED_CALL_MARKER, indicating the caller is immediately awaiting
this call. Look for uses of the IS_AWAITED() macro and this vectorcall
flag.
The Cinder JIT is a method-at-a-time custom JIT implemented in C++. It is
enabled via the -X jit flag or the PYTHONJIT=1 environment variable.
It supports almost all Python opcodes, and can achieve 1.5-4x speed
improvements on many Python performance benchmarks.
By default when enabled it will JIT-compile every function that is ever
called, which may well make your program slower, not faster, due to overhead
of JIT-compiling rarely-called functions. The option -X
jit-list-file=/path/to/jitlist.txt or
PYTHONJITLISTFILE=/path/to/jitlist.txt can point it to a text file
containing fully qualified function names (in the form
path.to.module:funcname or path.to.module:ClassName.method_name),
one per line, which should be JIT-compiled. We use this option to compile
only a set of hot functions derived from production profiling data. (A more
typical approach for a JIT would be to dynamically compile functions as they
are observed to be called frequently. It hasn't yet been worth it for us to
implement this, since our production architecture is a pre-fork webserver,
and for memory sharing reasons we wish to do all of our JIT compiling up
front in the initial process before workers are forked, which means we can't
observe the workload in-process before deciding which functions to
JIT-compile.)
The JIT lives in the Jit/ directory, and its C++ tests live in
RuntimeTests/ (run these with make testruntime). There are also some
Python tests for it in Lib/test/test_cinderjit.py; these aren't meant to
be exhaustive, since we run the entire CPython test suite under the JIT via
make testcinder_jit; they cover JIT edge cases not otherwise found in the
CPython test suite.
See Jit/pyjit.cpp for some other -X options and environment variables
that influence the behavior of the JIT. There is also a cinderjit module
defined in that file which exposes some JIT utilities to Python code (e.g.
forcing a specific function to compile, checking if a function is compiled,
disabling the JIT). Note that cinderjit.disable() only disables future
compilation; it immediately compiles all known functions and keeps existing
JIT-compiled functions.
The JIT first lowers Python bytecode to a high-level intermediate
representation (HIR); this is implemented in Jit/hir/. HIR maps
reasonably closely to Python bytecode, though it is a register machine
instead of a stack machine, it is a bit lower level, it is typed, and some
details that are obscured by Python bytecode but important for performance
(notably reference counting) are exposed explicitly in HIR. HIR is
transformed into SSA form, some optimization passes are performed on it, and
then reference counting operations are automatically inserted into it
according to metadata about the refcount and memory effects of HIR opcodes.
HIR is then lowered to a low-level intermediate representation (LIR), which
is an abstraction over assembly, implemented in Jit/lir/. In LIR we do
register allocation, some additional optimization passes, and then finally
LIR is lowered to assembly (in Jit/codegen/) using the excellent
asmjit library.
The JIT is in its early stages. While it can already eliminate interpreter loop overhead and offers significant performance improvements for many functions, we've only begun to scratch the surface of possible optimizations. Many common compiler optimizations are not yet implemented. Our prioritization of optimizations is largely driven by the characteristics of the Instagram production workload.
Strict modules is a few things rolled into one:
1. A static analyzer capable of validating that executing a module's top-level code will not have side effects visible outside that module.
2. An immutable StrictModule type usable in place of Python's default
module type.
3. A Python module loader capable of recognizing modules opted in to strict
mode (via an import __strict__ at the top of the module), analyzing them
to validate no import side effects, and populating them in sys.modules as
a StrictModule object.
Static Python is a bytecode compiler that makes use of type annotations to emit type-specialized and type-checked Python bytecode. Used along with the Cinder JIT, it can deliver performance similar to MyPyC or Cython in many cases, while offering a pure-Python developer experience (normal Python syntax, no extra compilation step). Static Python plus Cinder JIT achieves 18x the performance of stock CPython on a typed version of the Richards benchmark. At Instagram we have successfully used Static Python in production to replace all Cython modules in our primary webserver codebase, with no performance regression.
The Static Python compiler is built on top of the Python compiler module
that was removed from the standard library in Python 3 and has since been
maintained and updated externally; this compiler is incorporated into Cinder
in Lib/compiler. The Static Python compiler is implemented in
Lib/compiler/static/, and its tests are in
Lib/test/test_compiler/test_static.py.
Classes defined in Static Python modules are automatically given typed slots
(based on inspection of their typed class attributes and annotated
assignments in __init__), and attribute loads and stores against
instances of these types use new STORE_FIELD and LOAD_FIELD opcodes,
which in the JIT become direct loads/stores from/to a fixed memory offset in
the object, with none of the indirection of a LOAD_ATTR or
STORE_ATTR. Classes also gain vtables of their methods, for use by the
INVOKE_* opcodes mentioned below. The runtime support for these features
is located in StaticPython/classloader.h and StaticPython/classloader.c.
A static Python function begins with a hidden prologue which checks
that the supplied arguments' types match the type annotations, and raises
TypeError if not. Calls from a static Python function to another static
Python function will skip this opcode (since the types are already validated
by the compiler). Static to static calls can also avoid much of the overhead
of a typical Python function call. We emit an INVOKE_FUNCTION or
INVOKE_METHOD opcode which carries with it metadata about the called
function or method; this plus optionally immutable modules (via
StrictModule) and types (via cinder.freeze_type(), which we currently
apply to all types in strict and static modules in our import loader, but in
future may become an inherent part of Static Python) and compile-time
knowledge of the callee signature allow us to (in the JIT) turn many Python
function calls into direct calls to a fixed memory address using the x64
calling convention, with little more overhead than a C function call.
Static Python is still gradually typed, and supports code that is only
partially annotated or uses unknown types by falling back to normal Python
dynamic behavior. In some cases (e.g. when a value of statically-unknown type
is returned from a function with a return annotation), a runtime CAST
opcode is inserted which will raise TypeError if the runtime type does
not match the expected type.
Static Python also supports new types for machine integers, bools, doubles,
and vectors/arrays. In the JIT these are handled as unboxed values, and e.g.
primitive integer arithmetic avoids all Python overhead. Some operations on
builtin types (e.g. list or dictionary subscript or len()) are also
optimized.
Cinder supports gradual adoption of static modules via a strict/static module
loader that can automatically detect static modules and load them as static
with cross-module compilation. The loader will look for import __static__
and import __strict__ annotations at the top of a file, and compile
modules appropriately. To enable the loader, you have one of three options:
1. Explicitly install the loader at the top level of your application
via from compiler.strict.loader import install; install().
- Set
PYTHONINSTALLSTRICTLOADER=1in your env. - Run
./python -X install-strict-loader application.py.
Alternatively, you can compile all code statically by using
./python -m compiler --static some_module.py,
which will compile the module as static Python and execute it.
See CinderDoc/static_python.rst for more detailed documentation.