Friday, September 21, 2018

Inside cpyext: Why emulating CPython C API is so Hard


cpyext is PyPy's subsystem which provides a compatibility layer to compile and run CPython C extensions inside PyPy. Often people ask why a particular C extension doesn't work or is very slow on PyPy. Usually it is hard to answer without going into technical details. The goal of this blog post is to explain some of these technical details, so that we can simply link here instead of explaining again and again :).
From a 10.000 foot view, cpyext is PyPy's version of "Python.h". Every time you compile an extension which uses that header file, you are using cpyext. This includes extension explicitly written in C (such as numpy) and extensions which are generated from other compilers/preprocessors (e.g. Cython).
At the time of writing, the current status is that most C extensions "just work". Generally speaking, you can simply pip install them, provided they use the public, official C API instead of poking at private implementation details. However, the performance of cpyext is generally poor. A Python program which makes heavy use of cpyext extensions is likely to be slower on PyPy than on CPython.
Note: in this blog post we are talking about Python 2.7 because it is still the default version of PyPy: however most of the implementation of cpyext is shared with PyPy3, so everything applies to that as well.

C API Overview

In CPython, which is written in C, Python objects are represented as PyObject*, i.e. (mostly) opaque pointers to some common "base struct".
CPython uses a very simple memory management scheme: when you create an object, you allocate a block of memory of the appropriate size on the heap. Depending on the details, you might end up calling different allocators, but for the sake of simplicity, you can think that this ends up being a call to malloc(). The resulting block of memory is initialized and casted to to PyObject*: this address never changes during the object lifetime, and the C code can freely pass it around, store it inside containers, retrieve it later, etc.
Memory is managed using reference counting. When you create a new reference to an object, or you discard a reference you own, you have to increment or decrement the reference counter accordingly. When the reference counter goes to 0, it means that the object is no longer used and can safely be destroyed. Again, we can simplify and say that this results in a call to free(), which finally releases the memory which was allocated by malloc().
Generally speaking, the only way to operate on a PyObject* is to call the appropriate API functions. For example, to convert a given PyObject* to a C integer, you can use PyInt_AsLong(); to add two objects together, you can call PyNumber_Add().
Internally, PyPy uses a similar approach. All Python objects are subclasses of the RPython W_Root class, and they are operated by calling methods on the space singleton, which represents the interpreter.
At first, it looks very easy to write a compatibility layer: just make PyObject* an alias for W_Root, and write simple RPython functions (which will be translated to C by the RPython compiler) which call the space accordingly:
def PyInt_AsLong(space, o):
    return space.int_w(o)

def PyNumber_Add(space, o1, o2):
    return space.add(o1, o2)
Actually, the code above is not too far from the real implementation. However, there are tons of gory details which make it much harder than it looks, and much slower unless you pay a lot of attention to performance.

The PyPy GC

To understand some of cpyext challenges, you need to have at least a rough idea of how the PyPy GC works.
Contrarily to the popular belief, the "Garbage Collector" is not only about collecting garbage: instead, it is generally responsible for all memory management, including allocation and deallocation.
Whereas CPython uses a combination of malloc/free/refcounting to manage memory, the PyPy GC uses a completely different approach. It is designed assuming that a dynamic language like Python behaves the following way:
  • You create, either directly or indirectly, lots of objects.
  • Most of these objects are temporary and very short-lived. Think e.g. of doing a + b + c: you need to allocate an object to hold the temporary result of a + b, then it dies very quickly because you no longer need it when you do the final + c part.
  • Only small fraction of the objects survive and stay around for a while.
So, the strategy is: make allocation as fast as possible; make deallocation of short-lived objects as fast as possible; find a way to handle the remaining small set of objects which actually survive long enough to be important.
This is done using a Generational GC: the basic idea is the following:
  1. We have a nursery, where we allocate "young objects" very quickly.
  2. When the nursery is full, we start what we call a "minor collection".
    • We do a quick scan to determine the small set of objects which survived so far
    • We move these objects out of the nursery, and we place them in the area of memory which contains the "old objects". Since the address of the objects changes, we fix all the references to them accordingly.
  1. now the nursery contains only objects which "died young". We can discard all of them very quickly, reset the nursery, and use the same area of memory to allocate new objects from now.
In practice, this scheme works very well and it is one of the reasons why PyPy is much faster than CPython. However, careful readers have surely noticed that this is a problem for cpyext. On one hand, we have PyPy objects which can potentially move and change their underlying memory address; on the other hand, we need a way to represent them as fixed-address PyObject* when we pass them to C extensions. We surely need a way to handle that.

PyObject* in PyPy

Another challenge is that sometimes, PyObject* structs are not completely opaque: there are parts of the public API which expose to the user specific fields of some concrete C struct. For example the definition of PyTypeObject which exposes many of the tp_* slots to the user. Since the low-level layout of PyPy W_Root objects is completely different than the one used by CPython, we cannot simply pass RPython objects to C; we need a way to handle the difference.
So, we have two issues so far: objects can move, and incompatible low-level layouts. cpyext solves both by decoupling the RPython and the C representations. We have two "views" of the same entity, depending on whether we are in the PyPy world (the movable W_Root subclass) or in the C world (the non-movable PyObject*).
PyObject* are created lazily, only when they are actually needed. The vast majority of PyPy objects are never passed to any C extension, so we don't pay any penalty in that case. However, the first time we pass a W_Root to C, we allocate and initialize its PyObject* counterpart.
The same idea applies also to objects which are created in C, e.g. by calling PyObject_New(). At first, only the PyObject* exists and it is exclusively managed by reference counting. As soon as we pass it to the PyPy world (e.g. as a return value of a function call), we create its W_Root counterpart, which is managed by the GC as usual.
Here we start to see why calling cpyext modules is more costly in PyPy than in CPython. We need to pay some penalty for all the conversions between W_Root and PyObject*.
Moreover, the first time we pass a W_Root to C we also need to allocate the memory for the PyObject* using a slowish "CPython-style" memory allocator. In practice, for all the objects which are passed to C we pay more or less the same costs as CPython, thus effectively "undoing" the speedup guaranteed by PyPy's Generational GC under normal circumstances.

Crossing the border between RPython and C

There are two other things we need to care about whenever we cross the border between RPython and C, and vice-versa: exception handling and the GIL.
In the C API, exceptions are raised by calling PyErr_SetString() (or one of many other functions which have a similar effect), which basically works by creating an exception value and storing it in some global variable. The function then signals that an exception has occurred by returning an error value, usually NULL.
On the other hand, in the PyPy interpreter, exceptions are propagated by raising the RPython-level OperationError exception, which wraps the actual app-level exception values. To harmonize the two worlds, whenever we return from C to RPython, we need to check whether a C API exception was raised and if so turn it into an OperationError.
We won't dig into details of how the GIL is handled in cpyext. For the purpose of this post, it is enough to know that whenever we enter C land, we store the current thread id into a global variable which is accessible also from C; conversely, whenever we go back from RPython to C, we restore this value to 0.
Similarly, we need to do the inverse operations whenever you need to cross the border between C and RPython, e.g. by calling a Python callback from C code.
All this complexity is automatically handled by the RPython function generic_cpy_call. If you look at the code you see that it takes care of 4 things:
  1. Handling the GIL as explained above.
  2. Handling exceptions, if they are raised.
  3. Converting arguments from W_Root to PyObject*.
  4. Converting the return value from PyObject* to W_Root.
So, we can see that calling C from RPython introduce some overhead. Can we measure it?
Assuming that the conversion between W_Root and PyObject* has a reasonable cost (as explained by the previous section), the overhead introduced by a single border-cross is still acceptable, especially if the callee is doing some non-negligible amount of work.
However this is not always the case. There are basically three problems that make (or used to make) cpyext super slow:
  1. Paying the border-crossing cost for trivial operations which are called very often, such as Py_INCREF.
  2. Crossing the border back and forth many times, even if it's not strictly needed.
  3. Paying an excessive cost for argument and return value conversions.
The next sections explain in more detail each of these problems.

Avoiding unnecessary roundtrips

Prior to the 2017 Cape Town Sprint, cpyext was horribly slow, and we were well aware of it: the main reason was that we never really paid too much attention to performance. As explained in the blog post, emulating all the CPython quirks is basically a nightmare, so better to concentrate on correctness first.
However, we didn't really know why it was so slow. We had theories and assumptions, usually pointing at the cost of conversions between W_Root and PyObject*, but we never actually measured it.
So, we decided to write a set of cpyext microbenchmarks to measure the performance of various operations. The result was somewhat surprising: the theory suggests that when you do a cpyext C call, you should pay the border-crossing costs only once, but what the profiler told us was that we were paying the cost of generic_cpy_call several times more than what we expected.
After a bit of investigation, we discovered this was ultimately caused by our "correctness-first" approach. For simplicity of development and testing, when we started cpyext we wrote everything in RPython: thus, every single API call made from C (like the omnipresent PyArg_ParseTuple(), PyInt_AsLong(), etc.) had to cross back the C-to-RPython border. This was especially daunting for very simple and frequent operations like Py_INCREF and Py_DECREF, which CPython implements as a single assembly instruction!
Another source of slow down was the implementation of PyTypeObject slots. At the C level, these are function pointers which the interpreter calls to do certain operations, e.g. tp_new to allocate a new instance of that type.
As usual, we have some magic to implement slots in RPython; in particular, _make_wrapper does the opposite of generic_cpy_call: it takes a RPython function and wraps it into a C function which can be safely called from C, handling the GIL, exceptions and argument conversions automatically.
This was very handy during the development of cpyext, but it might result in some bad nonsense; consider what happens when you call the following C function:
static PyObject* foo(PyObject* self, PyObject* args)
{
    PyObject* result = PyInt_FromLong(1234);
    return result;
}
  1. you are in RPython and do a cpyext call to foo: RPython-to-C;
  2. foo calls PyInt_FromLong(1234), which is implemented in RPython: C-to-RPython;
  3. the implementation of PyInt_FromLong indirectly calls PyIntType.tp_new, which is a C function pointer: RPython-to-C;
  4. however, tp_new is just a wrapper around an RPython function, created by _make_wrapper: C-to-RPython;
  5. finally, we create our RPython W_IntObject(1234); at some point during the RPython-to-C crossing, its PyObject* equivalent is created;
  6. after many layers of wrappers, we are again in foo: after we do return result, during the C-to-RPython step we convert it from PyObject* to W_IntObject(1234).
Phew! After we realized this, it was not so surprising that cpyext was very slow :). And this was a simplified example, since we are not passing a PyObject* to the API call. When we do, we need to convert it back and forth at every step. Actually, I am not even sure that what I described was the exact sequence of steps which used to happen, but you get the general idea.
The solution is simple: rewrite as much as we can in C instead of RPython, to avoid unnecessary roundtrips. This was the topic of most of the Cape Town sprint and resulted in the cpyext-avoid-roundtrip branch, which was eventually merged.
Of course, it is not possible to move everything to C: there are still operations which need to be implemented in RPython. For example, think of PyList_Append: the logic to append an item to a list is complex and involves list strategies, so we cannot replicate it in C. However, we discovered that a large subset of the C API can benefit from this.
Moreover, the C API is huge. While we invented this new way of writing cpyext code, we still need to convert many of the functions to the new paradigm. Sometimes the rewrite is not automatic or straighforward. cpyext is a delicate piece of software, so it happens often that we make a mistake and end up staring at a segfault in gdb.
However, the most important takeaway is that the performance improvements we got from this optimization are impressive, as we will detail later.

Conversion costs

The other potential big source of slowdown is the conversion of arguments between W_Root and PyObject*.
As explained earlier, the first time you pass a W_Root to C, you need to allocate its PyObject* counterpart. Suppose you have a foo function defined in C, which takes a single int argument:
for i in range(N):
    foo(i)
To run this code, you need to create a different PyObject* for each value of i: if implemented naively, it means calling N times malloc() and free(), which kills performance.
CPython has the very same problem, which is solved by using a free list to allocate ints. So, what we did was to simply steal the code from CPython and do the exact same thing. This was also done in the cpyext-avoid-roundtrip branch, and the benchmarks show that it worked perfectly.
Every type which is converted often to PyObject* must have a very fast allocator. At the moment of writing, PyPy uses free lists only for ints and tuples: one of the next steps on our TODO list is certainly to use this technique with more types, like float.
Conversely, we also need to optimize the converstion from PyObject* to W_Root: this happens when an object is originally allocated in C and returned to Python. Consider for example the following code:
import numpy as np
myarray = np.random.random(N)
for i in range(len(arr)):
    myarray[i]
At every iteration, we get an item out of the array: the return type is a an instance of numpy.float64 (a numpy scalar), i.e. a PyObject'*: this is something which is implemented by numpy entirely in C, so completely opaque to cpyext. We don't have any control on how it is allocated, managed, etc., and we can assume that allocation costs are the same as on CPython.
As soon as we return these PyObject* to Python, we need to allocate their W_Root equivalent. If you do it in a small loop like in the example above, you end up allocating all these W_Root inside the nursery, which is a good thing since allocation is super fast (see the section above about the PyPy GC).
However, we also need to keep track of the W_Root to PyObject* link. Currently, we do this by putting all of them in a dictionary, but it is very inefficient, especially because most of these objects die young and thus it is wasted work to do that for them. Currently, this is one of the biggest unresolved problem in cpyext, and it is what causes the two microbenchmarks allocate_int and allocate_tuple to be very slow.
We are well aware of the problem, and we have a plan for how to fix it. The explanation is too technical for the scope of this blog post as it requires a deep knowledge of the GC internals to be understood, but the details are here.

C API quirks

Finally, there is another source of slowdown which is beyond our control. Some parts of the CPython C API are badly designed and expose some of the implementation details of CPython.
The major example is reference counting. The Py_INCREF / Py_DECREF API is designed in such a way which forces other implementation to emulate refcounting even in presence of other GC management schemes, as explained above.
Another example is borrowed references. There are API functions which do not incref an object before returning it, e.g. PyList_GetItem(). This is done for performance reasons because we can avoid a whole incref/decref pair, if the caller needs to handle the returned item only temporarily: the item is kept alive because it is in the list anyway.
For PyPy, this is a challenge: thanks to list strategies, lists are often represented in a compact way. For example, a list containing only integers is stored as a C array of long. How to implement PyList_GetItem? We cannot simply create a PyObject* on the fly, because the caller will never decref it and it will result in a memory leak.
The current solution is very inefficient. The first time we do a PyList_GetItem, we convert the whole list to a list of PyObject*. This is bad in two ways: the first is that we potentially pay a lot of unneeded conversion cost in case we will never access the other items of the list. The second is that by doing that we lose all the performance benefit granted by the original list strategy, making it slower for the rest of the pure-python code which will manipulate the list later.
PyList_GetItem is an example of a bad API because it assumes that the list is implemented as an array of PyObject*: after all, in order to return a borrowed reference, we need a reference to borrow, don't we?
Fortunately, (some) CPython developers are aware of these problems, and there is an ongoing project to design a better C API which aims to fix exactly this kind of problem.
Nonetheless, in the meantime we still need to implement the current half-broken APIs. There is no easy solution for that, and it is likely that we will always need to pay some performance penalty in order to implement them correctly.
However, what we could potentially do is to provide alternative functions which do the same job but are more PyPy friendly: for example, we could think of implementing PyList_GetItemNonBorrowed or something like that: then, C extensions could choose to use it (possibly hidden inside some macro and #ifdef) if they want to be fast on PyPy.

Current performance

During the whole blog post we claimed cpyext is slow. How slow it is, exactly?
We decided to concentrate on microbenchmarks for now. It should be evident by now there are simply too many issues which can slow down a cpyext program, and microbenchmarks help us to concentrate on one (or few) at a time.
The microbenchmarks measure very simple things, like calling functions and methods with the various calling conventions (no arguments, one arguments, multiple arguments); passing various types as arguments (to measure conversion costs); allocating objects from C, and so on.
Here are the results from the old PyPy 5.8 relative and normalized to CPython 2.7, the lower the better:



PyPy was horribly slow everywhere, ranging from 2.5x to 10x slower. It is particularly interesting to compare simple.noargs, which measures the cost of calling an empty function with no arguments, and simple.onearg(i), which measures the cost calling an empty function passing an integer argument: the latter is ~2x slower than the former, indicating that the conversion cost of integers is huge.
PyPy 5.8 was the last release before the famous Cape Town sprint, when we started to look at cpyext performance seriously. Here are the performance data for PyPy 6.0, the latest release at the time of writing:


The results are amazing! PyPy is now massively faster than before, and for most benchmarks it is even faster than CPython: yes, you read it correctly: PyPy is faster than CPython at doing CPython's job, even considering all the extra work it has to do to emulate the C API. This happens thanks to the JIT, which produces speedups high enough to counterbalance the slowdown caused by cpyext.
There are two microbenchmarks which are still slower though: allocate_int and allocate_tuple, for the reasons explained in the section about Conversion costs.

Next steps

Despite the spectacular results we got so far, cpyext is still slow enough to kill performance in most real-world code which uses C extensions extensively (e.g., the omnipresent numpy).
Our current approach is something along these lines:
  1. run a real-world small benchmark which exercises cpyext
  2. measure and find the major bottleneck
  3. write a corresponding microbenchmark
  4. optimize it
  5. repeat
On one hand, this is a daunting task because the C API is huge and we need to tackle functions one by one. On the other hand, not all the functions are equally important, and is is enough to optimize a relatively small subset to improve many different use cases.
Where a year ago we announced we have a working answer to run c-extension in PyPy, we now have a clear picture of what are the performance bottlenecks, and we have developed some technical solutions to fix them. It is "only" a matter of tackling them, one by one. It is worth noting that most of the work was done during two sprints, for a total 2-3 person-months of work.
We think this work is important for the Python ecosystem. PyPy has established a baseline for performance in pure python code, providing an answer for the "Python is slow" detractors. The techniques used to make cpyext performant will let PyPy become an alternative for people who mix C extensions with Python, which, it turns out, is just about everyone, in particular those using the various scientific libraries. Today, many developers are forced to seek performance by converting code from Python to a lower language. We feel there is no reason to do this, but in order to prove it we must be able to run both their python and their C extensions performantly, then we can begin to educate them how to write JIT-friendly code in the first place.
We envision a future in which you can run arbitrary Python programs on PyPy, with the JIT speeding up the pure Python parts and the C parts running as fast as today: the best of both worlds!

3 comments:

  1. Thanks fo this nice article!

    —Albert

    ReplyDelete
  2. Great work guys! I should benchmark some of my apps again - a couple of things that were dependent on C extensions didn't show much speedup previously.

    ReplyDelete

See also PyPy's IRC channel: #pypy at freenode.net, or the pypy-dev mailing list.
If the blog post is old, it is pointless to ask questions here about it---you're unlikely to get an answer.