Garbage Collection in PyPy
The overview and description of our garbage collection strategy and framework can be found in the EU-report on this topic. Please refer to that file for an old, but still more or less accurate, description. The present document describes the specific garbage collectors that we wrote in our framework.
Reminder: to select which GC you want to include in a translated RPython program, use the --gc=NAME option of translate.py. For more details, see the overview of command line options for translation.
The following overview is written in chronological order, so the "best" GC (which is the default when translating) is the last one below.
Classical Mark and Sweep collector. Also contained a lot of experimental and half-unmaintained features. Was removed.
Two arenas of equal size, with only one arena in use and getting filled with new objects. When the arena is full, the live objects are copied into the other arena using Cheney's algorithm. The old arena is then cleared. See `rpython/memory/gc/semispace.py`_.
On Unix the clearing is done by reading /dev/zero into the arena, which is extremely memory efficient at least on Linux: it lets the kernel free the RAM that the old arena used and replace it all with allocated-on-demand memory.
The size of each semispace starts at 8MB but grows as needed when the amount of objects alive grows.
This is a two-generations GC. See `rpython/memory/gc/generation.py`_.
It is implemented as a subclass of the Semispace copying collector. It adds a nursery, which is a chunk of the current semispace. Its size is computed to be half the size of the CPU Level 2 cache. Allocations fill the nursery, and when it is full, it is collected and the objects still alive are moved to the rest of the current semispace.
The idea is that it is very common for objects to die soon after they are created. Generational GCs help a lot in this case, particularly if the amount of live objects really manipulated by the program fits in the Level 2 cache. Moreover, the semispaces fill up much more slowly, making full collections less frequent.
This is a three-generations GC.
It is implemented as a subclass of the Generational GC. The Hybrid GC can handle both objects that are inside and objects that are outside the semispaces ("external"). The external objects are not moving and collected in a mark-and-sweep fashion. Large objects are allocated as external objects to avoid costly moves. Small objects that survive for a long enough time (several semispace collections) are also made external so that they stop moving.
This is coupled with a segregation of the objects in three generations. Each generation is collected much less often than the previous one. The division of the generations is slightly more complicated than just nursery / semispace / external; see the diagram at the start of the source code, in `rpython/memory/gc/hybrid.py`_.
Killed in trunk. The following documentation is for historical purposes only.
Inspired, at least partially, by Squeak's garbage collector, this is a single-arena GC in which collection compacts the objects in-place. The main point of this GC is to save as much memory as possible (to be not worse than the Semispace), but without the peaks of double memory usage during collection.
Unlike the Semispace GC, collection requires a number of passes over the data. This makes collection quite slower. Future improvements could be to add a nursery to Mark & Compact in order to mitigate this issue.
During a collection, we reuse the space in-place if it is still large enough. If not, we need to allocate a new, larger space, and move the objects there; however, this move is done chunk by chunk, and chunks are cleared (i.e. returned to the OS) as soon as they have been moved away. This means that (from the point of view of the OS) a collection will never cause an important temporary growth of total memory usage.
More precisely, a collection is triggered when the space contains more than N*M bytes, where N is the number of bytes alive after the previous collection and M is a constant factor, by default 1.5. This guarantees that the total memory usage of the program never exceeds 1.5 times the total size of its live objects.
The objects themselves are quite compact: they are allocated next to each other in the heap, separated by a GC header of only one word (4 bytes on 32-bit platforms) and possibly followed by up to 3 bytes of padding for non-word-sized objects (e.g. strings). There is a small extra memory usage during collection: an array containing 2 bytes per surviving object is needed to make a backup of (half of) the surviving objects' header, in order to let the collector store temporary relation information in the regular headers.
This is a simplification and rewrite of the ideas from the Hybrid GC. It uses a nursery for the young objects, and mark-and-sweep for the old objects. This is a moving GC, but objects may only move once (from the nursery to the old stage).
The main difference with the Hybrid GC is that the mark-and-sweep objects (the "old stage") are directly handled by the GC's custom allocator, instead of being handled by malloc() calls. The gain is that it is then possible, during a major collection, to walk through all old generation objects without needing to store a list of pointers to them. So as a first approximation, when compared to the Hybrid GC, the Minimark GC saves one word of memory per old object.
There are :ref:`a number of environment variables <minimark-environment-variables>` that can be tweaked to influence the GC. (Their default value should be ok for most usages.)
In more detail:
- The small newly malloced objects are allocated in the nursery (case 1). All objects living in the nursery are "young".
- The big objects are always handled directly by the system malloc(). But the big newly malloced objects are still "young" when they are allocated (case 2), even though they don't live in the nursery.
- When the nursery is full, we do a minor collection, i.e. we find which "young" objects are still alive (from cases 1 and 2). The "young" flag is then removed. The surviving case 1 objects are moved to the old stage. The dying case 2 objects are immediately freed.
- The old stage is an area of memory containing old (small) objects. It is handled by `rpython/memory/gc/minimarkpage.py`_. It is organized as "arenas" of 256KB or 512KB, subdivided into "pages" of 4KB or 8KB. Each page can either be free, or contain small objects of all the same size. Furthermore at any point in time each object location can be either allocated or freed. The basic design comes from obmalloc.c from CPython (which itself comes from the same source as the Linux system malloc()).
- New objects are added to the old stage at every minor collection. Immediately after a minor collection, when we reach some threshold, we trigger a major collection. This is the mark-and-sweep step. It walks over all objects (mark), and then frees some fraction of them (sweep). This means that the only time when we want to free objects is while walking over all of them; we never ask to free an object given just its address. This allows some simplifications and memory savings when compared to obmalloc.c.
- As with all generational collectors, this GC needs a write barrier to record which old objects have a reference to young objects.
- Additionally, we found out that it is useful to handle the case of big arrays specially: when we allocate a big array (with the system malloc()), we reserve a small number of bytes before. When the array grows old, we use the extra bytes as a set of bits. Each bit represents 128 entries in the array. Whenever the write barrier is called to record a reference from the Nth entry of the array to some young object, we set the bit number (N/128) to 1. This can considerably speed up minor collections, because we then only have to scan 128 entries of the array instead of all of them.
- As usual, we need special care about weak references, and objects with finalizers. Weak references are allocated in the nursery, and if they survive they move to the old stage, as usual for all objects; the difference is that the reference they contain must either follow the object, or be set to NULL if the object dies. And the objects with finalizers, considered rare enough, are immediately allocated old to simplify the design. In particular their __del__ method can only be called just after a major collection.
- The objects move once only, so we can use a trick to implement id() and hash(). If the object is not in the nursery, it won't move any more, so its id() and hash() are the object's address, cast to an integer. If the object is in the nursery, and we ask for its id() or its hash(), then we pre-reserve a location in the old stage, and return the address of that location. If the object survives the next minor collection, we move it there, and so its id() and hash() are preserved. If the object dies then the pre-reserved location becomes free garbage, to be collected at the next major collection.