Garbage collector documentation and configuration¶
PyPy’s default garbage collector is called incminimark - it’s an incremental, generational moving collector. Here we hope to explain a bit how it works and how it can be tuned to suit the workload.
Incminimark first allocates objects in so called nursery - place for young objects, where allocation is very cheap, being just a pointer bump. The nursery size is a very crucial variable - depending on your workload (one or many processes) and cache sizes you might want to experiment with it via PYPY_GC_NURSERY environment variable. When the nursery is full, there is performed a minor collection. Freed objects are no longer referencable and just die, just by not being referenced any more; on the other hand, objects found to still be alive must survive and are copied from the nursery to the old generation. Either to arenas, which are collections of objects of the same size, or directly allocated with malloc if they’re larger. (A third category, the very large objects, are initially allocated outside the nursery and never move.)
Since Incminimark is an incremental GC, the major collection is incremental, meaning there should not be any pauses longer than 1ms.
Before we discuss issues of “fragmentation”, we need a bit of precision. There are two kinds of related but distinct issues:
- If the program allocates a lot of memory, and then frees it all by dropping all references to it, then we might expect to see the RSS to drop. (RSS = Resident Set Size on Linux, as seen by “top”; it is an approximation of the actual memory usage from the OS’s point of view.) This might not occur: the RSS may remain at its highest value. This issue is more precisely caused by the process not returning “free” memory to the OS. We call this case “unreturned memory”.
- After doing the above, if the RSS didn’t go down, then at least future allocations should not cause the RSS to grow more. That is, the process should reuse unreturned memory as long as it has got some left. If this does not occur, the RSS grows even larger and we have real fragmentation issues.
There is a special function in the
gc module called
memory_pressure controls whether or not to report memory pressure from
objects allocated outside of the GC, which requires walking the entire heap,
so it’s disabled by default due to its cost. Enable it when debugging
mysterious memory disappearance.
Example call looks like that:
>>> gc.get_stats(True) Total memory consumed: GC used: 4.2MB (peak: 4.2MB) in arenas: 763.7kB rawmalloced: 383.1kB nursery: 3.1MB raw assembler used: 0.0kB memory pressure: 0.0kB ----------------------------- Total: 4.2MB Total memory allocated: GC allocated: 4.5MB (peak: 4.5MB) in arenas: 763.7kB rawmalloced: 383.1kB nursery: 3.1MB raw assembler allocated: 0.0kB memory pressure: 0.0kB ----------------------------- Total: 4.5MB
In this particular case, which is just at startup, GC consumes relatively
little memory and there is even less unused, but allocated memory. In case
there is a lot of unreturned memory or actual fragmentation, the “allocated”
can be much higher than “used”. Generally speaking, “peak” will more closely
resemble the actual memory consumed as reported by RSS. Indeed, returning
memory to the OS is a hard and not solved problem. In PyPy, it occurs only if
an arena is entirely free—a contiguous block of 64 pages of 4 or 8 KB each.
It is also rare for the “rawmalloced” category, at least for common system
The details of various fields:
- GC in arenas - small old objects held in arenas. If the amount “allocated”
is much higher than the amount “used”, we have unreturned memory. It is
possible but unlikely that we have internal fragmentation here. However,
this unreturned memory cannot be reused for any
malloc(), including the memory from the “rawmalloced” section.
- GC rawmalloced - large objects allocated with malloc. This is gives the
current (first block of text) and peak (second block of text) memory
malloc(). The amount of unreturned memory or fragmentation caused by
malloc()cannot easily be reported. Usually you can guess there is some if the RSS is much larger than the total memory reported for “GC allocated”, but do keep in mind that this total does not include malloc’ed memory not known to PyPy’s GC at all. If you guess there is some, consider using jemalloc as opposed to system malloc.
- nursery - amount of memory allocated for nursery, fixed at startup, controlled via an environment variable
- raw assembler allocated - amount of assembler memory that JIT feels responsible for
- memory pressure, if asked for - amount of memory we think got allocated via external malloc (eg loading cert store in SSL contexts) that is kept alive by GC objects, but not accounted in the GC
GC hooks are user-defined functions which are called whenever a specific GC event occur, and can be used to monitor GC activity and pauses. You can install the hooks by setting the following attributes:
- Called whenever a minor collection occurs. It corresponds to
- Called whenever an incremental step of a major collection occurs. It
- Called after the last incremental step, when a major collection is fully
done. It corresponds to
To uninstall a hook, simply set the corresponding attribute to
install all hooks at once, you can call
gc.hooks.set(obj), which will look
obj. To uninstall all the hooks at once, you
The functions called by the hooks receive a single
stats argument, which
contains various statistics about the event.
Note that PyPy cannot call the hooks immediately after a GC event, but it has
to wait until it reaches a point in which the interpreter is in a known state
and calling user-defined code is harmless. It might happen that multiple
events occur before the hook is invoked: in this case, you can inspect the
stats.count to know how many times the event occurred since the last
time the hook was called. Similarly,
stats.duration contains the
total time spent by the GC for this specific event since the last time the
hook was called.
On the other hand, all the other fields of the
stats object are relative
only to the last event of the series.
The attributes for
- The number of minor collections occurred since the last hook call.
- The total time spent inside minor collections since the last hook call. See below for more information on the unit.
- The duration of the fastest minor collection since the last hook call.
- The duration of the slowest minor collection since the last hook call.
- The amount of memory used at the end of the minor collection, in bytes. This include the memory used in arenas (for GC-managed memory) and raw-malloced memory (e.g., the content of numpy arrays).
- the number of pinned objects.
The attributes for
- See above.
- Integers which indicate the state of the GC before and after the step.
The value of
newstate is one of these constants, defined
STATE_FINALIZING. It is possible to get a string
representation of it by indexing the
The attributes for
- See above.
- The total number of major collections which have been done since the
start. Contrarily to
count, this is an always-growing counter and it’s not reset between invocations.
- Number of arenas used before and after the major collection.
- Total number of bytes used by GC-managed objects.
- Total number of bytes used by raw-malloced objects, before and after the major collection.
GcCollectStats has not got a
duration field. This is
because all the GC work is done inside
gc-collect-done is used only to give additional stats, but doesn’t do any
A note about the
duration field: depending on the architecture and
operating system, PyPy uses different ways to read timestamps, so
is expressed in varying units. It is possible to know which by calling
__pypy__.debug_get_timestamp_unit(), which can be one of the following
- The default on
x86machines: timestamps are expressed in CPU ticks, as read by the Time Stamp Counter.
- Timestamps are expressed in nanoseconds.
- On Windows, in case for some reason
tscis not available: timestamps are read using the win API
Unfortunately, there does not seem to be a reliable standard way for
tsc ticks into nanoseconds, although in practice on modern CPUs
it is enough to divide the ticks by the maximum nominal frequency of the CPU.
For this reason, PyPy gives the raw value, and leaves the job of doing the
conversion to external libraries.
Here is an example of GC hooks in use:
import sys import gc class MyHooks(object): done = False def on_gc_minor(self, stats): print 'gc-minor: count = %02d, duration = %d' % (stats.count, stats.duration) def on_gc_collect_step(self, stats): old = gc.GcCollectStepStats.GC_STATES[stats.oldstate] new = gc.GcCollectStepStats.GC_STATES[stats.newstate] print 'gc-collect-step: %s --> %s' % (old, new) print ' count = %02d, duration = %d' % (stats.count, stats.duration) def on_gc_collect(self, stats): print 'gc-collect-done: count = %02d' % stats.count self.done = True hooks = MyHooks() gc.hooks.set(hooks) # simulate some GC activity lst =  while not hooks.done: lst = [lst, 1, 2, 3]
incminimark garbage collector is configurable through
several environment variables:
- The nursery size.
Defaults to 1/2 of your last-level cache, or
4Mif unknown. Small values (like 1 or 1KB) are useful for debugging.
- If set to non-zero, will fill nursery with garbage, to help debugging.
- The size of memory marked during the marking step. Default is size of nursery times 2. If you mark it too high your GC is not incremental at all. The minimum is set to size that survives minor collection times 1.5 so we reclaim anything all the time.
- Major collection memory factor.
1.82, which means trigger a major collection when the memory consumed equals 1.82 times the memory really used at the end of the previous major collection.
- Major collection threshold’s max growth rate.
1.4. Useful to collect more often than normally on sudden memory growth, e.g. when there is a temporary peak in memory usage.
- The max heap size.
If coming near this limit, it will first collect more often, then
raise an RPython MemoryError, and if that is not enough, crash the
program with a fatal error.
Try values like
- The major collection threshold will never be set to more than
PYPY_GC_MAX_DELTAthe amount really used after a collection. Defaults to 1/8th of the total RAM size (which is constrained to be at most 2/3/4GB on 32-bit systems). Try values like
- Don’t collect while the memory size is below this limit. Useful to avoid spending all the time in the GC in very small programs. Defaults to 8 times the nursery.
- Enable extra checks around collections that are too slow for normal
1(on major collections) or
2(also on minor collections).
- The maximal number of pinned objects at any point in time. Defaults to a conservative value depending on nursery size and maximum object size inside the nursery. Useful for debugging by setting it to 0.