[Python-Dev] Real time behaviour of Pythons memory management; WAS: RE: Documentation about Python's GC,python-dev list messages referenced in Modules/gcmodule.c notreachable anymore (original) (raw)

Weber, Gregoire gregoire.weber at schmid-telecom.ch
Wed Jan 4 12:17:32 CET 2006

Garbage Collector findings

To understand pythons garbage collector better and to get a picture about the runtime behaviour and performance I did some experiments:

The attached script allocates a lot of circularly self referencing lists. Then it instantiates a one item list.

I tuned the loop counter in such a way that the subsequent
instantiation of the one item list triggers the garbage collector. The number of the circularly self referencing objects is around 90000.

Results (Pentium4 3GHz, Python 2.4.1 on cygwin/Win2k)

gc gen 1) | time msec 2) | unreachable 3) -----------+--------------+---------------- None | 0.000 | All 0 | 0.002 | All 1 | 0.007 | All 2 | 0.065 | All None | 0.000 | None 0 | 0.001 | None 1 | 0.003 | None 2 | 0.018 | None

--> Collecting a self referencing list costs about 0.7 usec (micro seconds) --> Checking a non collectable self ref list costs about 0.2 usec (micro seconds)

Legend:

  1. the generation the garbage collector has been triggered to collect (None means that GC wasn't triggered)
  2. time for instantiating a list with one entry in msec
  3. All: All of the circularly self referencing lists were unreachable (thus got collected by the gc) None: None of the circularly self referencing lists were unreachable (no garbage available to collect)

Questions

  1. Am I correct that in a system which instantiates a lot*) of containerish objects without destructing them again the GC may be triggered to evaluate all generations which may be very costy (see above)?

  2. In a system where only generation 0 and 1 get evaluated (because not so much object got instantiated without beeing destructed) the costs are relatively low. Correct?

) a lot means here: more than threshold0 * threshold1 * threshold2 objects, Python 2.4 default values are: 70010*10 = 70000

Gregoire

P.S.: Interestingely the counters found by experimenting seem to depend on the platform (my platform: Python 2.4.1, cygwin under Win2k, the counters aren't valid for native Python 2.4.2 under Win2k).

-------------- next part -------------- A non-text attachment was scrubbed... Name: gctest.py Type: application/octet-stream Size: 3936 bytes Desc: gctest.py Url : http://mail.python.org/pipermail/python-dev/attachments/20060104/a19e29c2/attachment.obj

More information about the Python-Dev mailing list