Re: [Sqlalchemy-tickets] [sqlalchemy] #2720: new C extensions for attribute accessing

[sqlalchemy <mi...@zz...>]

#2720: new C extensions for attribute accessing
-----------------------------------+---------------------------------------
      Reporter:  zzzeek            |      Owner:  zzzeek
          Type:  enhancement       |     Status:  new
      Priority:  medium            |  Milestone:  0.9.0
     Component:  cextensions       |   Severity:  very major - up to 2 days
    Resolution:                    |   Keywords:
Progress State:  not decided upon  |
-----------------------------------+---------------------------------------

Comment (by claudiofreire):

 Right. Well, {{{key in dict_}}} then {{{return dict_[key]}}} does the
 lookup twice, and catching the !KeyError has to set up a try/catch block,
 which is what is expensive about it (kinda like a stack bookmark CPython's
 VM can roll back to).

 But the real savings here come from:

  * {{{Entity.id}}} goes from CPython's VM (LOAD_ATTR) directly towards the
 C code, there's no frame setup or bytecodes involved (that's why the call
 doesn't show up in profiles). With these ultra-cheap functions, frame
 setup can be 90% of the cost.
  * {{{PyBlah_Bleh}}} calls are all native. While, true, not immensely
 quicker than equivalent VM bytecodes, the fact that they're just working
 with simple dict ops and non-property attributes makes the VM itself a
 hinderance. {{{Entity.id}}} was about a dozen opcodes, that's a dozen
 loops over a huge switch block that interprets the bytecode. Now it's a
 pair of native calls. That's '''way''' faster in comparison.

 While those two wouldn't normally matter much in the overall picture, the
 fact that properly optimized CPU-bound SQLAlchemy apps tend to be all
 about reading attributes (in my experience), changes things.

 I didn't try the timeit. I will as soon as I get back to my SA devel
 machine (ie: the office).

-- 
Ticket URL: <http://www.sqlalchemy.org/trac/ticket/2720#comment:15>
sqlalchemy <http://www.sqlalchemy.org/>
The Database Toolkit for Python