Re: [Modeling-users] Modeling/core v0.9pre11 released

[Sebastien B. <sbi...@us...>]

  While the bench script is not completely finished for now, I can give
  some performance indications for this release. For the benchs
  implemented by now, we have:

> ./compare_bench.py perf.0.9pre10 perf.0.9pre11=20
{'bench_addObj_toMany': ('428.705 usec', '196.659 usec', '2.18'),
 'bench_addObj_toOne': ('950.431 usec', '449.753 usec', '2.11'),
 'bench_fetch_persons': ('6.056 s', '5.384 s', '1.12'),
 'bench_fetch_raw_row': ('10.672 ms', '11.005 ms', '0.97'),
 'bench_fetch_raw_rows': ('550.871 ms', '526.122 ms', '1.05'),
 'bench_fetch_raw_rows_all_modified': ('1.803 s', '1.432 s', '1.26')}

(reported time is for python2.2, no optimization, with model caching enabled
for 0.9pre11 --new in this release)

The performance gains are quite similar w/ python -O:

> ./compare_bench.py perf.0.9pre10.opt perf.0.9pre11.opt=20
{'bench_addObj_toMany': ('424.029 usec', '190.539 usec', '2.23'),
 'bench_addObj_toOne': ('951.976 usec', '424.874 usec', '2.24'),
 'bench_fetch_persons': ('5.627 s', '5.030 s', '1.12'),
 'bench_fetch_raw_row': ('10.570 ms', '10.579 ms', '1.00'),
 'bench_fetch_raw_rows': ('555.589 ms', '521.151 ms', '1.07'),
 'bench_fetch_raw_rows_all_modified': ('1.800 s', '1.452 s', '1.24')}


Signification:

  The above lines share the same format:

    'bench_name': (time1, time2, gain=3Dtime1/time2),=20

  where times are as observed on my machine, and benchs' names are:

    - bench_addObj_toMany: addObjectToBothSidesOfRelationshipWithKey()
        on a to-many relationship
=20=20=20=20
    - bench_addObj_toOne: addObjectToBothSidesOfRelationshipWithKey() on
        a to-one relationship
=20=20=20=20
    - bench_fetch_persons: fetch 5000 simple objects (1 string
        attribute, 1 PK, 1 FK hence 1 to-one fault created for each
        object)
=20=20=20=20
    - bench_fetch_raw_row: fetch one single row
=20=20=20=20
    - bench_fetch_raw_rows: fetch 5000 raw rows (same simple objects/
        rows than above)
=20=20=20=20
    - bench_fetch_raw_rows_all_modified: fetch the same 5000 raw rows,
        where the corresponding objects were already fetched in the EC
        and are all modified (this is the worst situation for fetching
        raw rows)


  Rq.: Fetching seems to be equally fast. In fact, it was slowed down a
  little by the extra work needed to support to-many relationships w/o
  inverse. This loss of performance is then regained by caching models
  at runtime.

  The performance boost observed for addObjectToBothSides... is mostly
  due to model's properties caching --plus some "refactoring effect",
  but only for a small part of the perf gain.

  I'd like to thank Yannick in particular here, for his numerous
feedbacks and suggestions about the observed performance bottlenecks in
the framework.

-- S=E9bastien.