Datastore profiling

Tomeu Vizoso tomeu at tomeuvizoso.net
Wed Aug 13 06:18:29 EDT 2008


On Wed, Aug 13, 2008 at 12:10 PM, riccardo <riccardo.lucchese at gmail.com> wrote:
> On Wed, 2008-08-13 at 11:41 +0200, Tomeu Vizoso wrote:
>> On Sat, Aug 9, 2008 at 3:30 PM, riccardo <riccardo.lucchese at gmail.com> wrote:
>> > Hi,
>> >
>> >
>> > Testcase:
>> > Fill the datastore with `many small' objects.
>>
>> Which user-perceived action are you investigating?
> I thought the datastore is one of the `components' we wanted to profile.
> This was straightforward after writing the script that populates the DS
> that I wanted to use to profile the journal.

Sure it is, but we need to take in account the user perceived
performance when choosing what to optimize. We have very few resources
and need to focus on what will give most value to our users.

>> > The test was automated with the script at:
>> > http://dev.laptop.org/~rlucchese/datastore_fill/fill_ds
>>
>> Why the need to restart sugar?
> I only send one signal to the DS (to stop and exit); in fact it can be
> done also without restarting sugar and sending a `start' signal to the
> process (so that the profiler gets enabled)

Oh, I see. Yeah, adding a couple of dbus methods for start and stop
profiling would be quite easy.

>> >  + Sysprof statistics:
>> > http://dev.laptop.org/~rlucchese/datastore_fill/small_files__b20000000.sysprof
>> >
>> > I can't see anything relevant for libxapian and it's bindings (xapian
>> > dbg packages were installed)
>>
>> Perhaps because of the small amount of metadata? PROPS_WITH_PREVIEW in
>> http://dev.laptop.org/git?p=projects/datastore;a=blob;f=tests/test_sugar.py
>> would get you quite close to the real case.
> yeah, perhaps it makes sense to add more metadata also for the journal
> scrolling test ?

Oh, totally.

Thanks,

Tomeu



More information about the Devel mailing list