[ovirt-devel] ovirt-engine-sdk-python too slow
Juan Hernández
jhernand at redhat.com
Wed Nov 25 18:15:25 UTC 2015
On 11/25/2015 06:45 PM, Nir Soffer wrote:
> $ ./profile-stats -c myscript.prof
>
> Wed Nov 25 10:40:11 2015 myscript.prof
>
> 7892315 function calls (7891054 primitive calls) in 7.940 seconds
>
> Ordered by: internal time
> List reduced from 1518 to 20 due to restriction <20>
>
> ncalls tottime percall cumtime percall filename:lineno(function)
> 9086 2.693 0.000 6.706 0.001 inspect.py:247(getmembers)
> 1952494 1.394 0.000 1.880 0.000 inspect.py:59(isclass)
> 9092 1.030 0.000 1.030 0.000 {dir}
> 1952642 0.600 0.000 0.600 0.000 {getattr}
> 1972765 0.504 0.000 0.504 0.000 {isinstance}
> 3 0.334 0.111 0.334 0.111 {method 'perform' of
> 'pycurl.Curl' objects}
> 1883918 0.284 0.000 0.284 0.000 {method 'append' of 'list'
> objects}
> 9087 0.221 0.000 0.221 0.000 {method 'sort' of 'list'
> objects}
> 9051 0.172 0.000 6.911 0.001
> reflectionhelper.py:51(isModuleMember)
> 1 0.124 0.124 0.354 0.354 errors.py:17(<module>)
> 1 0.088 0.088 0.230 0.230 params.py:8(<module>)
> 8879 0.070 0.000 6.998 0.001 params.py:367(__setattr__)
> 1 0.047 0.047 5.182 5.182 api.py:23(<module>)
> 1 0.025 0.025 4.743 4.743 brokers.py:22(<module>)
> 1 0.023 0.023 0.030 0.030
> connectionspool.py:17(<module>)
> 1 0.022 0.022 0.053 0.053
> lxml.etree.pyx:1(PyMODINIT_FUNC PyInit_etree(void))
> 118 0.019 0.000 4.684 0.040 params.py:45277(__init__)
> 5 0.015 0.003 0.024 0.005 {built-in method strptime}
> 1 0.012 0.012 0.013 0.013 socket.py:45(<module>)
> 10 0.011 0.001 0.015 0.002 collections.py:288(namedtuple)
>
> So it is not the classes, it is the code inspecting them on import.
>
The script doesn't contain only the imports, it is also calling the
server, and we know parsing the result is slow, due to the excesive use
of "inspect", as I mentioned before:
[RFE][performance] - generate large scale list running to slow.
https://bugzilla.redhat.com/show_bug.cgi?id=1221238#c2
In the profiling information seems to corresponds to the script before
commenting out the part that lists all the VMs, as it looks like the
constructor of the VM class was called 21 times (you probably have 21 VMs):
21 0.004 1.308
build/bdist.linux-x86_64/egg/ovirtsdk/infrastructure/brokers.py:29139(VM)
> Nir
>
> On Wed, Nov 25, 2015 at 7:49 AM, John Hunter <zhjwpku at gmail.com
> <mailto:zhjwpku at gmail.com>> wrote:
>
> Hi Nir,
>
> Attachment is my script and its profile.
> Thanks a lot about your help!
>
> On Wed, Nov 25, 2015 at 5:49 AM, Nir Soffer <nsoffer at redhat.com
> <mailto:nsoffer at redhat.com>> wrote:
>
> On Tue, Nov 24, 2015 at 3:49 PM, John Hunter <zhjwpku at gmail.com
> <mailto:zhjwpku at gmail.com>> wrote:
>
>
>
> On Tue, Nov 24, 2015 at 9:15 PM, Juan Hernández
> <jhernand at redhat.com <mailto:jhernand at redhat.com>> wrote:
>
> On 11/24/2015 01:40 PM, John Hunter wrote:
> > Hi,
> >
> > On Tue, Nov 24, 2015 at 5:18 PM, Oved Ourfali <oourfali at redhat.com <mailto:oourfali at redhat.com>
> > <mailto:oourfali at redhat.com <mailto:oourfali at redhat.com>>> wrote:
> >
> > Hi
> >
> > I discussed it with Juan (cc-ed).
> >
> > There used to be a bug in the JDBC authenticion extension that
> > artificially delayed RESTAPI responses by 5 seconds:
> >
> > brute force prevention login delay should not be applied to successful
> > login requests
> > https://bugzilla.redhat.com/1255814
> >
> > That matches the description of the issue, but in theory it has been
> > fixed. I would suggest him to check that he is using the right version
> > of the extension.
> >
> > I did not use the extension ovirt-engine-extension-aaa-jdbc, and I don't
> > think this bug matches my problem, because even there is only one line
> > in the python script, it still cost like 3 seconds, I don't think this is a
> > reasonable time as when I import other package, it cost almost no time.
> >
> > Can you explain why this import line costs so much time?
> >
>
> If you are using 3.6 then you are using
> ovirt-engine-extension-aaa-jdbc,
> as it is enabled by default, but looks like it isn't
> related to your
> problem.
>
> That line takes a long time to execute because it has to
> process two
> large Python modules: the "params" module that contains
> a class per each
> type used by the API (393 classes) and the "brokers"
> module that
> contains a class per each resource used by the API (358
> classes). That
> makes a total of 751 classes. In my environment it takes
> 0.9 seconds,
> approx. You may want to use the python profile in your
> environment and
> share the results:
>
> $ cat > profile.py <<.
> import cProfile
> cProfile.run("from ovirtsdk.api import API")
> .
>
> $ python profile.py
>
> I won't be surprised to see this taking those 3 seconds
> in a slower
> environment.
>
> But even if this takes those 3 seconds it shouldn't be a
> big problem,
> because you shouldn't be running that "from ... import
> ..." line
> frequently. Your program should do this once only, when
> it starts.
>
> Assume that I have two functions to implement, one is to
> list all the vms belong
> to the user, and the other is to retrieve one vm's
> virt-viewer connection file, as
> far as I can see, I have to write two python scripts and
> import the ovirtsdk.api in both
> scripts, each script has to take the 3 seconds :(
>
>
> No, you have two functions, which can be in the same module, or
> in different modules, depending on how you want to organize your
> code.
>
> Python imports are done only once, on the first time you import
> a module anywhere. The module is stored in sys.modules, and the
> next import fetch the module object from sys.modules.
>
> Can you share a profile of a simple script importing the Python sdk?
>
> Do:
>
> python -m cProfile -o myscript.prof myscript.py
>
> The profile will be stored in myscript.prof.
>
> For viewing the profile, I recommend to use the profile-stats tool:
> https://github.com/oVirt/vdsm/blob/master/contrib/profile-stats
>
>
>
> How can I run the "from ... import ..." just once ?
>
>
> >
> > In addition we also know that retrieving large lists of objects with the
> > SDK is slow:
> >
> > [RFE][performance] - generate large scale list running to slow.
> > https://bugzilla.redhat.com/1221238
> >
> > We don't have a solution for that yet.
> >
> > CC-ing Juan in case you have additional questions.
> >
> >
> > On Mon, Nov 23, 2015 at 11:27 AM, John Hunter <zhjwpku at gmail.com <mailto:zhjwpku at gmail.com>
> > <mailto:zhjwpku at gmail.com <mailto:zhjwpku at gmail.com>>> wrote:
> >
> > Hi guys,
> >
> > I am using the ovirt-engine-sdk-python to communicate with the
> > ovirt-engine,
> > I am ok to list the vms but the processing time is too long,
> > like 4.5 seconds,
> > and this line:
> > from ovirtsdk.api import API
> > take almost 3 seconds.
> >
> > This seems a little bit longer than I expected it to be, so I am
> > asking is there
> > a quicker way to communicate with the ovirt-engine?
> >
>
--
Dirección Comercial: C/Jose Bardasano Baos, 9, Edif. Gorbea 3, planta
3ºD, 28016 Madrid, Spain
Inscrita en el Reg. Mercantil de Madrid – C.I.F. B82657941 - Red Hat S.L.
More information about the Devel
mailing list