Hi,

So after the a long weekend, I'm back with my results.
For the write, findByPK and findAll tests I get now good numbers.
See:
https://github.com/johannesschaefer/javaee_jsf_cdi_jpa_data_ds_project_template/blob/master/src/test/java/de/psi/metals/futurelab/repo/benchmark/SaveTest.java
https://github.com/johannesschaefer/javaee_jsf_cdi_jpa_data_ds_project_template/blob/master/src/test/java/de/psi/metals/futurelab/repo/benchmark/ReadTest.java
https://github.com/johannesschaefer/javaee_jsf_cdi_jpa_data_ds_project_template/blob/master/src/test/java/de/psi/metals/futurelab/repo/benchmark/ReadAllTest.java

The difference between delta spike and plain EM are just a few percent, in both 
directions ;-) .

But I wrote a new test case were I try to find entities by an query. 
https://github.com/johannesschaefer/javaee_jsf_cdi_jpa_data_ds_project_template/blob/master/src/test/java/de/psi/metals/futurelab/repo/benchmark/ReadQueryTest.java
So I compare
            TypedQuery< Material > query = eml.createQuery(
                "SELECT m FROM Material m WHERE grade = :grade AND width = 
:width AND thickness = :thickness",
                Material.class );
            query.setParameter( "grade", "AAA" );
            query.setParameter( "width", 5 );
            query.setParameter( "thickness", 5. );
List< Material > mats = query.getResultList();

to 
List< Material > mats = matRepo.findByGradeAndWidthAndThickness( "AAA", 5, 5. );

Here again the difference is quite high.
|   | iter 10    | iter 20    | iter 40    | iter 80    | iter 160   | iter 320 
  | iter 640   | iter 1280  | iter 2560  | iter 5120   | iter 10240  |
|====================================================================================================================================================|
| DS| 0.03988012 | 0.151870613| 0.144881044| 0.270389952| 0.526700787| 
1.023574545| 1.806960223| 3.426772405| 6.969935385| 13.963582543| 26.785764953|
| EM| 0.010984804| 0.021940339| 0.059921297| 0.087386918| 0.171045079| 
0.375059016| 0.747171594| 1.560946145| 2.968347174| 6.446844753 | 12.361550486|

So as you can see the DeltaSpike implementation needs at least the double 
amount of time.

Any hints for improving the performance?

Regards,
Johannes

-----Original Message-----
From: Schäfer, Johannes [mailto:[email protected]] 
Sent: Thursday, June 1, 2017 2:27 PM
To: [email protected]
Subject: RE: Performance of DeltaSpike Data

Right. Copy and paste error.
I added also a flush to the EM test.
Now I have similar numbers.
______________________________________________________________________________________________________________________________________________________
|   | iter 10    | iter 20    | iter 40    | iter 80    | iter 160   | iter 320 
  | iter 640   | iter 1280  | iter 2560  | iter 5120   | iter 10240   |
|=======================================================================
|=======================================================================
|=======|
| DS| 0.001588214| 0.004130191| 0.007351854| 0.014062036| 0.048373222| 
| DS| 0.593463008| 0.741351593| 1.697058004| 6.049719288| 34.101109279| 
| DS| 101.589077365|
| EM| 0.001385601| 0.002662861| 0.004092937| 0.108730649| 0.046299193| 
| EM| 0.106900289| 0.461147505| 1.688040769| 5.960683928| 25.604583163| 
| EM| 106.688041149|

It's a little bit strange for me, why I have to compare 
EntityPersistenceRepository.save with a em.persist + em.flush. I would expect 
that an simple EntityPersistenceRepository.save don't have a flush (there is a 
separate EntityPersistenceRepository.saveAndFlush). 

When I do a run with EntityPersistenceRepository.saveAndFlush I get the 
following numbers.
______________________________________________________________________________________________________________________________________________________
|   | iter 10    | iter 20    | iter 40    | iter 80    | iter 160   | iter 320 
  | iter 640   | iter 1280  | iter 2560  | iter 5120   | iter 10240   |
|=======================================================================
|=======================================================================
|=======|
| DS| 0.001703015| 0.003457728| 0.008079817| 0.019099994| 0.053865065| 
| DS| 0.940319597| 0.643245399| 2.292716685| 9.902395587| 40.84301017 | 
| DS| 172.179435413|
| EM| 0.001677545| 0.004168205| 0.005779986| 0.014491211| 0.031066334| 
| EM| 0.110747277| 0.4051742  | 1.925682412| 5.842606084| 23.540393571| 
| EM| 132.817886521|

So I have the feeling that there is still something wrong.

Thanks to Gerhard for his additional hints.
I committed all changes to the github repo.

Regards,
Johannes

-----Original Message-----
From: Gerhard Petracek [mailto:[email protected]]
Sent: Thursday, June 1, 2017 1:21 PM
To: [email protected]
Subject: Re: Performance of DeltaSpike Data

@johannes:
as mentioned yesterday you have to move EntityTransaction#begin and 
EntityTransaction#commit into the for-loop.

regards,
gerhard



2017-06-01 12:58 GMT+02:00 Thomas Andraschko <[email protected]>:

> Hi,
>
> ~1 year ago i did many optimizations in the data module and AFAIR DS 
> Data was only a little bit slower.
> After i compared my testcase with a benchmark from a user, the big 
> different came from the transaction handling which was different in 
> both testcases.
>
> Regards,
> Thomas
>
> 2017-06-01 12:33 GMT+02:00 Gerhard Petracek <[email protected]>:
>
> > hi johannes,
> >
> > after refactoring your initial code to ds-test-control i saw e.g. 
> > ~6s vs 7,5s for 2560 iterations.
> > i'll compare my local version with your new version (mentioned in 
> > your mail).
> >
> > regards,
> > gerhard
> >
> >
> >
> > 2017-06-01 11:35 GMT+02:00 Schäfer, Johannes <[email protected]>:
> >
> > > Hi,
> > >
> > > My company is thinking about using DeltaSpike Data. But before we 
> > > integrate this into our development I was asked to prepare some
> > benchmarks,
> > > comparing the usage of DeltaSpike Data with the usage of a plain 
> > > EntityManager.
> > > I prepared some benchmarks and I was surprised that there is a big 
> > > difference in the write performance. I already got some hints in 
> > > the
> > delta
> > > spike irc channel, but the performance is still bad.
> > > Based on a template from os890 I implemented my tests and prepared 
> > > a github project.
> > > https://github.com/johannesschaefer/javaee_jsf_
> cdi_jpa_data_ds_project_
> > > template
> > > Basically I'm talking about this test:
> > > https://github.com/johannesschaefer/javaee_jsf_
> cdi_jpa_data_ds_project_
> > > template/blob/master/src/test/java/de/psi/metals/futurelab/
> > > repo/benchmark/SaveTest.java
> > >
> > > It just saves an entity into a DB in a loop. Depending of the 
> > > number of iterations the difference is quite big.
> > >
> > > SaveTest
> > > ____________________________________________________________
> > > ____________________________________________________________
> > > _____________________________
> > > |   | iter 10    | iter 20    | iter 40    | iter 80    | iter 160   |
> > > iter 320   | iter 640   | iter 1280  | iter 2560  | iter 5120   | iter
> > > 10240  |
> > > |===========================================================
> > > ============================================================
> > > =============================|
> > > | DS| 0.004911746| 0.03597043 | 0.015765787| 0.016966639| 
> > > | DS| 0.043319612|
> > > 0.281807839| 1.308948835| 1.370535533| 8.186996818| 20.920141274| 
> > > 93.631768475|
> > > | EM| 0.004557839| 0.003256631| 0.005775416| 0.004834958| 
> > > | EM| 0.028243393|
> > > 0.035484616| 0.038600595| 0.088904458| 0.339158674| 0.745850523 |
> > > 0.853543234 |
> > >
> > > Also the difference between a run with 5120 and 10240 iteration is 
> > > not just the double amount of time, it is more than 4 times more.
> > >
> > > Do you have some hints to me what I'm doing wrong there?
> > >
> > > Regards
> > > Johannes
> > >
> > >
> > >
> >
>

Reply via email to