[e-dev-day 2015][4/4] update on efl performance benchmarking (cedric bail)

1Samsung Open Source Group

Cedric BailSenior Open Source EngineerSamsung Open Source Group

[email protected]

Update on EFL performance

mailto:[email protected]


The plan

● Benchmark memory usage, frame rate and time to first frame

● One week should be enough, right ?

● Use usual tool and experiment with a few new one

– Expedite

– Massif

– Smem

● Walk in a park !


:”(

● Not quite…

● Find memory leak

● Find massive memory usage increase

● Find massive performance regression

● Barely enough time to fix them !


Easiest one, the one I create !

● Massive leak of Cairo context in expedite

● 132MB of leaked data at the end of the test

● Kind of make any other peak impossible to see !

● That one is now fixed


Easiest one, the one I create !

1.8 1.9 1.10 1.11 1.12 1.13 1.14 1.15 1.160

2

4

6

8

10

12

14

Peak


Weird increase in PSS

● Write a script to mesure PSS

● Gathered the data

● Look at it in the plane

● 8-O

● Need to review the script first …

● That's not a fixed problem !


Weird increase in PSS

1.8 1.9 1.10 1.11 1.12 1.13 1.14 1.15 1.160

10000

20000

30000

40000

50000

60000

70000

80000

90000


Run expedite

● Instead of the usual raw benchmark, try to find the reliable test

● Run all expedite test twice on every version since 1.8

● That's 18 runs ! More than 2000 tests…

● Start process them in the plane…

● 8-O

● See major slow down !


Run expedite

1.8 1.9 1.10 1.11 1.12 1.13 1.14 1.15 1.160

200

400

600

800

1000

1200

Image Blend Solid Middle Border

Image Blend Solid Border

Rect Solid


Run expedite

● Let's blame Tom, because he is not here ! :-) Is he ?

● Then run expedite under callgrind with 1.8 and git

● Issue seems to be related to raster change in region_add

– 1.8: 40% of the benchmark, while drawing was 20%

– 1.16: 80% of the benchmark, while drawing is 5%

● Going to be fun to fix...


So what about the rest ?

● No time for time to first frame benchmarking !

● No time for power consumption benchmarking !

● Memory testing still need fixing


What did I learn

● Automating expedite benchmark is doable if we focus only on a subset

● Expedite does catch regression if we use it…

● How to integrate it with Jenkins and our infrastructure ?

● Maybe put more logic in it to benchmark memory by processing /proc/$PID/maps

● Still no test of elementary/edje at all

● If we don't see the issue when we have tools, then what about when we don't !


Opinions ? Questions ?

Thank you.


[e-dev-day 2015][4/4] update on efl performance benchmarking (cedric bail)

Software