g
Green Mining and the Perils of Mining Energy Profjles Abram Hindle (hindle1@ualberta.ca) Department of Computing Science University of Alberta http://softwareprocess.es/ with Karan Aggarawal*, Candy Pang*, Chenlei Zhang*, Kent Rasmussen*, Joshua Campbell*, Alex Wilson*, Jed Barlow*, Stephen Romansky*, w/ Eleni Stroulia, Bram Adams, Ahmed Hassan, Daniel German
energy
Programmers are responsible for Software Power Use!
In general, programmers lack knowledge about software energy consumption, but they are more knowledgeable about software energy consumption on mobile devices than on desktop computers. http://webdocs.cs.ualberta.ca/~hindle1/2014/green-programmers/ [Pang et al.]
MOV [bufgerloc], EAX
Energy Profjle of Single Versions
Android Firefox Energy Profile 1.4 1.4 1.2 1.2 Mean Watts Mean Watts 1.0 1.0 0.8 0.8 0.6 0.6 518a79e90444 3f0b94325b80 7d4268f8884c bfd85c9652fa fef5f202b2dc 7d4268f8884c fef5f202b2dc 518a79e90444 Android Fennec Revision Software Revision
Android Calculator Energy Profile
Energy Profjle: Same test, Same Device a p p _ 1 _ S u i t e 1 0 0 0 m s g J o u l e s 1 7 0 s e l u o J 1 6 0 1 5 0 B
Energy Profjle: Difgerent device, difgerent response 1 _ S u i t e 1 0 0 0 m s g J o u l e s 1 7 5 s e l u o J 1 5 0 A B C D A p p V e r s i o n s
Error rors Tha That s silver er ba band is t nd is the he manuf ufacturing cturing err error t r toler oleranc nce
Error rors Mult ultiple i le ins nstan ances ces of t f the e same m ame model o odel of ph f phone ne wi will pro ll produce di duce difger fgerent m t meas easurem urements ts du due t e to e o erro rror t r tolera olerances nces i in Man anufa ufactu cturi ring ng! This i impli mplies es a n a nee eed for for nor normali malizat ation n and rep nd repeated t ed tes ests ts!
Power Sup r Supply Ras aspberr rry P y Pi Ardui uino USB H USB Hub/P /PS Phone
Elaborate Job Queue [MSR 2014]
How does it work?
Fake the Battery
Control the USB Connection USB Computer T I P 1 2 7 A r d u i n o 1 k D i g i t a l 4 G N D USB Phone
Timeseries graph and Component Breakdown 12 10 8 Column 1 6 Column 2 Column 3 4 2 0 Row 1 Row 2 Row 3 Row 4
Aggregate Plotting How does my energy profjle change over time?
Aggregate by Component and Version! How does power use difger across scenarios and versions?
Green Miner This colour is what happens When you don't specify UI Look and feel requirements. [MSR 2014]
Per Component Breakdown
Sometimes we need to debug the tests with some context 1
Green Miner is an enabler ● Continuous Integration and Testing for Energy Regressions ● Repeatable, logged, measureable framework. ● Enables asking of questions and repeating experiments. ● Some examples to follow
Does Adblocking Save Power? Adblock Saves power Versus Nothing or very large MO MORE /etc/hosts files [GREENS 2014]
Does Adblocking Save Power? Adblock Adblocking saved Saves power Versus 0.01 to 0.06W Nothing per test or very large (~5J every 100 seconds) MO MORE /etc/hosts files [GREENS 2014]
Which UI Theme would use More Power? Black on White White on Black UI Theme UI Theme
What Causes Software Power Use? [Aggarwal et al.]
What causes Software Power Use? Syscalls matter but Syscalls can always Vary per run of the Same test! [Aggarwal et al.]
What causes Software Power Use? If the system call profjle Syscalls matter but changes signifjcantly from the Syscalls can always previous version, it is Vary per run of the Same test! probable (~50%) that the application’s energy consumption profjle has changed as well. [Aggarwal et al.]
What causes Software Power Use? ● Rule of thumb model: If a cumulative count of a syscall changes signfjcantly between versions there's a good chance of a signfjciant change in power use and energy consumption! [Aggarwal et al.]
Does logging matter? Bufgering hides, delays and defers The cost of IO [Di Nardio et al.]
Does logging matter? 1 . 0 Bufgering hides, delays and defers The cost of IO s t t Wa n 0 . 9 a e M 0 . 8 0 _ Wa i t f o r w a t t l o g1 _ S u i t e 1 0 0 0 m s g 2 _ S u i t e 1 0 0 m s g 3 _ S u i t e 1 0 m s g 4 _ S u i t e 1 m s g 5 _ S u i t e 0 1 m s g 6 _ S u i t e _ 0 0 1 m s g 7 _ I d l e t i m e A p p V e r s i o n s [Di Nardio et al.]
Methodological Bottleneck 1.4 Applied 1.2 Mean Watts 1.0 Exhaustively 0.8 0.6 518a79e90444 3f0b94325b80 7d4268f8884c bfd85c9652fa fef5f202b2dc Android Fennec Revision to all revisions Revisions are measured multiple times for stability
The Research Hurdle: Measuring every revision takes too long!
Pareto frontier: Most cost effective approximations per time spent
Pareto frontier: Most cost effective approximations per time spent Le Less r revision ions t tested Lower fj Lo fjnal a al accu curac acy Mo More r revisions ions t tested Hig igher fj r fjnal ac l accu curacy
Pareto frontier: Most cost effective approximations per time spent Half o lf of t f the he w work rk Sm Small Di ll Difg fgerence i in RMSE n RMSE
Pareto frontier: Most cost effective approximations per time spent Half o lf of t f the he w work rk Sm Small Di ll Difg fgerence i in RMSE n RMSE
Pareto frontier: Most cost effective approximations per time spent A t thir ird o of t f the he w work ork Sm Small Di ll Difg fgerence i in RMSE n RMSE
Pareto frontier: Most cost effective approximations per time spent 1/3 /3 t to 1 1/2 /20th of of t the W Work rk Larger Di r Difg fgerence in in RMSE RMSE
Pareto frontier: Most cost effective approximations per time spent Tim ime & Accu curac acy
Green Mining Needs You Got a question about power? Looking for collaborators. Green Miner is usable remotely.
Conclusions 1 measurement is not enough Energy profjles change Measurement Granularity Software evolves, and forks, -- it's not always now. http://greenmining.softwareprocess.es/ Green Miner is available For collaboration!
Recommend
More recommend