Performance Analysis, Scheduling and Synthesis of Embedded Systems - PowerPoint PPT Presentation

Performance Analysis, Scheduling and Synthesis of Embedded Systems Kim G. Larsen CISS – Aalborg University DENMARK

CISS in Numbers � National ICT Comptetence Center 2002: MDKK Ministry 31,5 MDKK North Jutland 8,5 MDKK Aalborg City 7,5 16,00 MDKK Companies 16,00 MDKK AAU � 45 projects � 20 CISS employees � 25 CISS associated researcher at 3 different research groups at AAU! � >20 industrial PhDs Kim L Ki Larsen [ [2] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

European Projects � ARTIST2 (NoE FP6) � coordinator for Testing & Verification Cluster � ARTIST Design (NoE FP7) � kick-off meeting end of January � co-coordinator of Modeling and Validation w Tom Henzinger) � Other new STREPs (FP7) Quasimodo, Multiform Kim L Ki Larsen [ [3] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Motivation – MPSoC CELL processor Ki Kim L Larsen [ [4] CDC Final Workshop CDC Final Workshop, Tallinn, J , Tallinn, Jan, 2008 n, 2008

Scheduling… in ES Tasks: Resources Computation times Deadlines Execution platform Dependencies PE, Memory Scheduling Principles (OS) Arrival patterns Networks EDF, FPS, RMS, DVS, .. uncertainties Drivers uncertainties Kim L Ki Larsen [ [5] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Issues � Performance Evaluation � Estimate resources (e.g. energy) required by given SP. � Scheduling & Synthesis Tasks Res. � Synthesize (optimal) SP ensuring given objective. SP � Scheduling : SP controls everything (including ex.time). � Synthesis : scheduling under � Schedulability Analysis uncertainties (e.g. execution � Verify that given SP time, availability of ensures deadlines. resources). Kim L Ki Larsen [ [6] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Approach – TA � Performance Evaluation � Estimate resources (e.g. energy) required by given CLASSI C CLASSI C CLASSI C SP. � Scheduling & Synthesis CORA CORA CORA Tasks Res. � Synthesize (optimal) SP ensuring given objective. TI GA TI GA SP TI GA � Scheduling : SP controls everything (including TALK: TALK TALK: ex.time). What can we do? What can we do? � Synthesis : scheduling under � Schedulability Analysis What can we do efficiently? What can we do efficiently? uncertainties (e.g. execution � Verify that given SP What can not be done? time, availability of What can not be done? ensures deadlines. What would we like to do? resources). What would we like to do? Kim L Ki Larsen [ [7] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

The UPPAAL Team @ AALborg @UPPsala � Kim G Larsen � Wang Yi � Gerd Behrman � Paul Pettersson � Arne Skou � John Håkansson � Brian Nielsen � Anders Hessel � Alexandre David � Pavel Krcal � Jacob I. Rasmussen � Leonid Mokrushin � Marius Mikucionis � Shi Xiaochun � Thomas Chatain @Elsewhere � Emmanuel Fleury, Didier Lime, Johan Bengtsson, Fredrik Larsson, Kåre J Kristoffersen, Tobias Amnell, Thomas Hune, Oliver Möller, Elena Fersman, Carsten Weise, David Griffioen, Ansgar Fehnker, Frits Vandraager, Theo Ruys, Pedro D’Argenio, J-P Katoen, Jan Tretmans, Judi Romijn, Ed Brinksma, Martijn Hendriks, Klaus Havelund, Franck Cassez, Magnus Lindahl, Francois Laroussinie, Patricia Bouyer, Augusto Burgueno, H. Bowmann, D. Latella, M. Massink, G. Faconti, Kristina Lundqvist, Lars Asplund, Justin Pearson... Kim L Ki Larsen [ [8] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

“Impact” UPPAAL downloads (total) Google: 1200000 1000000 UPPAAL: 134.000 SPIN Verifier: 242.000 800000 nuSMV: 57.700 600000 # > 2.900 400000 Google Scholar Citations 200000 (Rhapsody/Esterel < 5.000) 0 9907 0001 0007 0101 0107 0201 0207 0301 0307 0401 0407 0501 0507 0601 0607 0701 0707 Date Kim L Ki Larsen [ [9] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Timed Automata [ Alur & Dill’89] Resource Synchronization Guard Reset Sem antics: Invariant ( Idle , x= 0 ) � ( Idle , x= 2.5) d(2.5) � ( InUse , x= 0 ) use? � ( InUse , x= 5) d(5) � ( Idle , x= 5) done! � ( Idle , x= 8) d(3) � ( InUse , x= 0 ) use? Kim L Ki Larsen [ [10] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Composition Task Resource Synchronization Shared variable Sem antics: ( Idle , Init , B= 0, x= 0) � ( Idle , Init , B= 0 , x= 3.1415 ) d(3.1415) � ( InUse , Using , B= 6, x= 0 ) use � ( InUse , Using , B= 6, x= 6 ) d(6) � ( Idle , Done , B= 6 , x= 6 ) done Kim L Ki Larsen [ [11] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Task Graph Scheduling Optimal Static Task Scheduling P 2 P 1 � Task P ={P 1 ,.., P m } 2 ,3 � Machines M ={M 1 ,..,M n } Duration Δ : (P × M) → N ∞ � 1 6 ,1 0 � < : p.o. on P (pred.) 6 ,6 1 0 ,1 6 P 6 P 3 P 4 2 ,3 � A task can be executed only if all predecessors have completed � Each machine can process at most one task at a time P 7 P 5 � 2 ,2 Task cannot be preempted. 8 ,2 M = { M 1 ,M 2 } � Compute schedule with minimum completion-time! Kim L Ki Larsen [ [12] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Task Graph Scheduling Optimal Static Task Scheduling � P 2 P 1 Task P ={P 1 ,.., P m } 2 ,3 � 1 6 ,1 0 Machines M ={M 1 ,..,M n } Duration Δ : (P × M) → N ∞ � � < : p.o. on P (pred.) 6 ,6 1 0 ,1 6 P 6 P 3 P 4 2 ,3 P 7 P 5 2 ,2 8 ,2 E<> (Task1.End and … and Task7.End) M = { M 1 ,M 2 } Kim L Ki Larsen [ [13] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Experimental Results Symbolic A* Brand-&-Bound 60 sec Abdeddaïm, Kerbaa, Maler Ki Kim L Larsen [ [14] CDC Final Workshop CDC Final Workshop, Tallinn, J , Tallinn, Jan, 2008 n, 2008

Optimal Task Graph Scheduling Power-Optimality P 2 P 1 2 ,3 � Energy-rates : 1 6 ,1 0 C : M → N � Compute schedule with minimum completion-cost! 6 ,6 1 0 ,1 6 P 6 P 3 P 4 2 ,3 P 7 P 5 2 ,2 8 ,2 4 W 3 W Kim L Ki Larsen [ [15] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Priced Timed Automata Behrmann, Fehnker, et all (HSCC’01) Timed Automata + COST variable Alur, Torre, Pappas (HSCC’01) l 2 l 1 l 3 x · 2 3 · y 0 · y · 4 ☺ c’ = 4 c’ = 2 x: = 0 c + = 4 cost rate c + = 1 cost update y · 4 x: = 0 Kim L Ki Larsen [ [16] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Priced Timed Automata Behrmann, Fehnker, et all (HSCC’01) Timed Automata + COST variable Alur, Torre, Pappas (HSCC’01) l 2 l 1 l 3 x · 2 3 · y 0 · y · 4 ☺ c’ = 4 c’ = 2 x: = 0 c + = 4 cost rate c + = 1 cost update y · 4 x: = 0 TRACES ε (3) ( l 1 ,x= y= 0) ( l 1 ,x= y= 3) ( l 2 ,x= 0,y= 3) ( l 3 ,_,_) ∑ c= 1 7 12 1 4 Ki Kim L Larsen [ [17] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Priced Timed Automata Behrmann, Fehnker, et all (HSCC’01) Timed Automata + COST variable Alur, Torre, Pappas (HSCC’01) l 2 l 1 l 3 x · 2 3 · y 0 · y · 4 ☺ c’ = 4 c’ = 2 x: = 0 c + = 4 : m e : l b m cost rate o e r l P b c + = 1 o r P t s o c t s ) m o c u m ) m i x u a cost update m m i y · 4 x ( x: = 0 a m m u ( m m i n u i m m i e n TRACES h i m t d e n h i t F l d n n i o F l i t a n c o o i 3 t l a g c ε (3) n o 3 i h l g c n a e i h r c f a ( l 1 ,x= y= 0) ( l 1 ,x= y= 3) ( l 2 ,x= 0,y= 3) ( l 3 ,_,_) o e r f o ∑ c= 1 7 12 1 4 Efficient Implementation: ε (2.5) Efficient Implementation: ε (.5) CAV’0 1 and TACAS’0 4 ( l 1 ,x= y= 0) ( l 1 ,x= y= 2.5) ( l 2 ,x= 0,y= 2.5) ( l 2 ,x= 0.5,y= 3) ( l 3 ,_,_) CAV’0 1 and TACAS’0 4 10 1 1 4 ∑ c= 1 6 ε (3) ( l 1 ,x= y= 0) ( l 2 ,x= 0,y= 0) ( l 2 ,x= 3,y= 3) ( l 2 ,x= 0,y= 3) ( l 3 ,_,_) Competitive with MILP 1 6 0 4 ∑ c= 1 1 and commercial tool (Axxon) Kim L Ki Larsen [ [18] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008

Optimal Infinite Scheduling Maximize throughput: i.e. maximize Reward / Time in the long run! Ki Kim L Larsen [ [19] CDC Final Workshop CDC Final Workshop, Tallinn, J , Tallinn, Jan, 2008 n, 2008

Optimal Infinite Scheduling Minimize Energy Consumption: i.e. minimize Cost / Time in the long run Ki Kim L Larsen [ [20] CDC Final Workshop CDC Final Workshop, Tallinn, J , Tallinn, Jan, 2008 n, 2008

Optimal Infinite Scheduling Maximize throughput: i.e. maximize Reward / Cost in the long run Ki Kim L Larsen [ [21] CDC Final Workshop CDC Final Workshop, Tallinn, J , Tallinn, Jan, 2008 n, 2008

Cost Optimal Scheduling = Optimal Infinite Path Accumulated cost c 3 c n c 1 c 2 r 3 r n σ r 1 r 2 Accumulated reward ¬ (Task0.Err or Task1.Err or …) Value of path σ : val( σ ) = lim n →∞ c n /r n Optimal Schedule σ * : val( σ * ) = inf σ val( σ ) Kim L Ki Larsen [ [22] CDC Final Workshop, Tallinn, J CDC Final Workshop , Tallinn, Jan, 2008 n, 2008