Embedded Systems Lecture 11: Worst-Case Execution Time. Björn Franke University of Edinburgh

Transcription

1 Embedded Systems Lecture 11: Worst-Case Execution Time Björn Franke University of Edinburgh

2 Overview Motivation Worst-Case Execution Time Analysis Types of Execution Times Measuring vs. Analysing Flow Analysis Low-Level Analysis Calculation

3 Motivation: Characteristics of Real-Time Systems Concurrent control of separate system components Reactive behaviour Guaranteed response times Interaction with special purpose hardware Maintenance usually difficult Harsh environment Constrained resources Often cross-development Large and complex Often have to be extremely dependable

4 What is the Execution Time of a program? WCET: Worst-Case Execution Time BCET: Best-Case Execution Time ACET: Average-Case Execution Time [Wilhelm+08] The WCET/BCET is the longest/shortest execution time possible for a program. Must consider all possible inputs including perhaps inputs that violate specification.

5 Why may we care about the WCET? We are interested in WCET to... perform schedulability anaylsis ensure meeting deadlines assess resource needs for real-time systems WCET accuracy may be safety-critical!

6 And why may we care about the BCET? We are interested in BCET to... benchmark hardware assess code quality assess resource needs for non/soft real-time systems ensure meeting livelines (new starting points)

7 What is the Execution Time of a program? safe BCET estimates actual BCET possible execution times actual WCET safe WCET estimates 0 tighter tighter time [EngblomES01] Approaches for approximating WCET or BCET Measuring: Measure run time of program on target hardware Analysis: Compute estimate of run time, based on program analysis and model of target hardware Hybrid: Combine measurements with program analysis

8 Measuring WCET/BCET Execution time may depend on program inputs In this case, quality of measurements depends on judicious choice of inputs Execution time may depend on execution context (cache content, state of pipeline,...) Typically need to add safety margin to best/worst result measured Extensive testing/measurement still common practice

9 Measuring Program Run Times Call OS timing routines Account for cost of calls to timing routines themselves Access hardware timers directly Use external hardware Oscilloscope, Logic analyser Count emulator cycles High water marking Continuously record max execution times Standard feature of RTOSs May include this in shipped products Read at service intervals

10 Analysing WCET/BCET Instead of measuring execution times, compute them Advantages Can ensure safety of result Saves testing effort Disadvantages Try to be as tight as possible may not always succeed Typically requires extensive analysis effort Accuracy depends on Complexity of hardware Program structure Quality of hardware model Program analysis capabilities

11 Analysing WCET/BCET Program Source Flow Analysis Compiler Global Low- Level Analysis Calculation WCET Object Code Local Low- Level Analysis [EngblomES01]

12 Flow Analysis Analyse dynamic behaviour of program Number of loop iterations, Recursion depth, Input dependences, Infeasible paths, Function instances,... Get information from Static Analysis Manual Annotation Analysis level Object code Source code (may need non-trivial mapping to object code)

13 Flow Analysis Structurally possible flows (infinite) Basic finiteness Statically allowed Actual feasible paths WCET found here=desired result WCET found here=overestimation [EngblomES01]

14 Flow Analysis The set of structurally possible flows for a program, i.e. those given by the structure of the program, is usually infinite, since e.g. loops can be taken an arbitrary number of times The executions are made finite by bounding all loops with some upper limit on the number of executions (basic finiteness) Adding even more information, e. g. about the input data, allows the set of executions to be narrowed down further, to a set of statically allowed paths. This is the optimal outcome of the flow analysis.

15 Flow Analysis const int max = 100; foo (float x) { A: for(i = 1; i <= max; i++) { B: if (x > 5) C: x = x * 2; else D: x = x + 2; E: if (x < 0) F: b[i] = a[i]; G: bar (i) }} Loop bounds: Easy to find in this example; in general, very difficult to determine Infeasible paths: Can we exclude a path, based on data analysis? A-B-C-E-F-G is infeasible since if x>5, it is not possible that x * 2 < 0. Well, really? What about integer overflows? Must be sure that these do not happen in the example...

16 Low-Level Analysis Program Source Flow Analysis Compiler Global Low- Level Analysis Calculation WCET Object Code Local Low- Level Analysis [EngblomES01] Determine execution time for program parts Account for hardware effects (pipeline, caches...) Work on object code Exact analysis generally not possible

17 Low-Level Analysis Global Low-Level Analysis Considers execution time effects of machine features that reach across entire program Instruction/data caches, branch predictors, translation lookaside buffers (TLBs) Local Low-Level Analysis Considers machine features that affect single instruction & its neighbours Scalar/superscalar pipelines

18 Worst-Case Execution Time Analysis Flow Analysis Low-Level Analysis Calculation Local low-level analysis Pipelining Local Low-Level Analysis - Pipelining!!"# % 1235"# 1245"$!$#%()7898!)%).)-+!/0, 12345"6!"#%()*+),-)%).)-+!/0,!!""%&'!""$ &!'#%:0;)</,=%!>)%!/?/,=%+@/,=!/?)@%8,;%!/?/,=%)AA)-!@ [EngblomJ02] Fig. 1. Pipelining of instruction execution effect of two successive Pipeline Pipelining effect of two successive instructionsinstructions Pipeline overlap reduces overall computation time!!"b by%δ = 2 Pipeline overlap time by δ!="%&' "B reduces overall 1245"#computation 12D5"B!"!""# & Embedded Real-Time Systems!"#$%&'$()"*+ #$,--#.*/,($("0 1)%*2*$)*-3,43 $53*6*#$, "C SS 2010, Lecture D5"C!"#"%&'!!"# Types of Execution Times Slide 18!#"B (!!"#"%G6

19 $53*6*#$,03 Worst-Case Execution Time 1235"# 1234D5"EF 1245"$ 12345"6!!"# Types of Execution Times! "%&'!""$ Measuring vs.!" % Analyzing Flow Analysis & Low-Level Analysis!"#%1/?/,=%?0;)< Calculation Local Low-Level Analysis - Pipelining Local low-level analysis Pipelining Fig. 2.!$#%()7898!)%).)-+!/0,!'#%:0;)</,=%!>)%!/?/,=%+@/,= Example long timing effect!"#%()*+),-)%).)-+!/0,!/?)@%8,;%!/?/,=%)aa)-!@ Fig. 1. Pipelining of instruction execution I1... Im is different from the execution of the nodes I2... Im starting with node I2 (see Section 4.1 for more details). %&'()*+,--)./ %&'()*+'01)./ ("!""#$%"&'(!!"B %! ( $A the execution profile!;3 of BC is; differ- "&,-.)/*00 Figure 2 shows an example of a LTE, where $5BC000D$(!;= 1235"B 1245"# 12D5"B "%&' sequence!!"the ent when executed in isolation and when executed as part of ABC.!= Note $$(" $ $$$A diagrams, similar to the reservation that we illustrate pipeline execution using pipeline!=6!=< $$$5BC00D$$!""# & tables commonly used to describe the behavior$$$$000$ of pipelined processors. Time runs hor#!6!!< #!"#"%&' $$$-!E!3 cycle. The pipeline!""#$%"&'( izontally, with each 12345"C tick of time corresponding to a processor clock!!"# "%G6! 124D5"C!6> $$$$000$! <>!"#$%&'$()"*+on the vertical axis. Instructions progress from 5''-.)/2)!!#"B ( stages are shown upper left to lower * $$$5BC000D$ " "!?= #$,--#.*/,($("0 5''-.)/800 $$$$000$ & 1)%*2*$)*-3,43!>7!>: right, and each cycle of execution is shown as a square. $53*6*#$,03 5''-.)/*00 $$$-!E! & 7 %!: LTEs can in general be both negative and$$$$000$ positive, %and must be accounted for by $$$F!"#%1/?/,=%?0;)<!!:? 1234D5"EF!$#%H.)-+!/0,%!/?)@ a static WCET analysis method that wants to $$GH5!-C00D be safe and tight. Positive 7?timing effects ' ' [EngblomJ02] an underfig. and 2. Example long$-!etiming add execution time to a program, are critical to effect consider since otherwise!?@ $$$000$*! Pipelining of three instructions 3@ estimate effect of the WCET couldsuccessive result. Negative timing effects indicate potential savings!@;in $F ) ) Pipelining effect ofthem three instructions execution ignoring onlysuccessive makes the WCET I1 time,... Im and is different from the execution of thegh5!-c000d$ nodes I2... Iestimate with tight. node I2 m starting less (see Section 4.1 for more details). "#$%&'()'#*%+(,".$%/+(0-%)'#01%2341%#44#+1-,%56(2%5# Positive LTEs comprise a central three problem for WCET analysis, since theythan make it Reduction of combining instructions can be larger Reduction offigure combining can be 2: larger sum Attached of savingsflow F 2 shows an three example instructions of a LTE, where the execution profile ofthan BC iswith differfigure Scopes necessary to consider effects across more than just adjacent instructions assumption in sum of savings when combining them pair-wise! ent when executed in isolation and when executed as part of the sequence ABC. Note when combining them pair-wise! order to that create a safe analysis. To prove a pipeline particular WCET analysis method safe, it is we illustrate pipeline execution using diagrams, similar to the reservation The target chip fortime theruns present i Embedded Real-Time SS 2010, Lecture 15 have Slide necessary to address theused question ofsystems whether all LTEs been accounted for. tables commonly to describe the behavior ofpositive pipelined processors. hor- 19 implementation V850E, a typical 32-bit embedded tick of corresponding tosufficient a processor clock cycle. The pipeline A centralizontally, issue inwith thiseach paper is time therefore to give conditions forrisc when LTEs are microcon stages are shown on the vertical axis. Instructions progress[6]. fromthe uppercompiler left to lowerused is an IAR V8 chitecture guaranteed not to be positive. right, and each cycle of execution is shown as a square. C/Embedded C++ compiler [32].

20 Global Low-Level Analysis - Caches Instruction Caches Predictable from control flow Data Caches No simple way to predict accesses Very difficult analysis problem Unified Caches Very pessimistic as a result of combining instructions & data

21 Worst-Case Execution Time Analysis Measuring vs. Analyzing Flow Analysis Low-Level Analysis Calculation Global low-level analysis Caches Global Low-Level Analysis - Caches D#73+%.6(+E / '()$*+ 0012)!3!" )2)& +!"#"$%N(($.*O.4!"##$%!($$PQRO.S, )2)& )2)&. '()$4+ )!6 #!7! " 00*+2)!6!$ 0092)!6!!: $ $ $ $ $ $ )2)& - )2)& 4K899 "KJ8=> A!-+B43(C 3C5('*#43(C!"#&$%TU#$.*O.S!"#'$%%V-$PER!"#"$%()*)+,%-(..!"##$%()*)+,%+(/0 %%%%%%%%%%1-,-%2345!"#&$%()*)+,%+(/!"#'$%()*)+,%+(/!"#"$%&%'()*&+",*&-.-/0,"1$&"$21'#(,"1$ / + 4J89; "JL8=@ "KI8=?,. "LM8=9-4I89: "IL8=9 4L8; 4M89< &32,-'&)")-4"$-&($(456"6 Figure 3: Timing Graph with Execution [StappertEE01] Facts May split loops to loops differentiate between first and successive loop iterations May split to differentiate between first and successive loop iterations / / 9,60:41'. "" )2)& )2)& %5#+47 Must combine with pipelining effects effects Must combine with pipelining + Facts n is the NEC ontroller arv850/v850e )2)& + 9,60:41'. / 4!"" K "KJ!*?@ "; << 15 / + Systems 9,60:41'. )2)& Embedded Real-Time SS 2010, Lecture 3.456)-1*'7 1,6,-5*5.4(8 )2)& )2)& 9,60:41,'-*.0-% + 4!"; J =,6,-5*5.4(8 >,18*1,6)% Figure 4: Timing Effect Calculation Slide 21

22 WCET Calculation Program Source Flow Analysis Compiler Global Low- Level Analysis Calculation WCET Object Code Local Low- Level Analysis [EngblomES01] Task: Find the path that results in the longest execution time Several approaches in use Properties of approaches Program flow allowed Object code structure (optimisations?) Pipeline effect modelling Solution complexity

23 WCET Calculation Path-based Constraint-based Implicit Path Enumeration Technique - IPET Structure-based

24 WCET Calculation [Wilhelm+08]

25 Path-Based Bound Calculation Upper bound for a task is determined by computing bounds for different paths in the task, searching for the overall path with the longest execution time. Defining feature is that possible execution paths are represented explicitly. Natural within a single loop iteration, but problems with flow information extending across loop nesting levels. Number of paths is exponential in the number of branch points. Possibly requiring heuristic search methods.

26 Implicit Path Enumeration Program flow and basic block execution time bounds are combined into sets of arithmetic constraints. Each basic block and program flow edge in the task is given a time coefficient, expressing the upper bound of the contribution of that entity to the total execution time every time it is executed.

27 Structure-based Bound Calculation Upper bound is calculated in a bottom-up traversal of the syntax tree of the task combining bounds computed for constituents of statements according to combination rules for that type of statement. Not every control flow can be expressed through the syntax tree Assumes straight-forward correspondence between source structures and the target program Not easily admitting code optimisations In general, not possible to add additional flow information (as in IPET).

28 Summary Motivation Worst-Case Execution Time Analysis Types of Execution Times Measuring vs. Analysing Flow Analysis Low-Level Analysis WCET Calculation

29 Preview Real-Time Operating Systems MQX