Predictable multithreading of embedded applications using PRET-C

Transcription

1 Predictable multithreading of embedded applications using PRET-C Sidharta Andalam University of Auckland New Zealand Interactive Presentation March 2010 Andalam (UoA) PRET DATE'10 1 / 20

2 Layout 1 Introduction Problems and motivations 2 The Auckland Approach PRET-C based predictable programming Mapping Logical time to Physical Time Results 3 Conclusions and Roadmap Andalam (UoA) PRET DATE'10 2 / 20

3 Real-Time System Design Current design approaches of Time Critical Systems Rely on the well known theory of real-time scheduling. A set of tasks with timing parameters, which execute on RTOS. WCET derived through static analysis. Major issue with WCET analysis: modern processors render WCET virtually unknowable; even simple problems demand heroic eorts [Edwards and Lee, DAC 2007]. Andalam (UoA) PRET DATE'10 3 / 20

4 Why not the theory of real-time systems? Do we have to really rethink timing from scratch? [E. A. Lee, Computing needs time, CACM, May 2009]. Designers try to guarantee worst case timing using the well-known scheduling theory. Implemented on top of speculative architectures and imprecise preemption mechanism called interrupts. A real-time operating system or RTOS is a patched solution that provides RT scheduling on top of a speculative processor. This solution is both non-deterministic and not fail-safe: priority inversion problem has been a nagging issue, for example. Andalam (UoA) PRET DATE'10 4 / 20

5 Our Philosophy Notion of concurrency: concurrency is logical but execution is sequential very similar to synchronous languages [Benveniste'03]. Notion of time: time is logical and the mapping of logical to physical time is done using static analysis of code. Design approach: Auckland Reactive PRET (ARPRET) architectures are designed by simple customisation of soft-core processors. Andalam (UoA) PRET DATE'10 5 / 20

6 Overview of the solution Stages 1 PRET-C: simple synchronous extension to C (using macros). Code void main() { while(1) { abort PAR(sampler,display); when pre (reset); Andalam (UoA) PRET DATE'10 6 / 20

7 Overview of the solution Stages 1 PRET-C: simple synchronous extension to C (using macros). 2 TCCFG : intermediate format. Code void main() { while(1) { abort i=0 out=0.0 sample=0.0 (6) (6) i=0 PAR(sampler,display); sample= (6) readsensor() when pre (reset); TCCFG (16) (0) (15) (2) cnt==0 (16+5) (5) Jump out= buffer[i] (6) Andalam (UoA) PRET DATE'10 6 / 20

8 Overview of the solution Stages 1 PRET-C: simple synchronous extension to C (using macros). 2 TCCFG : intermediate format. 3 TFSM : FSM denoted with execution costs. Code void main() { while(1) { abort TCCFG TFSM (0) (15) i=0 out=0.0 sample=0.0 (6) (6) i=0 PAR(sampler,display); 7 EOT0 (2) EOT1 19 sample= (6) cnt==0 readsensor() when pre (reset); 14 EOT2 (16) (16+5) (5) Jump EOT3 out= 13 buffer[i] (6) EOT4 Andalam (UoA) PRET DATE'10 6 / 20

9 Overview of the solution Stages 1 PRET-C: simple synchronous extension to C (using macros). 2 TCCFG : intermediate format. 3 TFSM : FSM denoted with execution costs. 4 Model Checking : calculates the WCRT based on a set of TFSMs and Code a safety property. void main() { while(1) { abort TCCFG TFSM Model Checker (UPPAAL) (0) (15) EOT0 i=0 out=0.0 EOT1 sample=0.0 (6) (6) i=0 PAR(sampler,display); 7! gtick x=x 31 (2) EOT1 19 lt1=true sample=! gtick (6) cnt==0! gtick readsensor() b12 x=x 32 when pre (reset); x=x lt1=true lt1=true EOT2 gtick gtick (16) (16+5) 15 lt1= false lt1= false 32 b22 20 EOT2 b31 b13 (5) Jump EOT3! gtick! gtick x=x 37 x=x 36 gtick lt1=true lt1=true out= 13 buffer[i] lt1= false (6) b23 gtick EOT4 gtick lt1= false lt1= false Andalam (UoA) PRET DATE'10 6 / 20

10 Overview of the solution Stages 1 PRET-C: simple synchronous extension to C (using macros). 2 TCCFG : intermediate format. 3 TFSM : FSM denoted with execution costs. 4 Model Checking : calculates the WCRT based on a set of TFSMs and Code a safety property. void main() { while(1) { abort TCCFG TFSM Model Checker (UPPAAL) (0) (15) EOT0 i=0 out=0.0 EOT1 sample=0.0 (6) (6) i=0 PAR(sampler,display); 7! gtick x=x 31 (2) EOT1 19 lt1=true sample=! gtick (6) cnt==0! gtick readsensor() b12 x=x 32 when pre (reset); x=x lt1=true lt1=true EOT2 (16) (16+5) 15 Final Output gtick gtick lt1= false lt1= false 32 b22 20 EOT2 b31 b13 (5) Jump EOT3 WCRT analysis! gtick of the! gtick x=x 37 x=x 36 gtick lt1=true lt1=true out= 13 buffer[i] lt1= false (6) b23 gtick Reactive function. EOT4 gtick lt1= false lt1= false Andalam (UoA) PRET DATE'10 6 / 20

11 Motivation for PRET-C Precision Timed C (PRET-C) Simple set of synchronous extensions to C for: light-weight multithreading in C. all extensions implemented as C macros. provides thread-safe shared memory access. supports predictable programming by mapping logical time to physical time through static analysis. Andalam (UoA) PRET DATE'10 7 / 20

12 Synchronous extensions to C Statement ReactiveInput I ReactiveOutput O Meaning declares I as a reactive input coming from the environment declares O as a reactive output emitted to the environment PAR(T1,...,Tn) synchronously executes in parallel the n threads Ti, with higher priority of Ti over EOT [weak] abort P when pre C Ti+1 marks the end of a tick (local or global depending on its position) immediately kills P when C is true in the previous instant Table: PRET-C extensions to C. Andalam (UoA) PRET DATE'10 8 / 20

13 Example: Producer Consumer Initialization #include <pretc.h> #define N 1000 ReactiveInput (int, reset, 0); ReactiveInput (float, sensor, 0.0); ReactiveOutput(float, out, 0.0); int cnt=0; float buffer[n]; Andalam (UoA) PRET DATE'10 9 / 20

14 Example: Producer Consumer main void main() { while(1) { abort flush(buffer); PAR(sampler,display); when pre (reset); cnt=0; Monitors reset signal for preemption. Andalam (UoA) PRET DATE'10 10 / 20

15 Example: Producer Consumer main void main() { while(1) { abort flush(buffer); PAR(sampler,display); when pre (reset); cnt=0; Clears the buer and then spawns the sampler and the display threads. Andalam (UoA) PRET DATE'10 10 / 20

16 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick0: i=0 i=0 cnt =0 sample=0.0 out=0.0 buer={0.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

17 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick0: i=0 i=0 cnt =0 sample=1.0 out=0.0 buer={0.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

18 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick0: i=0 i=0 cnt =0 sample=1.0 out=0.0 buer={0.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

19 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 1: i=0 i=0 cnt =0 sample=1.0 out=0.0 buer={1.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

20 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 1: i=0 i=0 cnt =0 sample=1.0 out=0.0 buer={1.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

21 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 2: i=1 i=0 cnt =1 sample=2.0 out=0.0 buer={1.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

22 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 2: i=1 i=0 cnt =1 sample=2.0 out=1.0 buer={1.0, 0.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

23 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 3: i=1 i=0 cnt =1 sample=2.0 out=1.0 buer={1.0, 2.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

24 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 3: i=1 i=1 cnt =0 sample=2.0 out=1.0 buer={1.0, 2.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

25 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 4: i=2 i=1 cnt =1 sample=3.0 out=1.0 buer={1.0, 2.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

26 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 4: i=2 i=1 cnt =1 sample=3.0 out=1.0 buer={1.0, 2.0, 0.0 Andalam (UoA) PRET DATE'10 11 / 20

27 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 5: i=2 i=1 cnt =1 sample=3.0 out=1.0 buer={1.0, 2.0, 3.0 Andalam (UoA) PRET DATE'10 11 / 20

28 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 5: i=2 i=1 cnt =1 sample=3.0 out=2.0 buer={1.0, 2.0, 3.0 Andalam (UoA) PRET DATE'10 11 / 20

29 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 6: i=3 i=1 cnt =2 sample=4.0 out=2.0 buer={1.0, 2.0, 3.0 Andalam (UoA) PRET DATE'10 11 / 20

30 Example: Producer Consumer void sampler() { int i = 0; float sample; sample = read(sensor); while (cnt==n) buffer[i] = sample; i = (i + 1)% N cnt = cnt + 1; void display(){ int i = 0; float out; while (cnt==0) out = buffer[i] i = (i + 1)% N; cnt = cnt - 1; WriteLCD(out); Tick 6: i=3 i=2 cnt =1 sample=4.0 out=2.0 buer={1.0, 2.0, 3.0 Andalam (UoA) PRET DATE'10 11 / 20

31 Design Flow: PRET-C to WCRT Stages Overview 1 PRET-C to Assembly: standard gcc based compilers can be used. 2 Assembly to TCCFG : our code analyser. 3 TCCFG to Model Checker : our FSM generator (XML). 4 CTL temporal logic property checking: bounded integer checking. mb-gcc TCCFG gen FSM gen PRET-C Assembly (Microblaze) TCCFG Execution (Microblaze) Architecture Specifications Model for Model Checker (UPPAAL) Verifying CTL properties WCRT value Andalam (UoA) PRET DATE'10 12 / 20

32 Hardware support ARPRET FSL connecion Microblaze Predictable Functional Unit (PFU) (Hardware Support) FSL connecion ARPRET platform Hardware extension (PFU) to the Microblaze (GPP) in order to achieve better throughput while simplifying WCRT analysis. Fast Simplex Link (FSL) provides a predictable communication. Andalam (UoA) PRET DATE'10 13 / 20

33 Hardware support Predictable Functional Unit (PFU) PC (32bits) TDA (1bits) TSP (1bits) TLT(1bit) N PID (log 2 (N)bits) ALC (log 2 (M)bits) SC (32bits) Thread Table Scheduler WCRT Timer Valid (1bit) TID (log 2 (N)bits) PVI (log 2 (X)bits) M WS(1bit) PA (32bits) AL (log 2 (M)bits) PV (X bits) Abort Table Control and data Control and data from MB to MB Controller Logic Thread Table stores: priority, local tick. alive, suspension. spawn count. parent ID. Abort Table stores: type of preemption (Weak/Strong) nesting of preemptions. monitoring signal. preemption address. Andalam (UoA) PRET DATE'10 14 / 20

34 Benchmarking PRET-C execution: Hardware support (ARPRET). We have also developed a software model ( CEC-like linked-list based scheduler). Andalam (UoA) PRET DATE'10 15 / 20

35 Benchmarking PRET-C execution: Hardware support (ARPRET). We have also developed a software model ( CEC-like linked-list based scheduler). We have compared with other light-weight C extensions such as: SyncCharts in C. Protothreads. Esterel. Andalam (UoA) PRET DATE'10 15 / 20

36 Benchmarking Hardware vs Software Average Case Execution Time Worst Case Execution Time Hardware Software Hardware Software Clock cycles Clock cycles ABRO Channel Protocol Reactor Control Producer Consumer Smokers Robot Sonar 0 ABRO Channel Protocol Reactor Control Producer Consumer Smokers Robot Sonar On average, the throughput on the hardware is greater than the software implementation. About 28% better for ACET. About 26% better for WCET. Andalam (UoA) PRET DATE'10 16 / 20

37 Quantitative Comparison Throughput 800 Average Case Execution Time 1200 Worst Case Execution Time Software ProtoThreads Esterel SyncCharts Software ProtoThreads Esterel SyncCharts Clock cycles Clock cycles ABRO Channel Protocol Reactor Control Producer Consumer Smokers Robot Sonar 0 ABRO Channel Protocol Reactor Control Producer Consumer Smokers Robot Sonar PRET-C yields signicantly more (20% to 75%) ecient code compared to all others in both the average and worst case. Andalam (UoA) PRET DATE'10 17 / 20

38 Quantitative Comparison Memory Memory usage Software ProtoThreads Esterel SyncCharts Bytes ABRO Channel Protocol Reactor Control Producer Consumer Smokers Robot Sonar Memory usage of PRET-C is better than (2.5%)Prothreads and (26%) Esterel, while slightly worse (3.7%) than SC. Andalam (UoA) PRET DATE'10 18 / 20

39 Benchmarking Execution Time Example WCRT WCRT Over approximation (Model Checker) (Actual Execution) (%) ABRO Channel Protocol Reactor Control Producer-Consumer Smokers Robot Sonar Average 4.08 On an average, value obtained from UPPAAL over approximates the actual value by about 4%. Andalam (UoA) PRET DATE'10 19 / 20

40 Conclusions and Roadmap Contributions New synchronous language for predictable programming. Thread-safe shared memory access via simple semantics. Predictable preemption support. Hardware accelerator to improve the worst case behaviour. Mapping of logical time to physical time through static analysis. PRET-C excels both in the worst case and average case execution over other light-weight threading libraries. Future work To explore the trade-os of scratchpads versus caches. Support for parallel execution of PRET-C via new semantics. Andalam (UoA) PRET DATE'10 20 / 20