{sorma, petel, zebpe}@ida.liu.se Performance Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng University of Linköping, Sweden {sorma, petel, zebpe}@ida.liu.se
Overview Toolset consisting of three tools Computation model: set of annotated task graphs Main performance indicator: expected deadline miss ratio per task or task graph
Computation Model Period 2 4 6 12 9 3 Execution times execution time probab E Deadlines B C Late task policy F Scheduling policy D P1 P2
Internals Concurrently constructs and analyses the GSMP underlying the system. Method of supplementary variables + tricks for memory reduction. Constructs the GSMP. Approximates the GSMP with a much larger CTMC. Exploits the regular structure of the CTMC for on-the-fly generation of its infinitesimal generator. Recursively computes the finishing time distribution, ignoring some dependencies among some of the random variables modelling the system.
Limitations We assume non-preemptive execution in all three methods First method is efficient only in the case of monoprocessor applications Third method is limited to fixed-priority scheduling
Scalability Third method is polynomial O(NLCM/h|ETPDF|/h) First two methods build the GSMP underlying the application Analysis visits each state that described the behaviour of the system Theoretically exponential, but practically…
…Practically First method (for monoprocessors) 20 independent tasks
…Practically Second method (for multiprocessors) 60 dependent tasks on 2 processors 18 dependent tasks on 6 processors
Accuracy First method gives exact results Second method relies on approximating generalised distributions with Coxian distributions problems with non-smooth distributions Third method gives approximate results. Its accuracy is experimentally investigated
Case Study Tasks A and B have exponentially distributed execution times, average rate = 1/7 and 1/2 respectively Task C, D, and E execute for 4, 5, and 6 time units respectively Interprocessor communication takes 0.5 time units Tasks on the orange processor in descending order of priorities: E, C, D. A B E C D
Case Study (cont’d)
Conclusions Three performance analysis approaches for applications with stochastic task execution times Multiproc. Accuracy Exponential Design space exploration Exact approach No Exact Theoretically yes, practically not a problem ETPDF approx. Yes Good for smooth distributions Yes, in the number of processors Equation-based Coarser than ETPDF approx. No. Linear in the number of tasks.