1 / 28

William Hawkins and Tarek Abdelzaher Presented By: Farhana Dewan

Towards Feasibility Region Calculus: An End-to-end Schedulability Analysis of Real-Time Multistage Execution. William Hawkins and Tarek Abdelzaher Presented By: Farhana Dewan. Introduction System Model Generalized Stage Delay Theorem Proof of the Theorem Usage of Feasibility Region

jetta
Download Presentation

William Hawkins and Tarek Abdelzaher Presented By: Farhana Dewan

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Towards Feasibility Region Calculus: An End-to-end Schedulability Analysis of Real-Time Multistage Execution William Hawkins and TarekAbdelzaher Presented By: FarhanaDewan

  2. Introduction • System Model • Generalized Stage Delay Theorem • Proof of the Theorem • Usage of Feasibility Region • Simulation Results • Conclusion Outline CSC 8260

  3. Aperiodicdistributed system- system is large complex, workload irregular • Less attention than periodic counter part • This paper presents • Analytic framework for computing end-to-end feasibility • Fixed-priority scheduling • Based on generalized stage delay theorem • Maximum fraction of end-to-end deadline a task can spend at a resource as a function of utilization of that resource • Sum of such fractions are less than 1 • Feasibility region is considered as a volume in multi-dimensional space where each dimension is utilization of one resource • Extends uni-dimensional schedulable region to multi-dimensional representation for distributed systems • Generalizes concurrent infinitesimal tasks to arbitrary set of finite tasks Introduction CSC 8260

  4. Distributed real-time systems • Performance sensitive server farms • Radar data processing back ends • Sensor networks • Different classes of traffic traverse several stages of distribute processing • Task must exit the system within specified per-class end 2 end latency constraints • Utilization bound of resource for centralized system • U ≤ Ubound • For distributed systems resource stage i has utilization Ui • f(U1 …,Un) ≤ Cbound • Cbound systems capacity to meet deadlines Introduction CSC 8260

  5. Goal: • Simple schedulability analysis technique for distributed rtsto satisfy e2e timing constraints • Conditions are sufficient • Fast dynamic admission control • Acyclic resource system • No feedback cycle in overall task flow graph • Synthetic utilization • Non-acyclic resource system • Instantaneous utilization Introduction CSC 8260

  6. Distributed system, task Ti arrive, require execution to N (subset of) resources • Aijarrival time of Ti at stage j, 1≤j ≤N • Ai arrival time of task to the system, Ai1 • Di e2e deadline for Ti • Cijcomputation time of Ti at stage j • Set of current tasks V(t)={Ti|Ai≤t<Ai+Di} • Instantaneous utilization Uj • Synthetic utilization Uj System Model CSC 8260

  7. Urgency inversion factor αj for stage j • Less urgent task assigned greater priority • αj= min (Dlo /Dhi ) over all tasks executing at stage j such that priority(Thi)>priority(Tlo) • Blocking factor βij • Maximum amount of time task i can be blocked at stage j due to lower priority task holding critical resource • Maximum normalized blocking factor • γj = max (βij/Di) Definitions CSC 8260

  8. End to end schedulabiltiy condition: ΣjFj ≤1 Generalized Stage Delay Theorem CSC 8260

  9. Stage j processing n concurrent tasks • Instantaneous utilization at stage j for task Tm • To obtain lower bound, ignore lower priority tasks Proof CSC 8260

  10. Consider task Tn at stage j • Worst case delay at stage j, Qnj • B is the end of last processor gap • tf time at which Tn departs stage j • L = An –B offset of arrival of Tn on j • For worst case arrival scenario, L=0 • Max amount of time critical task is preempted by tasks with absolute deadline prior to tf, Proof (cont.) CSC 8260

  11. Busy period • Rearranging and substituting, we obtain instantaneous utilization Uj Proof (cont.) CSC 8260

  12. Worst case arrival sequence • T= A1j – Anj, Aij – Anj = T + Σh=1 Chj • Qnj = T + Σi=1 Cij Proof (cont.) CSC 8260

  13. Proof (cont.) CSC 8260

  14. Di is bounded by Dn/αj and Σ is minimized when Cij = C for i=1 to n-1 Proof (cont.) CSC 8260

  15. Delay: • To obtain fraction of deadline, divide by deadline • F to be worst case bound, last term must be maximized Proof (cont.) CSC 8260

  16. Generalized Stage Delay Theorem Corollary CSC 8260

  17. Stages of computation: resources can be CPU, communication links, disks • Scheduling policy at each resource • αj and γj must be pre computed • Admission controller: based on generalized stage delay theorem or corollary • Feasibility region calculus: to build admission controller • Each task arriving the system, utilization is added to Uj of each stage to be traversed by the task • Check fractional delay, if greater than 1, don’t admit, reverse the utilization modification • AC checks that the system operates in feasibility region • Complexity: linear in terms of number of stages, fractional stage delay in constant time Usage of Feasibility Region CSC 8260

  18. Simulator: distributed real-time system with arbitrary tasks • Admission Controller: for either of the cases • For each arriving task, its utilization is tentatively added to every stage j it will traverse during computation. • The generalized stage delay theorem, or its corollary if applicable, is used to check whether ΣjFj≤ 1 over the stages to be traversed. • If so, the task is admitted. If not, the task is rejected and its utilization is removed from further consideration. • Task granularity: ratio of total computation time and deadline • Load: sum of computation time of all tasks divided by simulation time Simulation Results CSC 8260

  19. Resources has increasing ids, a task leaving stage x never requires resource from stage i, 0≤i≤x • Pipeline, 1 to 5 stages, each task must be executed by each stage from 1 to 5 in order • Deadlines are drawn from uniform distribution • Task granularity is 1/100 • Load is varied from 60% to 200% • Corollary of generalized stage delay theorem is used • No task misses deadline • Each point in the plot is average of 100 simulation runs • Utilization is high for all offered loads, independent of no of stages, AC is not pessimistic Acyclic Task System CSC 8260

  20. CSC 8260

  21. CSC 8260

  22. Task may receive computation from same resource more than once (Ex- resource is database) • Each task in the experiment traverses more than 1 stage in the system • Task granularity 1/100, computation time approximately equal at each stages • Load is varied between 60% and 200% • Utilization of system with 1 stage is higher than that of 2,3 or 4 stages • Lower priority task suffer from delay, whether delay is from higher priority task in same stage or other • Stage delay corollary can be used as heuristic in admission controller to improve utilization Non-Acyclic Task System CSC 8260

  23. CSC 8260

  24. CSC 8260

  25. Stage delay theorem • System with very large number of concurrent task • Calculates feasibility region based on utilization in the stages • Generalized Stage delay theorem • Calculates feasibility region based on utilization and concurrent tasks in the stages • Two stage pipeline distributed rts • 6 task classes, arrival time and deadline from uniform distribution • For moderate number of tasks and very small granularity gsdt performs better Comparing Stage Delay Theorem with Generalized Stage Delay Theorem CSC 8260

  26. CSC 8260

  27. Presented: Analytical framework for computing the end-to-end feasibility regions of distributed aperiodic task systems under independent fixed-priority scheduling • Extended: the previous derivations of uni-dimensional schedulability regions for single processors • Generalized: the results for infinite number of concurrent liquid tasks to arbitrary sets of finite tasks • Applicable to more realistic acyclic and non-acyclic workloads Conclusion CSC 8260

  28. The results can be extended to: • Other categories of scheduling policies such as EDF • Systems that accept some percentage of deadline misses (soft real-time systems), relaxed schedulability conditions can be derived • System where tasks need multiple resources simultaneously Future Work CSC 8260

More Related