MC-ADAPT: Adaptive Task Dropping under Task-level Mode Switch - PowerPoint PPT Presentation

MC-ADAPT: Adaptive Task Dropping under Task-level Mode Switch Jaewoo Lee Department of Computer and Information Science University of Pennsylvania

MC-ADAPT Adaptive MC scheduling • Trends in MC scheduling - Earlier MC work drops all low-criticality tasks (LO- tasks) at mode switch - Recent MC scheduling work provides the degraded service for LO-task after mode switch  We consider to drop less LO-tasks • Our problem - How to minimize the dropping of LO-tasks? window engine window engine An example of automotive logging MP3 logging MP3 systems Adaptive MC scheduling Traditional MC scheduling

MC-ADAPT Approach High-criticality task HI-task 2 HI-task 1 • Selective task dropping under a L L fine-grained criticality mode - Existing work adopted system-level H H mode switch - We adopt task-level mode switch L L  Drop LO-tasks selectively L H

MC-ADAPT Challenges • How to drop LO-tasks under task-level mode switch? - Although different runtime scenarios need different task drop, design-time algorithms prepare offline drop decisions for a limited cases due to space complexity - Need a more flexible task drop decision • How to evaluate the quality of task dropping solution? - Hard to evaluate runtime performance (existing work often evaluates it with random simulation) - Need a formal characterization (in terms of the speedup factor)

MC-ADAPT Task Drop by an Online Test • Idea: drop LO-tasks by a simple online schedulability test • Algorithm EDF-AD ( A daptive task D ropping): VD coefficient (0<x≤1) - Schedule a HI-task with VD (= � in its LO-mode  VD  1) reserve time for HI-WCET, 2) bound the demand of LO-tasks at drop - Drop LO-tasks selectively by EDF-AD online test • EDF-AD online schedulability test: To be schedulabie, demand in the considered interval ≤ interval length ( � ) Demand before t* + � (runtime util. after t*) � t* (the rel. time of the last mode switch job) �� � � � �� �� �� � (VD) � /�� � � � � � � �� : LO-mode HI-tasks � � � �� : HI-mode HI-tasks � /� � � � � ≤ � · � � )/� � � �� : active LO-tasks (� � � · �� � )/� � � �� : dropped LO-tasks (� �

MC-ADAPT Task Drop by an Online Test active  dropped τ 1 (4,1,1,L) active τ 2 (4,1,1,L) τ 3 LO  HI (6,1,2,H) VD:3 LO τ 4 (6,1,3,H) VD:3 6 12 job release x= 0.5 By online test, drop τ 1 job (virtual) � � + � �� deadline � + � · � �� � ≤ 1 � �� � + � ��

MC-ADAPT Schedulability • EDF-VD [Baruah12]: VD-based scheduling algo. under system-level mode switch • EDF-AD : the proposed scheduling algo. under task-level mode switch • EDF-AD-E : enhanced EDF-AD algo. handling a corner case of EDF-AD deceasing the offline schedulability

MC-ADAPT The Deadline Miss Ratio of LO-tasks • Simulation: DMR varying utilization bound - The probability of mode switch: 0.4 - Each random system is schedulable by EDF-VD - Result: EDF-AD-E shows up to 42.5% lower DMR than EDF-VD The lower value is good

MC-ADAPT Speedup Factor for Task Drop the min. speedup α (α ≥1) s.t. algo. A performs the same as the optimal algo. for a problem • How to apply the speedup factor to evaluate the quality of task dropping? Schedule a given MC task set with a given mode-switch sequences by dropping the minimal # of LO-tasks • Idea: the speedup factor for the task dropping problem - The minimum speedup s.t. algo. A performs the same as OPT for any feasible task set and any mode switch sequence The optimal scheduling algo. with the optimal task dropping Algo. A sped up by α OPT HI-task 2 HI-task 1 HI-task 1 HI-task 2 LO-task LO-task LO-task LO-task LO-task LO-task An example task set and its example mode switch sequence (suppose that algo. A has a speedup factor of α )

Discussion • Scheduling algorithm for better speedup factor? - MC-ADAPT (based on EDF-VD) has a speedup factor of 1.618 for task drop  Speedup factor for task drop is no smaller than 1.333 since the task dropping problem is a generalization of MC scheduling problem - To improve the speedup factor, I conjecture that we need individual VD assignment approach  Global VD assignment approaches drop more tasks due to its inefficiency under task level mode switch  1.6 is the best speedup factor that current MC-ADAPT can achieve - Need online schedulability test (for task dropping) based on the individual VD assignment  [Ekberg12] based on demand analysis has pseudopolynomial complexity, which cannot used in runtime decision

Thank you Question?

MC-ADAPT Evaluation: Experimental Setup • Random task set generation according to [Baruah12] • Goal - Evaluate acceptance ratio, the higher the better - Evaluate Deadline Miss Ratio (DMR), the lower the better - For each setting, we generate 5,000 random systems and run 10,000 time units • Scheduling algorithms - EDF-VD [Baruah12]: VD-based scheduling algo. under system-level mode switch - EDF-AD : the proposed scheduling algo. under task-level mode switch

Offline Analysis • Offline schedulability test is derived from the online schedulability test: EDF-VD [Baruah12] EDF-AD � � � When the system � + � � starts � � � � + � � + � After mode switch � � � � � � � � � � � � � � � � � • Schedulability anomaly - Some task set is schedulable by EDF-VD but not by EDF-AD - The reason why the anomaly happens � although � � � /� ≥ � � � /� ≤ � � �  There are HI-tasks s.t. � �  VD coefficient � is derived to satisfy the latter inequality

EDF-AD-E (Enhanced) • The resolution of the schedulability anomaly - HI-mode-preferred tasks: HI-tasks s.t. � � � � - Do not execute the HI-mode-preferred tasks in LO-mode • EDF-AD-E scheduling algo. - Set the initial mode of HI-mode-preferred tasks to HI - Other rules are the same as EDF-AD • Offline schedulability analysis EDF-VD EDF-AD EDF-AD-E � � � When the � + � � system starts After mode- � � � ≤ 1 � � � � � ) � + ∑ � + � � � + ∑ � ) � · � � min ( � , � � ≤ 1 � · � � � · � � max ( � , � � ≤ 1 � � � � switch