Customer-Aware Task Allocation and Scheduling for Multi-Mode MPSoCs Lin Huang, Rong Ye and Qiang Xu CHhk REliable computing laboratory (CURE) The Chinese University of Hong Kong 1 TAS and Execution Modes • Task Allocation and Scheduling T0 Task Allocation & Graph T Scheduling 1 T2 T3 T4 P1 P2 • T2 T0 T1 T3 T4 MPSoC Platform P1 P2 Periodical Schedule Multi-Mode MPSoCs (multiple execution modes) • • • Communication service Audio/Video player Digital camera… 2 Personalized TAS • Prior Works [Huang etc., DATE’09, DATE’10] • TAS solutions are generated at design stage • A unified task schedule for each execution mode is constructed for all the products • Usage Strategy Deviation • The products, bought by different end users, experience different life stories. • Personalized TAS solution for each individual product can be more energy-efficient and/or reliable 3 Motivational Example • Consider • • A simple MPSoC product with 3 execution modes and 2 processor cores 10,000 sample products 4 Problem Formulation • Problem 1 [Design Stage] • Given – – – – – – • • q execution modes and a directed acyclic task graph for each mode; The joint probability density function; A platform-based MPSoC embedded system; Execution time table; Power consumption table; The target service life and the corresponding reliability requirement. To determine a periodical task schedule for each execution mode, such that the expected energy consumption over all products is minimized under the performance and reliability constraints Problem 2 [Online Adjustment] • Given – Interval length; – Usage strategy of a specific interval; – Task mapping flexibility constraints. • To achieve the same optimization as Problem 1 5 Proposed TAS at Design Stage • • Simulated annealing-based algorithm to minimize the expected energy consumption over all the products Solution representation Task Graph • Zone Representation Two kinds of moves • • • Task Schedule M1: Insert a task in the front of its sink, if no precedunce constraint between them M2: Change the resource assignment of a task Cost function m tL j sys R (tL ) exp( ( ) ) j 1 ( s ) 6 Proposed Online Adjustment • Overall flow • • • • Resort to similar technique as design stage; The main difference stays in particularly in the cost function. Since aging effect is a slow process, online adjustment is performed at regular intervals in range of days or months as a special task. Analytical model • A forgetful scheme to infer future usage strategy y (1 ) yu (1 ) yu 1 • u 1 y1 System reliability is given by u tL u tI tI ; yu , su ) exp( ( ) j) j ( s) j 1 1 j ( s ) m sys R (tL ; y, s | y1 , s1; 7 Experimental Results • Without mapping constraints Initial Solution Online Adjustment 8 Experimental Results • With mapping constraints Online Adjustment (25% tasks with constraints) Online Adjustment (50% tasks with constraints) 9 Conclusion • • Customer-aware TAS on multi-mode MPSoCs Two phases of proposed approach • • • Simulated annealing-based algorithm at design stage Usage-specific online adjustment Experimental results • • Based on hypothetical MPSoCs with various task graphs; Show the capability to significantly increase the lifetime reliability and energy reduction of MPSoC products. 10