Exploiting Dynamic Phase Distance Mapping for Phasebased Tuning

  • Slides: 6
Download presentation
Exploiting Dynamic Phase Distance Mapping for Phase-based Tuning of Embedded Systems Tosiron Adegbija and

Exploiting Dynamic Phase Distance Mapping for Phase-based Tuning of Embedded Systems Tosiron Adegbija and Ann Gordon-Ross+ Department of Electrical and Computer Engineering University of Florida, Gainesville, Florida, USA Also Affiliated with NSF Center for High-Performance Reconfigurable Computing + This work was supported by National Science Foundation (NSF) grant CNS-0953447

Introduction and Motivation • Embedded systems are pervasive and have stringent design constraints –

Introduction and Motivation • Embedded systems are pervasive and have stringent design constraints – Constraints: Energy, size, real time, cost, etc • System optimization is challenging due to numerous tunable parameters – Tunable parameters: parameters that can be changed • E. g. , cache size, associativity, line size, clock frequency, etc – Many combinations large design space – Phase-based tuning increases optimization potential • How to determine a phase’s best configuration without significant runtime overhead? 2

Previous Work: Phase Distance Mapping (PDM) New phase Previously characterized phase Base phase Cache

Previous Work: Phase Distance Mapping (PDM) New phase Previously characterized phase Base phase Cache characteristics Pb Phase Pi Cache configuration Cache characteristics Cb Pi d (Pb, Pi) Cache configuration ? Phase distance Distance windows Configuration distance Statically defined phase distance ranges Configuration change from base phase’s configuration 3

Phase Distance Mapping (PDM) • PDM’s limitations Statically defined Distance windows Needed to know/analyze

Phase Distance Mapping (PDM) • PDM’s limitations Statically defined Distance windows Needed to know/analyze applications a priori Unknown applications/ general purpose systems? Design time overhead!!! Base phase Most prominent application domain 4

Contributions Dyna. PDM: Dynamic Phase Distance Mapping Dynamically defined at runtime Adapts to runtime

Contributions Dyna. PDM: Dynamic Phase Distance Mapping Dynamically defined at runtime Adapts to runtime application changes Distance windows No need to know/analyze applications a priori Suitable for unknown applications/ general purpose systems Low overhead dynamic method for determining a phase’s best configuration! Design time overhead!!! Base phase Dynamically determined Most prominent application domain 5

1 0. 9 0. 8 0. 7 0. 6 0. 5 0. 4 0.

1 0. 9 0. 8 0. 7 0. 6 0. 5 0. 4 0. 3 0. 2 0. 1 0 47% Optimal PDM Dyna. PDM 28% Dyna. PDM’s configurations 1% of the optimal! - ate ro t ate -1 6 x 4 M s 3 -1 2 w 6 x 4 M 8 64 s 4 w 8 M -ro tat atew ro e 4 M 2 tat ro etat s 4 52 ew 1 0 k co lo -2 r-4 70 de M g -9 0 d eg 4 M w 1 -c 4 M -re heck ip as pk 4 M sem tch bl ec -tc y kp 8 x m 4 M ix ed -4 ip W re o s-6 rk e m d 5 M 4 w r -1 o 28 rk er M m 4 d 5 wo -3 2 M rke 4 w r or ke r m d 5 em 4 M pt ywl d hu ffd ea Av ll er ag e EDP normalized to the base cache configuration Results EDP savings calculated with respect to the base configuration Dyna. PDM achieved 28% average EDP savings overall Savings as high as 47% for 64 M-rotatew 2 On average, within 1% of the optimal EDP improved over PDM by 8% Dyna. PDM improved over PDM and eliminated design-time effort! 6