Low Power Design Prof. Dr. J. Henkel CES - Chair for Embedded - PowerPoint PPT Presentation

1 Thermal Management Low Power Design Prof. Dr. J. Henkel CES - Chair for Embedded Systems KIT, Germany Thermal Management – Part 2 (Thomas Ebi) http://ces.itec.kit.edu T. Ebi, KIT, SS13

2 Thermal Management Overview  Thermal modeling & Simulation  Multi-core architectures  Motivation Part 2  Reactive thermal management  Proactive thermal management  3D architectures  Thermal Management at CES http://ces.itec.kit.edu T. Ebi, KIT, SS13

3 Thermal Management The RC-Model P P 1 2 P RC equivalent thermal circuit for P 3 4 single component with heat dissipating, e.g. through packaging Voltage ≙ Temperature Current ≙ Heat dissipation RC equivalent thermal circuit for four component s with heat dissipating to outside through package (Cp, Rp) This gives us the thermal equation from last week as: dT T P dt R C C [Shi, 2010] http://ces.itec.kit.edu T. Ebi, KIT, SS13

4 Thermal Management The RC Model (cont) [Skadron, 2004] http://ces.itec.kit.edu T. Ebi, KIT, SS13

5 Thermal Management Thermal Simulation  Thermal simulators such as HotSpot calculate thermal distribution by solving equation of RC equivalent model  Accuracy of simulation dependent on the granularity of components  Block based: coarse granularity (CPU, cache, etc.), fast  Grid based: divides blocks into smaller parts, slower, more accurate temperature distribution, slow  Accuracy also dependent on the power input!  Instruction-based simulators count execution of instructions and know power consumption of each block  E.g. Wattch, m5+McPAt  Inaccurate but fast (Wattch inaccuracy up to 30%) [Brooks 2000]  Circuit-based simulators  Highly accurate but very slow http://ces.itec.kit.edu T. Ebi, KIT, SS13

6 Thermal Management Thermal Sensors: Thermal Diodes  Currently most common method for on-chip thermal measurement  Used by Intel, AMD, Xilinx, etc..  Xilinx Virtex 5 FPGA datasheet: Accuracy +/- 4°C  Analog circuitry  Needs A/D converter  Occupies large chip area [Long, 2008] http://ces.itec.kit.edu T. Ebi, KIT, SS13

7 Thermal Management Thermal Sensors: Ring Oscillator  Idea: analyze negative thermal side-effects to quantify temperature  Due to increased delay ring oscillators oscillate slower at higher temperatures  Oscillation frequency determined using a reference clock  Provide relative temperature values  Challenge: must be calibrated to obtain absolute values  Xilinx reference design: [src: Xilinx] Inverter Delay http://ces.itec.kit.edu T. Ebi, KIT, SS13

8 Thermal Management Thermal Sensors: Leakage based  Since leakage is temperature dependent, measuring leakage can also determine temperature Idea: measure the time a capacitor takes to discharge capacitance through leakage current 1. Input switches from low-to-high  M1 transitions from “on” to “off”  Charge stored in CL should remain, but slowly decreases due to leakage current 2. When voltage of CL falls below a threshold, [Ituero 2008] the inverter M3-M4 produces a low-to-high transition 3. Temperature can be determined by the delay between the input and output transitions http://ces.itec.kit.edu T. Ebi, KIT, SS13

9 Thermal Management Multi-core Motivation Hot Tile 4 Tile 3 Tile 8 Tile 2 Tile 7 Tile 12 Tile 11 Tile 16 Tile 1 Tile 6 Cold Tile 15 Tile 5 Tile 10 Tile 9 Tile 14 Tile 13 Spreading applications reduces thermal hotspots  Thermal hotspots! http://ces.itec.kit.edu T. Ebi, KIT, SS13

10 Thermal Management Example Platform: Intel’s SCC  24 Tiles each consisting of two Pentium cores  Two thermal sensors per tile (same principle as ring oscillators)  Frequency scaling per core (100-800MHz)  Voltage scaling per “voltage island” (4 Tiles per island, 1 island for on-chip mesh comm. network, 208 voltage levels)  Tile area: 18.7mm 2  1.3B transistors at 45nm process [src: intel] http://ces.itec.kit.edu T. Ebi, KIT, SS13

11 Sensors on the SCC  Half of the cores running the program, half in idle state Nikil Dutt and Jörg Henkel, Tutorial @ ASP-DAC 2013

12 Thermal Management Problems  Mutual heating  Heat conducts to surrounding areas  Thermal gradients  Variations of temperature across chip  Thermal cycling  Management may lead to periodic heating/cooling http://ces.itec.kit.edu T. Ebi, KIT, SS13

13 Thermal Management Multi-core thermal management  Classification of thermal management approaches:  Reactive approaches  Depend on the current temperature  Proactive approaches  Predict the temperature  Aim to balance temperature to avoid hotspots  Naïve reactive approache:  [Skadron, ISCA.2004] controls the temperature by:  Switching off the hottest core and turning on the coldest one,  but that leads to:  Thermal cycling and large spatial variations  Negative effect on the performance. http://ces.itec.kit.edu T. Ebi, KIT, SS13

14 Thermal Management Reactive approaches (cont’d)  [Coskun, 2007] proposed two OS-level methods that achieve temperature-aware task scheduling.  First method: Coolest-FLP  Depends on the current temperature and floor-plan. • Select the coolest processors For each ready job •Give priority to processors, whose neighbors are “idle”  Reduces the hot spots.  Second method: probabilistic method  Takes into consideration the analysis of the temperature history. • Calculates the probability for each core to receive the incoming job For each P n = P n-1 ± W ready job Weight depends on the core‟s history Previous probability  Achieves more balancing in the temperature and reduces the spatial variation in the temperature http://ces.itec.kit.edu T. Ebi, KIT, SS13

15 Thermal Management Reactive approaches (cont’d)  [Coskun ASPDAC 2008] uses Integer Linear Programming (ILP):  Models the applications as tasks graph  Results in optimal task scheduling for  Given set of tasks with deadlines and dependence constraints  Given temperature profiles.  Aims at reaching the best temporal and spatial distribution of temperature http://ces.itec.kit.edu T. Ebi, KIT, SS13

16 Thermal Management Reactive approaches (cont’d) Normal mode: Thermal balancing mode:  Processing demand < certain threshold.  Processing demand > certain threshold.  Goal: maximize energy savings with  Goal: prevent concentration of high meeting performance demands and power densities, then saving energy. thermal constraints. Yes Demand > α No Task assignment to the cores Global frequency assignment Core-Level frequency assignment Task assignment to the cores Calculating processing demand Calculating processing demand No Yes Demand < β http://ces.itec.kit.edu T. Ebi, KIT, SS13

17 Thermal Management Proactive Approach  [Coskun 2008 ] uses autoregressive moving average (ARMA) modeling to:  Predicting the future temperature from history  Apply thermal-aware job allocation method, which aims to:  Avoid reaching a set thermal threshold achieve and balance the temperature across the chip Temperature Data from Thermal Sensors ARMA Model Validation: Predictor (ARMA) Update Model if Necessary Temperature at time (Tcurrent+tn) for all cores Scheduler Temperature-Aware Allocation on Cores http://ces.itec.kit.edu T. Ebi, KIT, SS13

18 Thermal Management Proactive Approach  ARMA models autocorrelation in a time series y t - value at time t p q e t - noise/error at time t y ( a y ) e ( c e ) a - autoregressive coef. t i t i t i t i c - moving avrg. coef. i 1 i 1  Given a stationary stochastic process  y t can be predicted as weighted sum of past values and moving average of error term  Steps involved:  Identification: determine p and q  Estimation: determine coefficients a and c  Model checking: determine quality of estimated values http://ces.itec.kit.edu T. Ebi, KIT, SS13  aas

19 Thermal Management Proactive Approach  Benefits of ARMA model  Model is generated through automated process  Does not require in depth thermal knowledge  High accuracy achievable with large number of samples (>150)  Shortcomings  Workloads vary over time  temperature is not a stationary function!  Solution: Thermal sensors are used to check if model is still valid If not, model is updated at runtime  As such: requires thermal sensors on each core http://ces.itec.kit.edu T. Ebi, KIT, SS13

20 Thermal Management Multicore Strategies & Scalability  „Centralized‟ management scheme: Manager can use global knowledge but also forms bottleneck for communication as well as computation  central point of failure, limited scalability  „Fully distributed‟ scheme: No central bottlenecks. Management is limited by local knowledge  can result in local maxima/minima  Hierarchical scheme: Combines local management with access to global knowledge http://ces.itec.kit.edu T. Ebi, KIT, SS13