Modeling CPU Energy Consumption for Energy Efficient Scheduling

Size: px

Start display at page:

Download "Modeling CPU Energy Consumption for Energy Efficient Scheduling"

Nora Dean
5 years ago
Views:

1 Modeling CPU Energy Consumption for Energy Efficient Scheduling Abhishek Jaiantilal, Yifei Jiang, Shivakant Mishra University of Colorado - Boulder GCM '10 Proceedings of the 1st Workshop on Green Computing 2010 ACM

2 Outline Introduction Energy Model Overview Power Consumed and CPU Cycles Experimental Results Conclusions 2

3 Introduction (1/2) The processor is the component that consumes the most power. 3

4 Introduction (2/2) Dynamic Voltage and Frequency Scaling (DVFS) is used in CPU, referring as P-states. Per Core Power Gating (PCPG), or Dynamic Core Gating (DCG) is a hardware feature allowing the cores in a multicore CPU to shut themselves off. It is also called C-states. C0 - Active state C1 - Inactive state with the core not running on these idle cycles C3 - Inactive state with the cache saved C6 - All the PLL turned off 4

5 Energy Model Overview (1/3) Black Box approach PCPG is hardware controlled, so we use Black Box approach. Obtained the statistics of /proc/stat file A scheduling policy to limit these loops on few cores might not be the best compared with running them on all the cores. Still a low power profile. Lesser execution time. So we need to know the power consumption of a task 5

6 Energy Model Overview (2/3) Even though the processes are running at 100% load, the power consumed is different for different tasks. Because some of these tasks are float-cycle intensive and others are integer or memory cycle intensive. 6

7 Energy Model Overview (3/3) Modified Black Box approach If we know how much power a task is consuming, then we can fit a schedule that allowing for a shorter execution time and a lower energy consumption. We need the training data to choose the best task schedule depending on the tradeoff between the power consumption and the execution time. Disadvantages Need training data from all the possible tasks first Computers should have the same configuration 7

8 Power Consumed and CPU Cycles (1/7) System power consumption P(System) f(p CPU + P Memory + P Fans + P HDD + P Northbridge + P Southbridge + P Graphics + P(Other components)) f() = Efficiency of the Power supply 8

9 Power Consumed and CPU Cycles (2/7) Simplified system power consumption P(System) P CPU + P Memory + P Bias Bias = Power of Fans, Motherboard, North-bridge, South-bridge, Graphics, HDD, and Other Components. 9

10 Power Consumed and CPU Cycles (3/7) We proposed if we know the CPU cycle profile for a task, we can build a simple linear model to account the CPU load and energy consumed. P System Cycles FPU + Cycles INT + Cycles Memory + P(Bias) P(Task i ) Cycles FPU + Cycles IU + Cycles Cache N P System Power Task i i=1 + Bias 10

11 Power Consumed and CPU Cycles (4/7) We need to know the counts and the types of CPU cycles executed by a task. Dtrace for Solaris Oprofile Intel Vtune for Linux We used Vtune in an offline manner and sampled the application and store the cycle time over some period. (30 minutes~1 hour) 11

12 Power Consumed and CPU Cycles (5/7) Linear Regression Model Power Task i = F number offp cycles +I number of Int Cycles +M number of Memory Cycles F, I, and M are multiplier for watt cost of running a single FP, INT, or Memory cycle. But there is no direct way to find them. 12

13 Power Consumed and CPU Cycles (6/7) We use the statistical approach of minimizing the square error to find these unknown variables. min F,I,M Measured wattage Y Predicted wattage Y 2 Y = F Number offp cycles +I (Number of Int Cycles) +M Number of Memory Cycles F, I, M > 0, β = F I M + Bias = Xβ Once we know X, Y, then F, I, and M (stored in the β vector) can be obtained as: β = X T X + λi 1 X T Y 13

14 Power Consumed and CPU Cycles (7/7) We also used another statistical algorithm - Random Forests in our experiments. Random Forests is a popular machine learning/statistical approach that uses decision trees. It is a non-linear algorithm compared to the linear regression formulation. 14

15 Experimental Results (1/6) Regression Model Training We obtained training data from the following benchmarks first: memcpy While-float mprime Then we obtained separated test data for: SPECjvm While-Int While-Branch 15

16 Experimental Results (2/6) Results of Regression Model 16

17 Experimental Results (3/6) Energy Efficient Scheduler We proposed that we do not wake up a core from idle state until its needed. The cores that were not allocated any tasks were shut off. A core cannot execute more than a specific number of processor cycles. We used the average number of cycles executed to predict the energy consumed and then chose the best energy efficient schedule. The ideal case would be in an online fashion, based on the current load/cycle executed and evaluate the task schedule every second. 17

18 Experimental Results (4/6) 18

19 Experimental Results (5/6) 19

20 Experimental Results (6/6) 20

21 Conclusions We showed that a linear and Random Forests model can be used for predicting energy consumption. We also proposed a simple scheduler that utilizes this model to minimize power consumption but still maintain similar execution time. In the future, we propose to come up with a better mathematical model for scheduler. We also propose to use model in an online fashion and allowing the OS to limit processes that consume power greater than a fixed limit. 21

Abhishek Pandey Aman Chadha Aditya Prakash

Abhishek Pandey Aman Chadha Aditya Prakash System: Building Blocks Motivation: Problem: Determining when to scale down the frequency at runtime is an intricate task. Proposed Solution: Use Machine learning