Alternating Blind Identification of Power Sources

Alternating BPI is a tool that allows to blindly estimate the power at the core level using Expectation-Maximization. The tool is a plug and play, since it requires only one data file given as input. The data file is going to be used for training to identify two modeling matrices, A and B.

Data for training is collected by capturing thermal and power measurements of different cores through sys nodes while stressing the different cores. The stress applied to each core should have various patterns and run at different frequencies. For example, stressing clusters individually versus combined. The more combinations, the better the results. The user must then provide power and thermal data per cluster formatted into one CSV file with power in the first column, then thermal data per cluster for the remaining columns. For example, given 4 cores (clusters), the .CSV should take this form (only numbers, first row labels for reference):

Power	Core 1 thermals	Core 2 thermals	Core 3 thermals	Core 4 thermals
66.711	41275	40500	42800	44400
75.247	41475	40500	42000	40500
...	...	...	...	...

Offline power modelling is done through this tool by:

a. Identifying the A matrix using transient data ( t(k) = A t(k-1) + ε
b. Making an initial guess for the B matrix
c. Iteratively following the EM steps below:

• Given A and B matrices, find the power sources with quadratic optimization (Expectation step). ||B p(k)-(t(k)-A t(k-1))||_2, such that p(k) >= 0
• Given A, the power sources, and the temperatures, compute the B matrix (Maximization step). t(k) = A t(k-1) + B p(k) + ε

The tool will generate a p matrix containing power per core predictions using the BIPEM model as well as a figure showing the results of the BIPEM algorithm.

Sample output figure showing thermal per cluster, total power, and power per cluster using BIPEM

Usage

The main_BPI.m script is mostly plug and play.

1. The user needs to input .csv file into input directory. The file should contain the raw data from clusters stress experiment, with the first column including power data, and the remaining columns containing thermal data for each core (details above).
2. Change the variable filein at the top of the script to the input file.
3. The script will attempt to find transient traces. Follow instructions on the command line until all transient traces (specifically, traces where thermal temperature is drastically decreasing). For coarse grain tuning, enter [C] to change the initial guess by changing the sensitivity of sudden changes (previous values are shown to help with relative tuning). For fine grain transient detection:
   i. Enter [A] to add base indices of missing transient traces. Simply enter space-separated integer values for each x-axis index (input only the base index).
   ii. Enter [R] to remove extra indices of transient traces. Simply enter space-separated integer values for each y-axis thermal value (input only the left most value).
   iii. Enter [Y] if all transient traces are accounted for and are ready to proceed with BPI.

Once transient traces are ready, the rest of the script will run without user intervention. The script will preprocess transient traces for BPI training, then plot the results. Take note that step '7/9 EM-steps' may take a while to run. The plot is automatically saved into the output folder (named for the time when the script is executed) along with variables.mat which contains the matrices A, B, p, and the following data data, filein, transient_indices, num_cores to allow for future usage without waiting for training.

Demo

demo_final.mp4

Video demo showing full flow, including how to add and remove transient traces.

0:00 - input .csv file
0:09 - adding missing transient states (if the initial guess was not sensitive enough) using data cursor tool to find index
0:30 - verifying all transient states are accounted and running the BPI algorithm
0:54 - final output figure showing power per core
1:11 - removing extra transient states (if the initial guess was too sensitive)

Requirements

MATLAB R2019a or later is required, as some functions such as writematrix are introduced in this version.

License

All the source codes in this folder are licensed under the following LICENSE.