Technologies for predicting power usage of a data center

• US 10,234,833 B2
• Filed: 12/30/2016
• Issued: 03/19/2019
• Est. Priority Date: 07/22/2016
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Ahuja, Nishi, Khanna, Rahul, Daniel, Abishai, Sheng, Zhijie
• Primary Examiner(s): Patel, Ramesh B
• Application Number:
  US15/396,012
• Publication Number:
  US 20180024578A1
• Time in Patent Office: 809 Days
• CPC Class Codes: G05B 13/0265 the criterion being a learning criterion
• CPC Class Codes: G06F 1/206 comprising thermal management

Abstract

Technologies for predicting the power usage of a data center are disclosed. A data center manager gathers sensor data from the compute devices of the data center. The sensor data indicates factors such as power used by the compute device and the intake air inlet temperature. The data center manager trains a machine-learning-based algorithm based on training sensor data, and then applies the machine-learning-based algorithm to sensor data as it is being gathered. The machine-learning-based algorithm can predict a change in future power usage of the data center, and control a cooling unit to compensate before the power usage even begins to change.

Claims

1. A data center manager for prediction of power usage of a data center, the data center manager comprising:

one or more processors;

one or more memory devices having stored therein a plurality of instructions that, when executed by the one or more processors, causes the data center manager to;

receive sensor data from one or more sensors from one or more sleds of the data center;

acquire one or more sensor data samples based on the sensor data;

acquire one or more parameters of a machine-learning-based algorithm;

predict, based on the one or more sensor data samples, a future power usage of a current workload of the data center with use of the one or more parameters of the machine-learning-based algorithm, wherein the future power usage of the current workload is different from a current power usage of the current workload; and

control at least one cooling unit of the data center based on the predicted future power usage.

2. The data center manager of claim 1, wherein to acquire one or more sensor data samples based on the sensor data comprises to:

average, for each of the one or more sensors, the corresponding sensor data over a period of time;

calculate, for at least one sensor of the one or more sensors, a moving average convergence/divergence (MACD) of the corresponding sensor data;

calculate, for at least one sensor of the one or more sensors, a bias of the corresponding sensor data; and

generate the one or more sensor data samples based on the averaged sensor data, the MACD of the sensor data, and the bias of the sensor data.

3. The data center manager of claim 1, wherein the plurality of instructions further causes the data center manager to:

acquire a plurality of training sensor data samples;

label each training sensor data sample of the plurality of training sensor data samples to be in a class of a plurality of classes; and

train, with use of the plurality of labeled training sensor data samples, the machine-learning-based algorithm to generate the one or more parameters.

4. The data center manager of claim 3, wherein to label each training sensor data sample to be in a class of the plurality of classes comprises to, for each training sensor data sample of the plurality of training sensor data samples:

determine a bias of a training sensor data sample that is subsequent to the corresponding training sensor data sample;

compare the bias to a fixed value;

label, if the bias is greater than the fixed value, the corresponding training sensor data sample to be in a first class of the plurality of classes; and

label, if the bias is less than or equal to the fixed value, the corresponding training sensor data sample to be in a second class of the plurality of classes.

5. The data center manager of claim 1, wherein to predict the future power usage comprises to:

classify, with use of the machine-learning-based algorithm, a sensor data sample of the one or more sensor data samples into a first class or a second class;

predict, if the sensor data sample is classified into the first class, the future power usage based on the power usage of a first period of time associated with the first class; and

predict, if the sensor data sample is classified into the second class, the future power usage based on the power usage of a second period of time associated with the second class and shorter than the first period of time.

6. The data center manager of claim 5, wherein to predict the future power usage based on the power usage in the first period of time comprises to predict the future power usage based on the power usage over the previous hour and wherein to predict the future power usage based on the power usage in the second period of time comprises to predict the future power usage based on the power usage over the previous fifteen minutes.

7. The data center manager of claim 1, wherein the one or more sensors comprise one or more of an intake air temperature sensor, an exhaust temperature sensor, a platform power sensor, and a processor power sensor.

8. A method for predicting power usage of a data center by a data center manager, the method comprising:

receiving, by the data center manager, sensor data from one or more sensors from one or more sleds of the data center;

acquiring, by the data center manager, one or more sensor data samples based on the sensor data;

acquiring, by the data center manager, one or more parameters of a machine-learning-based algorithm;

predicting, by the data center manager and based on the one or more sensor data samples, a future power usage of the data center with use of the one or more parameters of the machine-learning-based algorithm; and

controlling, by the data center manager, at least one cooling unit of the data center based on the predicted future power usage.

9. The method of claim 8, wherein acquiring one or more sensor data samples based on the sensor data comprises:

averaging, for each of the one or more sensors, the corresponding sensor data over a period of time;

calculating, for at least one sensor of the one or more sensors, a moving average convergence/divergence (MACD) of the corresponding sensor data;

calculating, for at least one sensor of the one or more sensors, a bias of the corresponding sensor data; and

generating the one or more sensor data samples based on the averaged sensor data, the MACD of the sensor data, and the bias of the sensor data.

10. The method of claim 8, further comprising:

acquiring, by the data center manager, a plurality of training sensor data samples;

labeling, by the data center manager, each training sensor data sample of the plurality of training sensor data samples to be in a class of a plurality of classes; and

training, by the data center manager and with use of the plurality of labeled training sensor data samples, the machine-learning-based algorithm to generate the one or more parameters.

11. The method of claim 10, wherein labeling each training sensor data sample to be in a class of the plurality of classes comprises, for each training sensor data sample of the plurality of training sensor data samples:

determining a bias of a training sensor data sample that is subsequent to the corresponding training sensor data sample;

comparing the bias to a fixed value;

labeling, if the bias is greater than the fixed value, the corresponding training sensor data sample to be in a first class of the plurality of classes; and

labeling, if the bias is less than or equal to the fixed value, the corresponding training sensor data sample to be in a second class of the plurality of classes.

12. The method of claim 8, wherein predicting the future power usage comprises:

classifying, with use of the machine-learning-based algorithm, a sensor data sample of the one or more sensor data samples into a first class or a second class;

predicting, if the sensor data sample is classified into the first class, the future power usage based on the power usage of a first period of time associated with the first class; and

predicting, if the sensor data sample is classified into the second class, the future power usage based on the power usage of a second period of time associated with the second class and shorter than the first period of time.

13. The method of claim 8, wherein the one or more sensors comprise one or more of an intake air temperature sensor, an exhaust temperature sensor, a platform power sensor, and a processor power sensor.

14. One or more non-transitory computer-readable media comprising a plurality of instructions stored thereon that, when executed, causes a data center manager to:

receive sensor data from one or more sensors from one or more sleds of a data center;

acquire one or more sensor data samples based on the sensor data;

acquire one or more parameters of a machine-learning-based algorithm;

predict, based on the one or more sensor data samples, a future power usage of a current workload of the data center with use of the one or more parameters of the machine-learning-based algorithm, wherein the future power usage of the current workload is different from a current power usage of the current workload; and

control at least one cooling unit of the data center based on the predicted future power usage.

15. The one or more non-transitory computer-readable media of claim 14, wherein to acquire one or more sensor data samples based on the sensor data comprises to:

average, for each of the one or more sensors, the corresponding sensor data over a period of time;

calculate, for at least one sensor of the one or more sensors, a moving average convergence/divergence (MACD) of the corresponding sensor data;

calculate, for at least one sensor of the one or more sensors, a bias of the corresponding sensor data; and

generate the one or more sensor data samples based on the averaged sensor data, the MACD of the sensor data, and the bias of the sensor data.

16. The one or more non-transitory computer-readable media of claim 14, wherein the plurality of instructions further causes the data center manager to:

acquire a plurality of training sensor data samples;

label each training sensor data sample of the plurality of training sensor data samples to be in a class of a plurality of classes; and

train, with use of the plurality of labeled training sensor data samples, the machine-learning-based algorithm to generate the one or more parameters.

17. The one or more non-transitory computer-readable media of claim 16, wherein to label each training sensor data sample to be in a class of the plurality of classes comprises to, for each training sensor data sample of the plurality of training sensor data samples:

determine a bias of a training sensor data sample that is subsequent to the corresponding training sensor data sample;

compare the bias to a fixed value;

label, if the bias is greater than the fixed value, the corresponding training sensor data sample to be in a first class of the plurality of classes; and

label, if the bias is less than or equal to the fixed value, the corresponding training sensor data sample to be in a second class of the plurality of classes.

18. The one or more non-transitory computer-readable media of claim 14, wherein to predict the future power usage comprises to:

classify, with use of the machine-learning-based algorithm, a sensor data sample of the one or more sensor data samples into a first class or a second class;

predict, if the sensor data sample is classified into the first class, the future power usage based on the power usage of a first period of time associated with the first class; and

predict, if the sensor data sample is classified into the second class, the future power usage based on the power usage of a second period of time associated with the second class and shorter than the first period of time.

19. The one or more non-transitory computer-readable media of claim 14, wherein the one or more sensors comprise one or more of an intake air temperature sensor, an exhaust temperature sensor, a platform power sensor, and a processor power sensor.

20. A data center manager for prediction of power usage of a data center, the data center manager comprising:

circuitry for receiving sensor data from one or more sensors from one or more sleds of the data center;

means for acquiring one or more sensor data samples based on the sensor data;

circuitry for acquiring one or more parameters of a machine-learning-based algorithm;

means for predicting, based on the one or more sensor data samples, a future power usage of a current workload of the data center with use of the one or more parameters of the machine-learning-based algorithm, wherein the future power usage of the current workload is different from a current power usage of the current workload; and

means for controlling at least one cooling unit of the data center based on the predicted future power usage.

21. The data center manager of claim 20, wherein the means for acquiring one or more sensor data samples based on the sensor data comprises:

means for averaging, for each of the one or more sensors, the corresponding sensor data over a period of time;

means for calculating, for at least one sensor of the one or more sensors, a moving average convergence/divergence (MACD) of the corresponding sensor data;

means for calculating, for at least one sensor of the one or more sensors, a bias of the corresponding sensor data; and

means for generating the one or more sensor data samples based on the averaged sensor data, the MACD of the sensor data, and the bias of the sensor data.

22. The data center manager of claim 20, further comprising:

means for acquiring a plurality of training sensor data samples;

means for labeling each training sensor data sample of the plurality of training sensor data samples to be in a class of a plurality of classes; and

means for training, with use of the plurality of labeled training sensor data samples, the machine-learning-based algorithm to generate the one or more parameters.

23. The data center manager of claim 22, wherein the means for labeling each training sensor data sample to be in a class of the plurality of classes comprises, for each training sensor data sample of the plurality of training sensor data samples:

means for determining a bias of a training sensor data sample that is subsequent to the corresponding training sensor data sample;

means for comparing the bias to a fixed value;

means for labeling, if the bias is greater than the fixed value, the corresponding training sensor data sample to be in a first class of the plurality of classes; and

means for labeling, if the bias is less than or equal to the fixed value, the corresponding training sensor data sample to be in a second class of the plurality of classes.

24. The data center manager of claim 20, wherein the means for predicting the future power usage comprises:

means for classifying, with use of the machine-learning-based algorithm, a sensor data sample of the one or more sensor data samples into a first class or a second class;

means for predicting, if the sensor data sample is classified into the first class, the future power usage based on the power usage of a first period of time associated with the first class; and

means for predicting, if the sensor data sample is classified into the second class, the future power usage based on the power usage of a second period of time associated with the second class and shorter than the first period of time.

25. The data center manager of claim 20, wherein the one or more sensors comprise one or more of an intake air temperature sensor, an exhaust temperature sensor, a platform power sensor, and a processor power sensor.