Augmented power control within a datacenter using predictive modeling
In disclosed techniques, augmented power control within a datacenter uses predictive modeling. A power usage by a first data rack within a datacenter is measured, using one or more processors, over a first period of time. A predicted power usage by the first data rack over a second period of time is generated on a computing device, wherein the second period of time is subsequent to the first period of time. A power correlation model is calculated that correlates a power prediction to an actual power usage. The predicted power usage for the second period of time is refined, based on the predicted power usage and the power correlation model. The refining is accomplished using a power prediction model comprising the predicted power usage and the power correlation model. Datacenter power structure is configured, based on the refined power usage prediction.
1. A computer-implemented method for power management comprising:
measuring, using one or more processors, a power usage by a first data rack within a datacenter over a first period of time;
generating, on a computing device, a predicted power usage by the first data rack over a second period of time, wherein the second period of time is subsequent to the first period of time;
calculating, on the computing device using a neural network, a power correlation model that correlates a power prediction to an actual power usage for the datacenter;
refining, on the computing device, the predicted power usage for the second period of time, based on the predicted power usage and the power correlation model resulting in a refined power usage prediction, wherein the refining includes adjusting at least one coefficient or weight in the power correlation model and wherein the refining includes adding a power buffer margin calibrated over time; and
configuring, using the one or more processors, a power structure for the datacenter, based on the refined power usage prediction.
2. The method of claim 1 wherein the refining is accomplished using a power prediction model comprising the predicted power usage and the power correlation model.
3. The method of claim 2 wherein the refining further comprises updating the power prediction model based on the refining.
4. The method of claim 1 further comprising refining predicted power usage for a third period of time wherein the third period of time is subsequent to the second period of time and wherein the refining is based on the power correlation model.
5. The method of claim 1 wherein the power buffer margin is used in a first power rack that was part of the measuring the power usage.
6. The method of claim 1 wherein the configuring is used for implementing dynamic redundancy within the datacenter.
7. The method of claim 6 wherein the implementing dynamic redundancy is accomplished using a policy for a power switch within the datacenter based on the refined power usage prediction.
8. The method of claim 7 wherein the dynamic redundancy comprises providing redundant power to data racks within the datacenter for a given period of time.
9. The method of claim 1 further comprising measuring a power usage by a second data rack within the datacenter over the first period of time.
10. The method of claim 9 further comprising aggregating the power usage by a first data rack with the power usage by a second data rack over the first period of time.
11. The method of claim 9 further comprising measuring power usage by further data racks within the datacenter over the first period of time.
12. The method of claim 11 further comprising aggregating the power usage by further data racks with the power usage by a first data rack.
13. The method of claim 12 further comprising generating a predicted power usage for the datacenter over the second period of time based on the power usage, by a first data rack and the further data racks, that was aggregated.
14. The method of claim 13 further comprising determining an aggregated refined power usage prediction for the datacenter based on the predicted power usage that was aggregated and the power correlation model.
15. The method of claim 1 wherein the power correlation model comprises coefficients, tree logic, or algorithm logic based on historical experience with predicted power usage and actual power usage within the datacenter.
16. The method of claim 1 further comprising identifying peak power patterns and further refining the refined power usage prediction based on the peak power patterns.
17. The method of claim 1 further comprising determining metrics for accuracy of the refined power usage prediction.
18. The method of claim 17 further comprising determining a permissible accuracy associated with the metrics for accuracy on the refined power usage prediction.
19. The method of claim 1 further comprising performing learning to determine the power correlation model.
20. The method of claim 1 wherein the predicted power usage is based on applications running on the first data rack.
21. The method of claim 1 wherein the first period of time and the second period of time are for a granular period of time.
22. The method of claim 21 wherein the granular period of time includes a sampling rate and a sample duration to measure power usage.
23. The method of claim 1 further comprising altering a datacenter configuration based on the refined power usage prediction.
24. The method of claim 1 wherein the one or more processors are on the computing device.
25. The method of claim 1 further comprising performing quantile regression as part of the refining the predicted power usage.
26. The method of claim 25 wherein the quantile regression is accomplished using Markov models.
27. The method of claim 1 further comprising identifying a spike in power usage and refining the predicted power usage based on the spike in power usage.
28. The method of claim 1 wherein predictive modeling is used to refine the predicted power usage based on a quadratic regression model.
29. A computer program product embodied in a non-transitory computer readable medium for power management, the computer program product comprising code which causes one or more processors to perform operations of:
measuring a power usage by a first data rack within a datacenter over a first period of time;
generating a predicted power usage by the first data rack over a second period of time, wherein the second period of time is subsequent to the first period of time;
calculating, using a neural network, a power correlation model that correlates a power prediction to an actual power usage for the datacenter;
refining the predicted power usage for the second period of time, based on the predicted power usage and the power correlation model resulting in a refined power usage prediction, wherein the refining includes adjusting at least one coefficient or weight in the power correlation model and wherein the refining includes adding a power buffer margin calibrated over time; and
configuring a power structure for the datacenter, based on the refined power usage prediction.
30. A computer system for power management comprising:
a memory which stores instructions;
one or more processors attached to the memory wherein the one or more processors, when executing the instructions which are stored, are configured to:
measure, using the one or more processors, a power usage by a first data rack within a datacenter over a first period of time;
generate, on a computing device, a predicted power usage by the first data rack over a second period of time, wherein the second period of time is subsequent to the first period of time;
calculate, on the computing device using a neural network, a power correlation model that correlates a power prediction to an actual power usage for the datacenter;
refine, on the computing device, the predicted power usage for the second period of time, based on the predicted power usage and the power correlation model resulting in a refined power usage prediction, wherein refining the predicted power usage includes adjusting at least one coefficient or weight in the power correlation model and wherein the refining includes adding a power buffer margin calibrated over time; and
configure, using the one or more processors, a power structure for the datacenter, based on refined power usage prediction.