Power consumption regulation and control method, apparatus and device, and readable storage medium

1. A power regulation and control method, comprising:

in response to a node power value of an artificial intelligence (AI) computing node being greater than a warning power value, acquiring, by a baseboard management controller (BMC), chip power values of computing chips in the AI computing node;

obtaining a grouping result by grouping, according to the chip power values of the computing chips, the computing chips; and

in response to the node power value being greater than a power capping value, querying a power regulation and control policy corresponding to the grouping result, and regulating power limit values of the computing chips according to the power regulation and control policy, such that a sum of all the power limit values is in a target range, wherein the warning power value is less than the power capping value, the power regulation and control policy is preset on the basis of the target range, and the target range is used for regulating and controlling energy efficiency values of the computing chips in the AI computing node;

wherein the obtaining a grouping result by grouping, according to the chip power values of the computing chips, the computing chips comprises:

acquiring an optimal power value in the target range, determining a buffer section having the optimal power value served as a mid-value, dividing computing chips having chip power values that are not in the buffer section and are greater than the optimal power value into a first group, and dividing computing chips having chip power values that are not in the buffer section and not greater than the optimal power value into a second group;

calculating a first power gradient for the first group, a second power gradient for the second group, and chip power gradients for the computer chips in the buffer section; and

calculating a first similarity between any one of the chip power gradients and the first power gradient and a second similarity between the chip power gradient and the second power gradient, in response to determining that the first similarity is greater than the second similarity, adding a computing chip corresponding to the chip power gradient into the first group, in response to determining that the first similarity is less than the second similarity, adding the computing chip corresponding to the chip power gradient into the second group, and taking the first group and the second group as the grouping result.

2. The power regulation and control method according to claim 1 , wherein before acquiring, by a baseboard management controller (BMC), chip power values of computing chips in the AI computing node, the method further comprises:

monitoring, by the BMC, the node power value in real time, and in response to the node power value being greater than the warning power value, starting a power regulation and control function.

3. The power regulation and control method according to claim 2 , wherein the monitoring, by the BMC, the node power value in real time comprises:

controlling the BMC to monitor the node power value through a power management bus (PMBUS).

4. The power regulation and control method according to claim 1 , wherein the acquiring, by a baseboard management controller (BMC), chip power values of computing chips in the AI computing node comprises:

obtaining respective chip power values by reading current values of the computing chips in the AI computing node by the BMC.

5. The power regulation and control method according to claim 4 , wherein the obtaining respective chip power values by reading current values of the computing chips in the AI computing node by the BMC comprises:

polling, by the BMC, current sensors corresponding to the computing chips for many times within a preset time period to obtain corresponding chip power values.

6. The power regulation and control method according to claim 1 , wherein the target range is determined by:

acquiring, by a performance benchmark test tool, energy efficiency values of any one of the computing chips in the AI computing node under different chip power values, wherein the computing chips in the AI computing node are identical;

calculating a ratio of each of the energy efficiency values to a thermal design power (TDP) energy efficiency value; and

adding energy efficiency value corresponding to the ratio greater than a preset threshold into an energy efficiency regulation and control section, and determining chip power values corresponding to the energy efficiency regulation and control section as the target range.

7. The power regulation and control method according to claim 6 , wherein the computing chip is a graphics processing unit (GPU), a neural network processing unit (NPU), a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC).

8. The power regulation and control method according to claim 1 , wherein the first power gradient for the first group is a sequence formed by a difference of adjacent chip power values of any one of the computing chips in the first group within a preset time period; the second power gradient for the second group is a sequence formed by a difference of adjacent chip power values of any one of the computing chips in the second group within a preset time period; each of the chip power gradients is a sequence formed by a difference of adjacent chip power values of the computing chips in the buffer section within a preset time period.

9. The power regulation and control method according to claim 1 , wherein the AI computing node is a device for running a machine learning algorithm, and comprises a plurality of computing chips.

10. The power regulation and control method according to claim 1 , wherein the node power value is an actual total power value of the AI computing node, the warning power value and the power capping value are both preset values.

11. A power regulation and control device, comprising:

a memory configured to store a computer program; and

a processor configured to execute the computer program to:

in response to a node power value of an artificial intelligence (AI) computing node being greater than a warning power value, acquire, by a baseboard management controller (BMC), chip power values of computing chips in the AI computing node;

obtain a grouping result by grouping, according to the chip power values of the computing chips, the computing chips; and

in response to the node power value being greater than a power capping value, query a power regulation and control policy corresponding to the grouping result, and regulate power limit values of the computing chips according to the power regulation and control policy, such that a sum of all the power limit values is in a target range, wherein the warning power value is less than the power capping value, the power regulation and control policy is preset on the basis of the target range, and the target range is used for regulating and controlling energy efficiency values of the computing chips in the AI computing node;

wherein the processor is further configured to:

acquire an optimal power value in the target range, determine a buffer section having the optimal power value served as a mid-value, divide computing chips having chip power values that are not in the buffer section and are greater than the optimal power value into a first group, and divide computing chips having chip power values that are not in the buffer section and not greater than the optimal power value into a second group;

calculate a first power gradient for the first group, a second power gradient for the second group, and chip power gradients for the computer chips in the buffer section; and

calculate a first similarity between any one of the chip power gradients and the first power gradient and a second similarity between the chip power gradient and the second power gradient, in response to determining that the first similarity is greater than the second similarity, add a computing chip corresponding to the chip power gradient into the first group, in response to determining that the first similarity is less than the second similarity, add the computing chip corresponding to the chip power gradient into the second group, and take the first group and the second group as the grouping result.

12. The power regulation and control device according to claim 11 , wherein the processor is further configured to:

monitor, by the BMC, the node power value in real time, and in response to the node power value being greater than the warning power value, starting a power regulation and control function.

13. The power regulation and control device according to claim 12 , wherein the processor is further configured to:

control the BMC to monitor the node power value through a power management bus (PMBUS).

14. The power regulation and control device according to claim 11 , wherein the processor is further configured to:

obtain respective chip power values by reading current values of the computing chips in the AI computing node by the BMC.

15. A non-transitory readable storage medium storing a computer program that, when executed by a processor, causes the processor to: in response to a node power value of an artificial intelligence (AI) computing node being greater than a warning power value, acquire, by a baseboard management controller (BMC), chip power values of computing chips in the AI computing node; obtain a grouping result by grouping, according to the chip power values of the computing chips, the computing chips; and in response to the node power value being greater than a power capping value, query a power regulation and control policy corresponding to the grouping result, and regulate power limit values of the computing chips according to the power regulation and control policy, such that a sum of all the power limit values is in a target range, wherein the warning power value is less than the power capping value, the power regulation and control policy is preset on the basis of the target range, and the target range is used for regulating and controlling energy efficiency values of the computing chips in the AI computing node; wherein the processor is further configured to: acquire an optimal power value in the target range, determine a buffer section having the optimal power value served as amid-value, divide computing chips having chip power values that are not in the buffer section and are greater than the optimal power value into a first group, and divide computing chips having chip power values that are not in the buffer section and not greater than the optimal power value into a second group; calculate a first power gradient for the first group, a second power gradient for the second group, and chip power gradients for the computer chips in the buffer section; and calculate a first similarity between any one of the chip power gradients and the first power gradient and a second similarity between the chip power gradient and the second power gradient, in response to determining that the first similarity is greater than the second similarity, add a computing chip corresponding to the chip power gradient into the first group, in response to determining that the first similarity is less than the second similarity, add the computing chip corresponding to the chip power gradient into the second group, and take the first group and the second group as the grouping result.