Power management method based on artificial intelligence for data storage system

Abstract

A power management method based on artificial intelligence for data storage systems collects multiple first-system-level time-series data points of application units requesting access in idle states, and uses this data to train an encoder of an autoencoder model to obtain an idle state feature vector representing the idle state. The power management method of this invention continuously collects multiple second-system-level time-series data points of application units requesting access during the real-time monitoring phase. It generates real-time state feature vectors through the encoder, calculates the statistical distance between the real-time state feature vectors and the idle state feature vectors, and determines whether the application unit requesting access is in an active or idle state based on whether the statistical distance exceeds a preset threshold, thereby managing the power consumption of the data storage system.

Claims

1. An AI-based power management method is applied to a data storage system comprising a plurality of data storage devices, wherein an access-demanding application unit accesses the data storage system. The AI-based power management method comprises the following steps: (a) during a model training phase, collecting a plurality of first system-level time-series data of the access-demanding application unit in an idle state; (b) performing an enhancement process on the plurality of first system-level data to form a plurality of first fixed-time data windows; (c) training an encoder of an autoencoder model based on the plurality of first fixed-time data windows to generate an idle state feature vector representing the idle state; (d) during a real-time monitoring phase, continuously collecting a plurality of second system-level time-series data of the access-demanding application unit; (e) continuously extracting a second fixed-time data window from the plurality of second system-level time-series data at a predetermined frequency; (f) processing the second fixed-time data window using the encoder. (g) Calculate a statistical distance between the instantaneous state feature vector and the idle state feature vector; (h) If the statistical distance exceeds a predetermined threshold and the number of times the statistical distance exceeds the predetermined threshold reaches a predetermined number, determine that the access request application unit is in an active state, and control the plurality of data storage devices to enter or remain in a first power state; and (i) If the statistical distance does not exceed the preset threshold, determine that the access request application unit is in the idle state, and after a waiting time, control the plurality of data storage devices to enter a second power state, wherein the first power consumption of one of the first power states is greater than the second power consumption of one of the second power states.

2. The AI-based power management method as described in claim 1, wherein the plurality of first system-level time series data includes at least one selected from a group consisting of a central processing unit utilization rate, a user idle time, a data read rate, a data write rate, and a change in the number of logged-in users.

3. The AI-based power management method as described in claim 2, wherein in step (b), an offset of the CPU utilization is introduced to perform the enhanced processing on the plurality of first system-level information.

4. The AI-based power management method as described in claim 1, wherein in step (c), the encoder processes the plurality of first fixed-time data windows to generate a plurality of potential spatial vectors, calculates the median of each of the potential spatial vectors, and combines the plurality of medians of the plurality of potential spatial vectors to form the idle state feature vector.

5. The AI-based power management method as described in claim 1, wherein the statistical distance is a Mahalanobis distance or a robust Mahalanobis distance.

6. The AI-based power management method as described in claim 1 further includes the following steps: dynamically adjusting the latency, wherein the latency is adjusted based on a frequency ratio of an actual frequency of power state transition of one of the plurality of data storage devices to a preset warranty frequency.

7. The AI-based power management method as described in claim 6, wherein the waiting time is increased when the frequency ratio is greater than 1, and the waiting time is decreased when the frequency ratio is less than 1.

8. The AI-based power management method as described in claim 1, wherein the first power state is an operating state and the second power state is a standby state.

9. The AI-based power management method as described in claim 1 further includes the following steps: When the plurality of data storage devices are in the second power state, if an input / output request command is received, the input / output request command is stored in a queue, the plurality of storage devices are actively woken up first, and after the plurality of data storage devices enter the first power state, the input / output request command stored in the queue is executed.

10. The AI-based power management method as described in claim 1, wherein the plurality of data storage devices are configured as a clustered disk (JBOD) architecture or a fault-tolerant disk array (RAID) architecture.

Images