Computing task scheduling based on an intrusiveness metric
View Patent ↗Examples of computing task scheduling based on an intrusiveness metric are described. In an example, an intrusiveness metric that indicates an impact of a computing task on performance of a computing device may be determined with an intrusiveness machine learning model. The intrusiveness metric may be sent to a scheduler device to determine distribution of additional computing tasks according to a scheduling machine learning model.
1 . A method by a computing device, comprising:
determining an intrusiveness metric that indicates an impact of a computing task on performance of a computing device, the intrusiveness metric being determined with an intrusiveness machine learning model; and
sending the intrusiveness metric to a scheduler device to determine distribution of additional computing tasks according to a scheduling machine learning model, wherein the scheduler device further receives additional intrusiveness metrics from additional computing devices and determines the distribution of the additional computing tasks also according to the intrusiveness metric and the additional intrusive metrics, and wherein the additional intrusiveness metrics are determined by additional intrusiveness machine learning models on the additional computing devices.
2 . The method of claim 1 , wherein the computing task comprises training a machine learning training task received from the scheduler device.
3 . The method of claim 1 , wherein the intrusiveness metric characterizes intrusiveness of the computing task on other simultaneous processes performed by the computing device.
4 . The method of claim 1 , wherein the intrusiveness metric comprises a vector generated by the intrusiveness machine learning model.
5 . The method of claim 1 , wherein the computing device includes a machine learning isolator implemented by a coprocessor to determine the intrusiveness metric.
6 . A scheduler device, comprising:
a memory;
a processor coupled to the memory, wherein the processor is to:
receive, from a plurality of computing devices, intrusiveness metrics that indicate an impact of computing tasks on performance of the computing devices, the intrusiveness metrics being determined by an intrusiveness machine learning model on the computing devices; and
determine distribution of additional computing tasks according to a scheduling machine learning model and the intrusiveness metrics.
7 . The scheduler device of claim 6 , wherein the scheduling machine learning model is trained to indicate which computing devices can perform a computing task.
8 . The scheduler device of claim 6 , wherein the scheduling machine learning model is used by the scheduler device to learn which types of machine learning training tasks can be performed on which computing devices.
9 . The scheduler device of claim 6 , wherein the scheduling machine learning model is used by the scheduler device to learn how well different computing devices can perform a given machine learning training task.
10 . The scheduler device of claim 6 , wherein the processor is to train the scheduling machine learning model to schedule additional computing tasks based on the received intrusiveness metrics.
11 . The scheduler device of claim 6 , wherein the processor is to analyze the intrusiveness metrics to select computing devices for a next iteration of a machine learning training task.
12 . The scheduler device of claim 6 , wherein the processor is to avoid assigning additional computing tasks to computing devices that report intrusiveness metrics with high intrusiveness.
13 . The scheduler device of claim 6 , wherein the processor is to remove irrelevant computing devices from a future machine learning training task to minimize overall cost while not degrading machine learning model performance.
14 . A non-transitory tangible computer-readable medium storing executable code, comprising:
code to cause a processor to receive a machine learning training task from a scheduler device;
code to cause the processor to determine an intrusiveness metric using an intrusiveness machine learning model in response to performing the machine learning training task; and
code to cause the processor to report the intrusiveness metric to the scheduler device, wherein, the scheduler device further receives additional intrusiveness metrics determined by additional intrusiveness machine learning models on additional computing devices, and determines distribution of additional computing tasks according to a scheduling machine learning model, the intrusiveness metric, and the additional intrusiveness metric.
15 . The computer-readable medium of claim 14 , wherein the intrusiveness machine learning model is trained to differentiate cache purges that indicate intrusiveness of the machine learning training task.
16 . The computer-readable medium of claim 14 , wherein the computing task comprises training a machine learning training task received from the scheduler device.
17 . The computer-readable medium of claim 14 , wherein the intrusiveness metric characterizes intrusiveness of the machine learning task on other simultaneous processes performed by a computing device.
18 . The computer-readable medium of claim 14 , wherein the intrusiveness metric comprises a vector generated by the intrusiveness machine learning model.
19 . The computer-readable medium of claim 14 , wherein the intrusiveness machine learning model includes a machine learning isolator implemented by a coprocessor to determine the intrusiveness metric.