Deep reinforcement learning for workflow optimization

1. A method, comprising:

obtaining a specification of a plurality of concurrent workflows in a shared computing environment, wherein the specification of a given one of the plurality of concurrent workflows comprises a plurality of states of the given workflow and one or more control variables indicating an allocation of one or more resources for the given workflow in the shared computing environment;

evaluating, using at least one processing device, a plurality of values of the one or more control variables indicating the allocation of the one or more resources, for an execution of said plurality of concurrent workflows, using at least one reinforcement learning agent, wherein said evaluating comprises observing said plurality of states, including a current state comprising a current configuration of said plurality of concurrent workflows and said shared computing environment, and obtaining, from the at least one reinforcement learning agent, an expected utility score for a plurality of combinations of said control variables for the execution of said plurality of concurrent workflows given an allocation of the one or more resources of the shared computing environment corresponding to said combination of said control variables in said current state, wherein the at least one reinforcement learning agent traverses said plurality of states and is trained to select a particular action for a given state, wherein the particular action corresponds to the allocation of the one or more resources of the shared computing environment for the given state; and

initiating an adjustment of the allocation of the one or more resources of the shared computing environment reflecting the combination of the control variables having the expected utility score, from the at least one reinforcement learning agent, that satisfies one or more predefined score criteria.

2. The method of claim 1 , further comprising the step of applying the allocation of the one or more resources of the shared environment.

3. The method of claim 1 , further comprising the step of updating said at least one reinforcement learning agent by further training a model with the states that result from said allocation as new training samples.

4. The method of claim 1 , wherein said expected utility score further comprises an expected cost depending on one or more of an execution time of the given workflow and a consumption of resources in said shared computing environment.

5. The method of claim 1 , wherein said states further comprise provenance data of said plurality of workflows.

6. The method of claim 1 , wherein said states further comprise telemetry data of said shared computing environment.

7. The method of claim 1 , wherein said at least one reinforcement learning agent comprises a Deep Q-Learning agent using a Q-Deep Neural Network (QDNN) as a representation of a Q-Function, and wherein said obtaining the expected utility score for the plurality of combinations of said control variables comprises selecting an action at random and computing a cost-to-go from the expected utility score of the selected action updated by an observation of the current state, and wherein an updating of the at least one reinforcement learning agent comprises a training of the QDNN given new samples in iterative epochs.

8. The method of claim 7 , wherein the values of the expected utility score are given by predictions from a Deep Neural Network for a predefined number of training epochs.

9. The method of claim 7 , wherein the obtaining the expected utility score for the plurality of combinations of said control variables further comprises the step of updating the cost-to-go from a previous state with a future value given by a pretrained neural network.

10. The method of claim 7 , wherein said computation of the cost-to-go from the expected utility score of the selected action updated by the observation of the current state additionally comprises the storage of input/output pairs in a database of samples and wherein a training batch for said training of the QDNN is comprised of new samples from the database processed in iterative epochs.

11. The method of claim 7 , wherein estimates of substantially optimal future values in said computation of the cost-to-go from the expected utility score of the selected action updated by the observation of the current state are given by the outputs of the QDNN with an architecture that yields multiple outputs configuring the expected utility scores for substantially all actions.

12. The method of claim 1 , wherein the one or more control variables comprise one or more of a number of processing cores allocated to a given workflow and an amount of memory allocated to the given workflow.

13. A system, comprising:

a memory; and

at least one processing device, coupled to the memory, operative to implement the following steps:

obtaining a specification of a plurality of concurrent workflows in a shared computing environment, wherein the specification of a given one of the plurality of concurrent workflows comprises a plurality of states of the given workflow and one or more control variables indicating an allocation of one or more resources for the given workflow in the shared computing environment;

evaluating, using at least one processing device, a plurality of values of the one or more control variables indicating the allocation of the one or more resources, for an execution of said plurality of concurrent workflows, using at least one reinforcement learning agent, wherein said evaluating comprises observing said plurality of states, including a current state comprising a current configuration of said plurality of concurrent workflows and said shared computing environment, and obtaining, from the at least one reinforcement learning agent, an expected utility score for a plurality of combinations of said control variables for the execution of said plurality of concurrent workflows given an allocation of the one or more resources of the shared computing environment corresponding to said combination of said control variables in said current state, wherein the at least one reinforcement learning agent traverses said plurality of states and is trained to select a particular action for a given state, wherein the particular action corresponds to the allocation of the one or more resources of the shared computing environment for the given state; and

initiating an adjustment of the allocation of the one or more resources of the shared computing environment reflecting the combination of the control variables having the expected utility score, from the at least one reinforcement learning agent, that satisfies one or more predefined score criteria.

14. The system of claim 13 , further comprising the step of updating said at least one reinforcement learning agent by further training a model with the states that result from said allocation as new training samples.

15. The system of claim 13 , wherein said expected utility score further comprises an expected cost depending on one or more of an execution time of the given workflow and a consumption of resources in said shared computing environment.

16. The system of claim 13 , wherein said at least one reinforcement learning agent comprises a Deep Q-Learning agent using a Q-Deep Neural Network (QDNN) as a representation of a Q-Function, and wherein said obtaining the expected utility score for the plurality of combinations of said control variables comprises selecting an action at random and computing a cost-to-go from the expected utility score of the selected action updated by an observation of the current state, and wherein an updating of the at least one reinforcement learning agent comprises a training of the QDNN given new samples in iterative epochs.

17. The system of claim 16 , wherein the values of the expected utility score are given by predictions from a Deep Neural Network for a predefined number of training epochs.

18. The system of claim 16 , wherein the obtaining the expected utility score for the plurality of combinations of said control variables further comprises the step of updating the cost-to-go from a previous state with a future value given by a pretrained neural network.

19. The system of claim 16 , wherein said computation of the cost-to-go from the expected utility score of the selected action updated by the observation of the current state additionally comprises the storage of input/output pairs in a database of samples and wherein a training batch for said training of the QDNN is comprised of new samples from the database processed in iterative epochs.

20. A computer program product, comprising a tangible machine-readable storage medium having encoded therein executable code of one or more software programs, wherein the one or more software programs when executed by at least one processing device perform the following steps:

obtaining a specification of a plurality of concurrent workflows in a shared computing environment, wherein the specification of a given one of the plurality of concurrent workflows comprises a plurality of states of the given workflow and one or more control variables indicating an allocation of one or more resources for the given workflow in the shared computing environment;

evaluating, using at least one processing device, a plurality of values of the one or more control variables indicating the allocation of the one or more resources, for an execution of said plurality of concurrent workflows, using at least one reinforcement learning agent, wherein said evaluating comprises observing said plurality of states, including a current state comprising a current configuration of said plurality of concurrent workflows and said shared computing environment, and obtaining, from the at least one reinforcement learning agent, an expected utility score for a plurality of combinations of said control variables for the execution of said plurality of concurrent workflows given an allocation of the one or more resources of the shared computing environment corresponding to said combination of said control variables in said current state, wherein the at least one reinforcement learning agent traverses said plurality of states and is trained to select a particular action for a given state, wherein the particular action corresponds to the allocation of the one or more resources of the shared computing environment for the given state; and

initiating an adjustment of the allocation of the one or more resources of the shared computing environment reflecting the combination of the control variables having the expected utility score, from the at least one reinforcement learning agent, that satisfies one or more predefined score criteria.