Latency in edge computing

1. A computer-implemented method for edge device self-orchestration in a spatial network community, the method comprising:

initializing a first deep neural network (DNN) for a local deep reinforcement learning (DRL) network of a first local cluster of edge devices including a first edge device;

receiving, by the first edge device in the spatial network community, a first task into a local task queue of tasks to be performed by the first edge device;

assigning, by the first edge device, a first class to the first task based on deadline requirements;

determining, by the local DRL network, to execute the first task by remote execution based on constraints represented by the assigned first class;

responsive to determining to execute the first task by remote execution, sending, by the first edge device, a request for help signal to a sub-set of edge devices in the spatial network community, the sub-set of edge devices having edge device profiles demonstrating characteristics aligned with the constraints represented by the assigned first class; and

offloading, by the first edge device, the first task to a selected edge device based on a dataset returned by the selected edge device responsive to the request for help signal.

2. The method of claim 1 , wherein the first class defines a delay-sensitivity level of the first task.

3. The method of claim 1 , further comprising:

initializing a second DNN for a remote DRL network including a plurality of local clusters including the first local cluster;

wherein:

the response to the request for help signal is performed by the remote DRL network;

the local DRL network is based on online deep policy gradient algorithm with importance sampling with reference to an old sample data of an existing policy; and

the remote DRL network is based on a DRL-based offloading schedule with the old sample data being replaced with new sample data for each training epoch.

4. The method of claim 1 , wherein:

responses from the sub-set of edge devices to the request for help signal include edge device state information for each responding edge device; and

selecting the select edge device based on the edge device state information provided by the select edge device.

5. The method of claim 1 , further comprising:

identifying the sub-set of edge devices according to geographic proximity to the first edge device, the assigned first class indicating the first task is a delay-sensitive task; and

analyzing, for each responsive edge device, respective responses to the request for help signal;

wherein:

the select edge device is selected from among the sub-set of edge devices by application of a cost function with respect to minimizing cost of the first edge device in view of the select edge device.

6. The method of claim 5 , wherein:

the sub-set of edge devices and the first edge device are each at a series of road traffic signals, each road traffic signal including one of the edge devices, and the series of road traffic signals making up the spatial network community; and

the geographic proximity is within the series of traffic signals at a traffic flow intersection where the first edge device is located.

7. The method of claim 1 , wherein:

the first class is one of three classes of tasks, delay-sensitive tasks, delay-constrained tasks, and delay-tolerant tasks; and

offloading tasks to edge devices is based, at least in part, on characteristics of the edge device as defined by the three classes of tasks, delay-sensitive tasks requiring performance within a geographic cluster of edge devices, delay-constrained tasks permitting performance by an edge server within the spatial community of edge devices, and the delay-tolerant tasks permitting performance by a cloud server.

8. A computer program product for edge device self-orchestration in a spatial network community, the computer program product comprising a non-transitory computer readable storage medium having a set of instructions stored therein which, when executed by a processor, causes the processor to perform a method comprising:

initializing a first deep neural network (DNN) for a local deep reinforcement learning (DRL) network of a first local cluster of edge devices including a first edge device;

receiving, by the first edge device in the spatial network community, a first task into a local task queue of tasks to be performed by the first edge device;

assigning, by the first edge device, a first class to the first task based on deadline requirements;

determining, by the local DRL network, to execute the first task by remote execution based on constraints represented by the assigned first class;

responsive to determining to execute the first task by remote execution, sending, by the first edge device, a request for help signal to a sub-set of edge devices in the spatial network community, the sub-set of edge devices having edge device profiles demonstrating characteristics aligned with the constraints represented by the assigned first class; and

offloading, by the first edge device, the first task to a selected edge device based on a dataset returned by the selected edge device responsive to the request for help signal.

9. The computer program product of claim 8 , wherein the first class defines a delay-sensitivity level of the first task.

10. The computer program product of claim 8 , further causing the processor to perform a method comprising:

initializing a second DNN for a remote DRL network including a plurality of local clusters including the first local cluster;

wherein:

the response to the request for help signal is performed by the remote DRL network;

the local DRL network is based on online deep policy gradient algorithm with importance sampling with reference to an old sample data of an existing policy; and

the remote DRL network is based on a DRL-based offloading schedule with the old sample data being replaced with new sample data for each training epoch.

11. The computer program product of claim 8 , wherein:

responses from the sub-set of edge devices to the request for help signal include edge device state information for each responding edge device; and

selecting the select edge device based on the edge device state information provided by the select edge device.

12. The computer program product of claim 8 , further causing the processor to perform a method comprising:

identifying the sub-set of edge devices according to geographic proximity to the first edge device, the assigned first class indicating the first task is a delay-sensitive task; and

analyzing, for each responsive edge device, respective responses to the request for help signal;

wherein:

the select edge device is selected from among the sub-set of edge devices by application of a cost function with respect to minimizing cost of the first edge device in view of the select edge device.

13. The computer program product of claim 12 , wherein:

the sub-set of edge devices and the first edge device are each at a series of road traffic signals, each road traffic signal including one of the edge devices, and the series of road traffic signals making up the spatial network community; and

the geographic proximity is within the series of traffic signals at an traffic flow intersection where the first edge device is located.

14. A computer system for edge self-orchestration in a spatial network community, the computer system comprising:

a processor set; and

a computer readable storage medium;

wherein:

the processor set is structured, located, connected, and/or programmed to run program instructions stored on the computer readable storage medium; and

the program instructions which, when executed by the processor set, cause the processor set to perform a method comprising:

initializing a first deep neural network (DNN) for a local deep reinforcement learning (DRL) network of a first local cluster of edge devices including a first edge device;

receiving, by the first edge device in the spatial network community, a first task into a local task queue of tasks to be performed by the first edge device;

assigning, by the first edge device, a first class to the first task based on deadline requirements;

determining, by the local DRL network, to execute the first task by remote execution based on constraints represented by the assigned first class;

responsive to determining to execute the first task by remote execution, sending, by the first edge device, a request for help signal to a sub-set of edge devices in the spatial network community, the sub-set of edge devices having edge device profiles demonstrating characteristics aligned with the constraints represented by the assigned first class; and

offloading, by the first edge device, the first task to a selected edge device based on a dataset returned by the selected edge device responsive to the request for help signal.

15. The computer system of claim 14 , wherein the first class defines a delay-sensitivity level of the first task.

16. The computer system of claim 14 , further causing the processor to perform a method comprising:

initializing a second DNN for a remote DRL network including a plurality of local clusters including the first local cluster;

wherein:

the response to the request for help signal is performed by the remote DRL network;

the local DRL network is based on online deep policy gradient algorithm with importance sampling with reference to an old sample data of an existing policy; and

the remote DRL network is based on a DRL-based offloading schedule with the old sample data being replaced with new sample data for each training epoch.

17. The computer system of claim 14 , wherein:

responses from the sub-set of edge devices to the request for help signal include edge device state information for each responding edge device; and

selecting the select edge device based on the edge device state information provided by the select edge device.

18. The computer system of claim 14 , further causing the processor to perform a method comprising:

identifying the sub-set of edge devices according to geographic proximity to the first edge device, the assigned first class indicating the first task is a delay-sensitive task; and

analyzing, for each responsive edge device, respective responses to the request for help signal;

wherein:

the select edge device is selected from among the sub-set of edge devices by application of a cost function with respect to minimizing cost of the first edge device in view of the select edge device.

19. The computer system of claim 18 , wherein:

the sub-set of edge devices and the first edge device are each at a series of road traffic signals, each road traffic signal including one of the edge devices, and the series of road traffic signals making up the spatial network community; and

the geographic proximity is within the series of traffic signals at an traffic flow intersection where the first edge device is located.

20. The computer system of claim 14 , wherein:

the first class is one of three classes of tasks, delay-sensitive tasks, delay-constrained tasks, and delay-tolerant tasks; and

offloading tasks to edge devices is based, at least in part, on characteristics of the edge device as defined by the three classes of tasks, delay-sensitive tasks requiring performance within a geographic cluster of edge devices, delay-constrained tasks permitting performance by an edge server within the spatial community of edge devices, and the delay-tolerant tasks permitting performance by a cloud server.