METHOD FOR SCHEDULING INTERFERENCE WORKLOADS ON EDGE NETWORK RESOURCES

摘要

A method for scheduling inference workloads on edge network resources by a scheduler (100) comprising a DNN agent (300) dispatching streams of queries (120) from a set of sources (110). A state St is sampled (1) from the environment comprising the incoming workloads (310) of queries (120) from the applications running in the sources (110), and the pool of resources of the cloud-edge network (320). The state St is forwarded (2) through the DNN agent (300) using reinforcement learning, which outputs (3) action At and performing action At on the environment contributes (4) to reward rt+1 obtained at the next step (5) of interaction by the RL-based DNN agent (300). The agent (300) receives an observation of the environment St at time t and, one step later, the reward rt+1 as a ratio of queries that are served successfully and determines the action At to perform, which, in part, yields to next state St+1.