Abstract: The disclosure herein describes elastically managing the execution of workers of multi-worker workloads on accelerator devices. A first worker of a workload is executed on an accelerator device during a first time interval. A first context switch point is identified when the first worker is in a first worker state. At the identified context switch point, a first memory state of the first worker is stored in a host memory and the accelerator device is configured to a second memory state of the second worker. The second worker is executed during a second time interval and a second context switch point is identified at the end of the second time interval when the second worker is in a state that is equivalent to the first worker state. During the intervals, collective communication operations between the workers are accumulated and, at the second context switch point, the accumulated operations are performed.

Abstract: A refinement system automatically identifies whether a detected error in a target system during abstract interpretation is a false error or a true error and adjusts the interpretation to prevent the false error. The target system is represented as a transition system with an initial state and state transitions and a specification that the target system is to satisfy. The refinement system iteratively performs steps of the abstract interpretation using a widening operator. When the state of a step does not satisfy the specification, the refinement system identifies a step whose widening operator was the source of the state that did not satisfy the specification and applies a more precise operator that eliminates the problem with the widening. The refinement system then starts re-performing the steps starting at that step.

Abstract: A refinement system automatically identifies whether a detected error in a target system during abstract interpretation is a false error or a true error and adjusts the interpretation to prevent the false error. The target system is represented as a transition system with an initial state and state transitions and a specification that the target system is to satisfy. The refinement system iteratively performs steps of the abstract interpretation using a widening operator. When the state of a step does not satisfy the specification, the refinement system identifies a step whose widening operator was the source of the state that did not satisfy the specification and applies a more precise operator that eliminates the problem with the widening. The refinement system then starts re-performing the steps starting at that step.