Abstract: Generally disclosed herein is an approach for smart topology-aware link disabling and user job rescheduling strategies for online network repair of broken links in high performance networks used in supercomputers that are common in Machine Learning (ML) and High-Performance Computing (HPC) applications. While a disabled link is repaired online, user jobs may continue to run. The broken links may be detected as part of pre-flight checks before the user jobs run and/or during the job run time via a distributed failure detection and mitigation software stack which includes a centralized network controller and multiple agents running on each node. The network controller may ensure that the user jobs are rerouted to healthy links within the same network until the broken links are fixed and tested by the repair workflows, in which case the broken links are enabled again by the network controller for future user jobs.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media for performing preflight checks of a distributed computing system, are described. In one aspect, a method includes assigning a computing workload to a first subset of hardware accelerator machines each having one or more hardware accelerators. A preflight check on the first subset is performed before performing the computing workload to verify the functionality of each machine in the first subset. For each hardware accelerator machine of the first subset, a program code package is installed, including a task action based at least in part on characteristics of the computing workload. The task action including a sequence of operations is performed on the hardware accelerator machine to determine whether the task action fails. Whenever the task action fails, the computing workload is re-assigned to a second subset of hardware accelerator machines different from the first subset.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media for performing preflight checks of a distributed computing system, are described. In one aspect, a method includes assigning a computing workload to a first subset of hardware accelerator machines each having one or more hardware accelerators. A preflight check on the first subset is performed before performing the computing workload to verify the functionality of each machine in the first subset. For each hardware accelerator machine of the first subset, a program code package is installed, including a task action based at least in part on characteristics of the computing workload. The task action including a sequence of operations is performed on the hardware accelerator machine to determine whether the task action fails. Whenever the task action fails, the computing workload is re-assigned to a second subset of hardware accelerator machines different from the first subset.

Abstract: Generally disclosed herein is an approach for smart topology-aware link disabling and user job rescheduling strategies for online network repair of broken links in high performance networks used in supercomputers that are common in Machine Learning (ML) and High-Performance Computing (HPC) applications. While a disabled link is repaired online, user jobs may continue to run. The broken links may be detected as part of pre-flight checks before the user jobs run and/or during the job run time via a distributed failure detection and mitigation software stack which includes a centralized network controller and multiple agents running on each node. The network controller may ensure that the user jobs are rerouted to healthy links within the same network until the broken links are fixed and tested by the repair workflows, in which case the broken links are enabled again by the network controller for future user jobs.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media for performing preflight checks of a distributed computing system, are described. In one aspect, a method includes assigning a computing workload to a first subset of hardware accelerator machines each having one or more hardware accelerators. A preflight check on the first subset is performed before performing the computing workload to verify the functionality of each machine in the first subset. For each hardware accelerator machine of the first subset, a program code package is installed, including a task action based at least in part on characteristics of the computing workload. The task action including a sequence of operations is performed on the hardware accelerator machine to determine whether the task action fails. Whenever the task action fails, the computing workload is re-assigned to a second subset of hardware accelerator machines different from the first subset.