International Business Machines Corporation

Junfeng Liu

Kuan Feng

Zhichao Su

Khalid Ahmed

Hai Long W Wen

Yi Zhao

Qing Xu

Xiao Ming Bao

King Shing K Lui

Su Lei

Embodiments for efficient machine and deep learning hyperparameter tuning in a distributed computing system. Runtime metrics of each training iteration are collected to identify candidate jobs to merge during an execution phase. The candidate jobs are grouped into job groups, and the job groups containing the candidate jobs are merged together subsequent to each iteration boundary for execution during the execution phase.

Job merging for machine and deep learning hyperparameter tuning

Embodiments for accelerator allocation and reclamation for deep learning jobs in a computing cluster. Metrics are recorded of each accelerator of a set of accelerators allocated to a deep learning job including computing a gain of computational power by an additional allocation of new accelerators and computing a cost of transferring data among the new accelerators and the set of allocated accelerators. Ones of the new accelerators are allocated to the deep learning job or ones of the set of allocated accelerators assigned to perform the deep learning job are reclaimed upon determining an optimal accelerator topology by comparing the gain of computation power and the cost of transferring data.

Smart accelerator allocation and reclamation for deep learning jobs in a computing cluster

Embodiments for dynamic accelerator scheduling and grouping for deep learning jobs in a computing cluster. An efficiency metric of each job executing in the computing cluster is calculated to generate a prioritized job queue. Accelerator re-grouping execution plans are then generated based on the prioritized job queue, the accelerator re-grouping execution plans associated with a target cluster topology to be achieved according to the placement of selected jobs from the prioritized job queue in relation to a location of respective ones of a plurality of accelerators within the computing cluster. One of the accelerator re-grouping execution plans is executed to allocate the selected jobs to the respective ones of the plurality of accelerators to thereby shift the computing cluster to the target cluster topology.

Dynamic accelerator scheduling and grouping for deep learning jobs in a computing cluster

Embodiments for leveraging directed acyclic graph (DAG) information to group tasks for execution, by at least one processor. An Input/Output (I/O) cost for each task in a set of tasks determined to be within a same input host set is determined, and tasks within the set of tasks are grouped into task groups according to an applied allocation time. The task groups are launched commensurate with a calculated delay time from each input host set so as to overlap I/O phases and central processing unit (CPU) phases to improve CPU efficiency. The launched task groups are executed by an executor according to the calculated delay time.

Leveraging directed acyclic graph (DAG) information to group tasks for execution

Embodiments for graphical processing unit (GPU) resource sharing in a computing cluster, by a processor device. Resource-specific stages are dynamically generated in a directed acyclic graph (DAG) using a DAG interpreter for a set of tasks by creating equivalence stages in the DAG having an associated inserted set of shuffle stages, the equivalence stages created based on a determined cost of each stage of the set of shuffle stages. Backlog tasks are scheduled and tasks within the set of tasks are shifted among respective stages in the equivalence stages according to the determined cost to avoid overlapping allocation of the GPU resource during central processing unit (CPU) execution of the respective tasks of the set of tasks.

Cluster graphical processing unit (GPU) resource sharing efficiency by directed acyclic graph (DAG) generation

Embodiments for leveraging directed acyclic graph (DAG) information to group tasks for execution, by at least one processor device. For a set of tasks, an input host set is determined for each task using a shuffle type and hosts used during a previous stage, and the tasks in the set of tasks determined to be within a same input host set are classified together. An Input/Output (I/O) cost for each task in the set of tasks is calculated and tasks within the set of tasks are grouped into task groups according to an applied allocation time. The task groups are launched commensurate with a calculated delay time from each input host set so as to overlap I/O phases and central processing unit (CPU) phases to improve CPU efficiency, the launched task groups being executed by an executor according to the calculated delay time.

Growing community of inventors

Markham, Canada

Junfeng Liu