The present disclosure discloses a method, an apparatus and a storage medium for training a deep learning framework, and relates to the artificial intelligence field such as deep learning and big data processing. The specific implementation solution is: acquiring at least one task node in a current task node cluster, that meets a preset opening condition when a target task meets a training start condition; judging whether a number of nodes of the at least one task node is greater than or equal to a preset number; synchronously training the deep learning framework of the target task by the at least one task node according to sample data if the number of nodes is greater than the preset number; and acquiring a synchronously trained target deep learning framework when the target task meets a training completion condition.

The present disclosure relates to systems and methods for sharing compute resources. The systems and methods may include identifying a plurality of workloads to complete by a deadline. The systems and methods may include generating a performance prediction for each workload of the plurality of workloads. The systems and methods may use the performance prediction to calculate a number of compute resources required for the plurality of workloads to complete by the deadline. The systems and methods may schedule the plurality of workloads across the number of compute resources.

Workloads are scheduled on a common set of resources distributed across a cluster of hosts using at least two schedulers that operate independently. The resources include CPU, memory, network, and storage, and the workloads may be virtual objects, including VMs, and also operations including live migration of virtual objects, network file copy, reserving spare capacity for high availability restarts, and selecting hosts that are to go into maintenance mode. In addition, the at least two independent schedulers are assigned priorities such that the higher priority scheduler is executed to schedule workloads in its inventory on the common set of resources before the lower priority scheduler is executed to schedule workloads in its inventory on the common set of resources.

A processing system is configured to dynamically carry out processes. A method for monitoring the processing system includes steps of determining a number of processes running on the processing system; of determining a maximum expected number of processes; of determining that more processes than expected are running; and of deactivating the processing system.

A system according one exemplary embodiment may receive instructions regarding a callee function that include a description of properties associated with arguments for the callee function, create a persistent representation of the callee function based on the received description of the callee function, receive instructions from a source describing a set of properties for a data collection, create a persistent representation of the data collection based on the set of properties for the data collection, receive an updated property for the data collection, and in response to receiving the updated property for the data collection: update the persistent representation of the callee function by associating the persistent representation of the callee function with the updated property of the data collection and update the persistent representation of the data collection by associating the persistent representation the data collection with the updated property of the data collection.

A deep neural network (DNN) module utilizes parallel kernel and parallel input processing to decrease bandwidth utilization, reduce power consumption, improve neuron multiplier stability, and provide other technical benefits. Parallel kernel processing enables the DNN module to load input data only once for processing by multiple kernels. Parallel input processing enables the DNN module to load kernel data only once for processing with multiple input data. The DNN module can implement other power-saving techniques like clock-gating (i.e. removing the clock from) and power-gating (i.e. removing the power from) banks of accumulators based upon usage of the accumulators. For example, individual banks of accumulators can be power-gated when all accumulators in a bank are not in use, and do not store data for a future calculation. Banks of accumulators can also be clock-gated when all accumulators in a bank are not in use, but store data for a future calculation.

A scheduler, a method of operating the scheduler, and an accelerator apparatus including the scheduler are disclosed. A method of operating a scheduler to perform scheduling on models to be executed in an accelerator, the method includes receiving at least one execution request for a first model and a second model that are executed independently from each other in the accelerator, and performing layer-unit scheduling on the first model and the second model based on workload characteristics of the first model and the second model.

A method for controlling transactional processing system having transactions that include multiple tasks, a throughput limit a transaction processing time limit includes allocating a plurality of threads to be used by multiple tasks to achieve a throughput approximating the throughput limit. The method assigns the multiple tasks to the plurality of threads and assigns respectively different processing delays to the plurality of threads. The processing delays span an interval less than the transaction processing time limit. The method processes the multiple tasks within the transaction processing time limit by executing the plurality of threads at times determined by the respective processing delays.

Optimized memory usage and management is crucial to the overall performance of a neural network (NN) or deep neural network (DNN) computing environment. Using various characteristics of the input data dimension, an apportionment sequence is calculated for the input data to be processed by the NN or DNN that optimizes the efficient use of the local and external memory components. The apportionment sequence can describe how to parcel the input data (and its associated processing parameters—e.g., processing weights) into one or more portions as well as how such portions of input data (and its associated processing parameters) are passed between the local memory, external memory, and processing unit components of the NN or DNN. Additionally, the apportionment sequence can include instructions to store generated output data in the local and/or external memory components so as to optimize the efficient use of the local and/or external memory components.

Provided is a parallelized and modular planning system that controls multiple actors in performing different tasks simultaneously without conflict based on plans that are modularly created and/or updated. The system may update different parts of different plans at the same time without redefining those plans anew. The system may include a first subsystem that generates a first plan based on an assignment of a first task to a first actor, while a second subsystem provides a path for a second plan assigned to a second actor, while a third subsystem determines access ordering by which a third actor accesses a first resource from a path of a third plan, and while a fourth subsystem controls operation of a fourth actor in accessing a second resource from a path of a fourth plan based on an access ordering determined for the second resource by the third subsystem.