International Business Machines Corporation

Platform Computing Corporation

Yonggang Hu

Zhenhua Hu

Alicia Elena Chin

Rui Zhang

Prasenjit Sarkar

Reshu Jain

Jason T S Lam

Zhimin Lin

Onkar S Parmar

Ajith Shanmuganathan

Li Feng Zhang

Xiaoqiao Meng

Chao Xue

Jian Tan

Shicong Meng

Amir Asif

Vida Movahedi

Ke Wei Sun

Rong Yan

Yonghua Lin

Yu Song

Michael Feiman

Lin Dong

Xiaoxing Wang

A one-shot neural architecture search method referred to as MergeNAS by merging different types of convolutions into a single operation. This mergence approach reduces the search cost to roughly half a GPU-day as well as alleviates the over-fitting problem caused by a traditional differentiable architecture search (DARTS) approach by reducing the number of redundant parameters.

Merge operations for darts

Aspects of the invention include systems and methods to obtain meta features of a dataset for training in a deep learning application. A method includes selecting an initial search space that defines a type of deep learning architecture representation that specifies hyperparameters for two or more neural network architectures. The method also includes applying a search strategy to the initial search space. One of the two or more neural network architectures are selected based on a result of an evaluation according to the search strategy. A new search space is generated with new hyperparameters using an evolutionary algorithm and a mutation type that defines one or more changes in the hyperparameters specified by the initial search space, and, based on the mutation type, the new hyperparameters are applied to the one of the two or more neural networks or the search strategy is applied to the new search space.

Search space exploration for deep learning

Methods and systems for training a machine learning model include training a machine learning model using training data. A status of the machine learning model's training is determined based on an accuracy curve of the machine learning model over the course of the training. Parameters of the training are adjusted based on the status. Training of the machine learning model is completed using the adjusted parameters.

Automatic monitoring and adjustment of machine learning model training

A method for multi-layer quality of service (QoS) management in a distributed computing environment includes receiving a workload and identifying a workload QoS class for the workload. The workload is scheduled to be run compute nodes. The workload QoS class is translated to a plurality of different resource layer QoS classes, and the plurality of QoS classes are communicated to a workload execution manager of the compute nodes. The workload execution manager extends the plurality of different resource layer QoS classes to respective resource layers involved in the running of the workload.

Multi-layer QOS management in a distributed computing environment

A system for multi-layer quality of service (QoS) management in a distributed computing environment includes a management node hosting a workload scheduler configured to receive a workload and identify a workload QoS class for the workload. The workload scheduler schedules running of the workload on compute nodes, translate the workload QoS class to a plurality of different resource layer QoS classes, and communicates the plurality of QoS classes to a workload execution manager of the compute nodes. The workload execution manager is configured to extend the plurality of different resource layer QoS classes to respective resource layers involved in the running of the workload.

Multi-layer QoS management in a distributed computing environment

A method, computer program product, and system for optimizing service pools supporting resource sharing and enforcing SLAs, to minimize service restart. A computer processor determines a first resource to be idle, wherein a service instance continues to occupy the first resource that is idle. The processor adds the first resource to a resource pool, wherein the service instance continues to occupy the first resource as a global standby service instance on the first resource. The processor receives a request for a resource, wherein the request for the resource includes a global name associated with a service that corresponds to the global standby service instance, and the processor allocates, from the resource pool, the first resource having the global standby service instance, based on the request for the resource that includes the global name associated with the service corresponding to the global standby service instance.

Minimizing service restart by optimally resizing service pools

A computer-implemented method, according to one embodiment, includes: splitting a video stream into a plurality of sub-streams, and encoding each of the sub-streams. A packet is also transmitted to a receiver, where the packet includes a number of the encoded sub-streams. Moreover, the number of the encoded sub-streams that are included in the packet transmitted to the receiver corresponds to a quality of service associated with a service cost incurred by the receiver. Other systems, methods, and computer program products are described in additional embodiments.

Video encoding and transcoding for multiple simultaneous qualities of service

A single workload scheduler schedules sessions and tasks having a tree structure to resources, wherein the single workload scheduler has scheduling control of the resources and the tasks of the parent-child workload sessions and tasks. The single workload scheduler receives a request to schedule a child session created by a scheduled parent task that when executed results in a child task; the scheduled parent task is dependent on a result of the child task. The single workload scheduler receives a message from the scheduled parent task yielding a resource based on the resource not being used by the scheduled parent task, schedules tasks to backfill the resource, and returns the resource yielded by the scheduled parent task to the scheduled parent task based on receiving a resume request from the scheduled parent task or determining dependencies of the scheduled parent task have been met.

Computing session workload scheduling and management of parent-child tasks using a blocking yield API to block and unblock the parent task

A system for multi-layer quality of service (QoS) management in a distributed computing environment includes a management node hosting a workload scheduler operable to receive a workload and identify a workload QoS class for the workload. The system also includes a plurality of distributed compute nodes. A workload scheduler schedules running of the workload on the compute nodes, translates the workload QoS class to a storage level QoS class, and communicates the storage level QoS class to a workload execution manager of the compute nodes. The workload scheduler communicates the storage level QoS class to one or more storage managers where the storage managers manage storage resources, and the storage managers are operable to extend the storage level QoS class to the storage resources to support the workload QoS class.

A technique for multi-layer quality of service (QoS) management in a distributed computing environment includes receiving a workload to run in a distributed computing environment. A workload quality of service (QoS) class for the workload is identified, and the workload QoS class is translated to a storage level QoS class. The workload is scheduled to run on a compute node of the environment, and the storage level QoS class is communicated to a workload execution manager of the compute node. The storage level QoS class is communicated to one or more storage managers of the environment where the storage managers manage storage resources in the environment. The storage managers extend the storage level QoS class to the storage resources to support the workload QoS class.

A computer-implemented method, according to one embodiment, includes: splitting a video stream into a plurality of sub-streams. Each of the sub-streams individually provide a unique version of the video stream having a lower video quality. However, merging more than one of the sub-streams provides a merged version of the video stream having a relatively higher video quality. Each of the sub-streams are also encoded. A packet which includes a number of the encoded sub-streams is transmitted to a receiver, where the number of the encoded sub-streams included in the packet transmitted to the receiver corresponds to a quality of service associated with a service cost incurred by the receiver. Other systems, methods, and computer program products are described in additional embodiments.

A computer processor allocates a service instance to a first resource of a resource pool, in which the service instance corresponds to performing a first task and for the service instance to continue to occupy the first resource subsequent to task completion, and determines the first resource to be idle. The first resource is returned to a resource pool, with the service instance occupying the first resource as a global standby service instance of a first service. In response to a request of a resource to perform a task from a second application, in which the request for the resource includes a global name associated with the first service of the global standby service instance, occupying the first resource returned to the resource pool, allocating from the resource pool, the first resource occupied by the global standby service instance, based on the global name associated with the first service.

A computer-implemented method, according to one embodiment, includes: splitting a video stream into a plurality of sub-streams, encoding each of the sub-streams in parallel, and transmitting a packet to a receiver, wherein the packet includes a number of the encoded sub-streams. Each of the sub-streams individually provide a unique version of the video stream having a lower video quality, while merging more than one of the sub-streams provides a merged version of the video stream having a relatively higher video quality. Moreover, the number of the encoded sub-streams included in the packet transmitted to the receiver corresponds to a quality of service associated with the receiver. Other systems, methods, and computer program products are described in additional embodiments.

A system for multi-layer quality of service (QoS) management in a distributed computing environment includes a management node hosting a workload scheduler operable to receive a workload and identify a workload QoS class for the workload. The system also includes a plurality of distributed compute nodes. A workload scheduler is operable to schedule running of the workload on the compute nodes. The workload scheduler is operable to translate the workload QoS class to a storage level QoS class and communicate the storage level QoS class to a workload execution manager of the compute nodes. The workload scheduler communicates the storage level QoS class to one or more storage managers where the storage managers manage storage resources, and the storage managers are operable to extend the storage level QoS class to the storage resources to support the workload QoS class.

Growing community of inventors

Richmond Hill, Canada

Yonggang Hu