International Business Machines Corporation

Kyndryl, Inc.

James E Olson

Ramani Ranjan Routray

Gabriel Alatorre

Aameek Singh

Mu Qiao

Ann Corrao

John Jason Auvenshine

Yang Song

Bernhard Julius Klingenberg

Stanley C Wood

Nagapramod S Mandagere

Sunhwan Lee

Lokesh M Gupta

Matthew G Borlick

Micah Robison

Rakesh Jain

Richard P Oubre, Jr

Usman Ahmed

Sandeep Gopisetty

Kavita Chavda

Laura A Richardson

David P Goodman

Larry Steven McGimsey

Richard Harry Hopkins

Beth Ann Peterson

Juan A Coronado

Eric Kevin Butler

Lisa R Martinez

Pritpal Arora

Laxmikantha Sai Nanduru

Mark V Chitti

Anil Kumar Narigapalli

Alan C Skinner

Mark Vincent Chitti

Richard H Hopkins

Barry B Becker, Jr

Barry Becker

A machine learning module is trained by receiving inputs comprising attributes of a computing environment, where the attributes affect a likelihood of failure in the computing environment. In response to an event occurring in the computing environment, a risk score that indicates a predicted likelihood of failure in the computing environment is generated via forward propagation through a plurality of layers of the machine learning module. A margin of error is calculated based on comparing the generated risk score to an expected risk score, where the expected risk score indicates an expected likelihood of failure in the computing environment corresponding to the event. An adjustment is made of weights of links that interconnect nodes of the plurality of layers via back propagation to reduce the margin of error, to improve the predicted likelihood of failure in the computing environment.

Perform preemptive identification and reduction of risk of failure in computational systems by training a machine learning module

A computer-implemented method, according to one embodiment, includes logically partitioning a storage system into a plurality of compartment constructs, and mapping hosts in communication with the storage system to the compartment constructs, thereby enabling interoperability among the hosts and the compartment constructs. The interoperability of the hosts and the compartment constructs is analyzed, and the interoperability is based on storage software and/or firmware versions being run by the hosts. The method further includes defining, based on the analysis, risk profiles for applications run on the hosts, and determining, based on the risk profiles, recommendations for assignment and mapping of the hosts with the compartment constructs. Ownership of storage objects is assigned to the compartment constructs based on the recommendations. Each of the storage objects define a logical partition of one of the hosts and a logical partition of a storage volume of the storage system.

Dynamically assigning storage objects to compartment constructs of a storage system to reduce application risk

Input on a plurality of attributes of a computing environment is provided to a machine learning module to produce an output value that comprises a risk score that indicates a likelihood of a potential malfunctioning occurring within the computing environment. A determination is made as to whether the risk score exceeds a predetermined threshold. In response to determining that the risk score exceeds a predetermined threshold, an indication is transmitted to indicate that potential malfunctioning is likely to occur within the computing environment. A modification is made to the computing environment to prevent the potential malfunctioning from occurring.

Using a machine learning module to perform preemptive identification and reduction of risk of failure in computational systems

A method for distributing data among storage devices. The method comprising one or more processors receiving a first graph workload that executes within a networked computing environment. The method further includes identifying data from the first graph workload that is utilized during the execution of the first graph workload that includes a plurality of data packets. The method further includes creating a first graph workload model representative of the graph structure of the first graph workload and determining two or more partitions that are representative of a distribution of the identified data utilized by the first graph workload based, at least in part, on the first graph workload model. The method further includes allocating a plurality of network accessible storage devices among the two or more partitions and copying a first set of data packets of the plurality of data packets to a network accessible storage device.

Assigning storage locations based on a graph structure of a workload

A method and associated systems for optimized orchestration of a data-migration project. A data-migration orchestration system represents a hierarchical organization of each dataset to be migrated as a tree, where each leaf node of the tree represents data to be migrated and where a path between the leaf node and the root node represents a hierarchical directory pathname of sensitive data represented by the leaf node. Each tree is assigned a sensitivity signature that is proportional to the relative sensitivity and access frequency of the dataset represented by that tree. The signatures are organized into clusters as a function of the distances between each signature, and each signature is associated with a soft migration cost specific to that signature's cluster. A soft cost for migrating an application that requires multiple datasets may be determined by adding the migration costs associated with each of the multiple datasets.

Optimized orchestration of complex application-migration projects based on data sensitivity

A method and associated systems for a workload-aware thin-provisioning system that allocates physical storage to virtual resources from pools of physical storage volumes. The system receives constraints that limit the amount of storage that can be allocated from each pool and the total workload that can be directed to each pool. It also receives lists of previous workloads and allocations associated with each volume at specific times in the past. The system then predicts future workloads and allocation requirements for each volume by regressing linear equations derived from the received data. If the predicted values indicate that a pool will at a future time violate a received constraint, the system computes the minimum costs to move each volume of the offending pool to a less-burdened pool. It then selects the lowest-cost combination of volume and destination pool and then moves the selected volume to the selected pool.

Workload-aware storage-allocation system

A method and associated systems for identifying and correcting suboptimal storage-reclamation processes. A storage-management system uses information received in system-generated storage-reclamation reports to assign each user a set of reclamation scores. Each score identifies how effectively the user has been able to reclaim lost storage at particular times. These scores are organized into user-specific profiles that each consists of a chronological sequence of one user's scores. If a user's profile is 'good' (that is, if the user's scores are consistently high) or 'improving' (if scores are increasing over time), the system then determines whether that user's reclamation efforts have successfully reduced the amount of reclaimable storage controlled by the user. If not, the system infers that a suboptimal storage-reclamation process interfered with the user's reclamation efforts. The system then undertakes corrective action to identify and resolve the cause of the suboptimal process.

Identifying and correcting suboptimal storage-reclamation processes

Optimized orchestration of data-migration projects with soft migration costs based on file-specific migration feasibilities

Low-cost storage-allocation system

Workload-aware thin-provisioning storage-allocation system

A method for inferring an application type, based on an operational characteristic I/O pattern of a storage volume. One or more processors determine at least one operational characteristic of each storage volume of a storage volume group associated with an application. One or more labels are assigned for each storage volume, based on the operational characteristics of each storage volume. At least one template is received that includes labels of storage volume characteristics of known application types. One or more processors infer a type of application associated with the storage volume group, based on a best-fit match of the aggregate labels of the storage volumes of the storage volume group to the labels included in the templates of storage volume characteristics of known application types.

Inferring application type based on input-output characteristics of application storage resources

Operational characteristic I/O patterns of each storage volume of a storage volume group, and at least one operational characteristic of each storage volume associated with an application, are determined. Unsupervised learning is used to form clusters of storage volumes of similar characteristics. Labels are generated, assigned, and aggregated for each characteristic of each storage volume. Templates are received that includes labels of storage volume characteristics of known application types. An application type associated with the storage volume group is inferred, based on a best-fit match of the aggregate labels of the storage volumes of the storage volume group to the labels included in the templates of storage volume characteristics of known application types.

An approach for reactive throttling of heterogeneous migration sessions running concurrently in a virtualized cloud environment is disclosed. The heterogeneous migration sessions are characterized by several attributes such as the type of function performed by the migration session and the value that is attained by performing the migration session. An input/output (I/O) path to any resource in the virtualized cloud environment computing that is experiencing excessive resource congestion is identified. Bottlenecks along the I/O path are then determined. Virtual-to-physical resource mapping information of resources in the virtualized cloud computing environment are used to obtain a list of the heterogeneous migration sessions contributing to the bottlenecks along the identified I/O path.

Reactive throttling of heterogeneous migration sessions in a virtualized cloud environment

Growing community of inventors

Seymour, CT, United States of America

James E Olson