International Business Machines Corporation

Charles E Hackett

Andrew M Sica

Michael C Cox

Kathleen M Walsh

Daniel Nieves

Dominic Schroeder

Christopher Chappell

John C Goodyear

Edward A King

Processing within a computing environment is facilitated by obtaining, by a decision tree inference accelerator, a decision node result vector based on processing an input data vector through decision nodes of a decision tree. Further, the processing is facilitated by accelerating, by the decision tree inference accelerator, leaf node processing of the decision node result vector. The decision tree inference accelerator facilitates, based on predetermined path vectors through the decision tree to each leaf node of the plurality of leaf nodes, processing of the decision node result vector by the plurality of leaf nodes in parallel, and determining therefrom which leaf node of the plurality of leaf nodes is selected.

Parallel inference processing by decision tree leaf nodes

An aspect of the invention includes receiving machine learning (ML) training data that includes a plurality of features for a plurality of observations. The ML training data is broken into a plurality of non-overlapping subsets of features and observations. A first ML algorithm is trained based on a first subset of the features and observations, and a second ML algorithm is trained based on a second subset of the features and observations. The training of the first ML algorithm overlaps in time with the training of the second ML algorithm. The first and second ML algorithms are tested. Either the first or second ML algorithm is selected based at least in part on results of the testing. The selected ML algorithm is retained as a trained ML algorithm for predicting one or more of the plurality of features based on one or more others of the plurality of features.

Parallel ensemble of machine learning algorithms

A computer-implemented method includes monitoring, by an agent, resource consumption at a system running at an on-premises data center, where the agent is embedded into the system. Compliance data is determined by measuring, by the agent, the resource consumption according to one or more cost factors of a consumption-based pricing agreement. The compliance data is useable to determine a consumption-based price of operating the system according to the consumption-based pricing agreement.

Ensuring compliance in on-premise compute models

Embodiments include method, systems and computer program products for operating a computer system using a workload manager to control a dynamic thread mode switch. The method includes receiving a plurality of workloads. A workload manager determines which of the plurality of workloads are eligible to use a processor core operating in a multi-threaded mode and which workloads are response sensitive and use a processor core operating in a single-thread mode. A hypervisor determines a state of operation for one or more cores of one or more processors and causes a core switch for at least one or more cores in response to a determination that workloads eligible to use a processor core operating in a multi-threaded mode exist or workloads that are response sensitive and a determination that the state of operation for one or more cores of one or more processors should change to facilitate handling the associated workloads.

Workload manager control of dynamic thread mode switch

Systems, methods, and computer-readable media for building ensemble members of a Support Vector Machine (SVM) ensemble in parallel and executing processing in parallel on data allocated to each ensemble member are disclosed. The parallel construction and processing of data of each ensemble member allows a single large SVM calculation to be replaced with many smaller SVM calculations performed in parallel, and thus, may reduce the computational resources required to classify datasets.

Parallel ensemble of support vector machines

In one embodiment, the invention provides a method of configuring a migration from a plurality of distributed virtual servers to a system of blades, the method comprising: identifying among the plurality of distributed virtual servers a plurality of candidate virtual servers for inclusion in the migration; collecting performance data for each candidate virtual server; consolidating the performance data for each candidate virtual server into a usage profile; normalizing the usage profile to allow representation of a statistical measure of resource usage by the candidate virtual server; and constructing a configuration for the plurality of candidate virtual servers within the system of blades.

Growing community of inventors

Greenlawn, NY, United States of America

Charles E Hackett