Adaptive resource allocation using neural networks

• US 5,704,012 A
• Filed: 05/31/1995
• Issued: 12/30/1997
• Est. Priority Date: 10/08/1993
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Bigus, Joseph Phillip
• Primary Examiner(s): Hafiz, Tariq R.
• Assistant Examiner(s): Rhodes, Jason W.
• Application Number:
  US08/454,977
• Time in Patent Office: 944 Days
• Field of Search: 395/22, 395/20, 395/21, 395/24, 395/27, 382/155, 382/156, 382/157, 382/158, 382/159
• US Class Current: 706/19
• CPC Class Codes: G06F 2209/501 Performance criteria
• CPC Class Codes: G06F 2209/5017 Task decomposition
• CPC Class Codes: G06F 9/5011 the resources being hardware resources other than CPUs, Servers and Terminals

Abstract

In a system comprising a plurality of resources for performing useful work, a resource allocation controller function, which is customized to the particular system'"'"'s available resources and configuration, dynamically allocates resources and/or alters configuration to accommodate a changing workload. Preferably, the resource allocation controller is part of the computer'"'"'s operating system which allocates resources of the computer system. The resource allocation controller uses a controller neural network for control, and a separate system model neural network for modelling the system and training the controller neural network. Performance data is collected by the system and used to train the system model neural network. A system administrator specifies computer system performance targets which indicate the desired performance of the system. Deviations in actual performance from desired performance are propagated back through the system model and ultimately to the controller neural network to create a closed loop system for resource allocation.

Claims

1. A method for controlling the response of a computer system to a workload and configuration, comprising the steps of:

gathering performance data for jobs in a plurality of job classes from said computer system, said performance data including workload data and configuration data for a plurality of time intervals, and further including computer system response data for said plurality of intervals, wherein the jobs in the plurality of job classes require different amounts of computer system resources;

constructing a neural network, said neural network having a set of inputs corresponding to the workload and configuration of said computer system, and having at least one output corresponding to response of said computer system;

training said neural network with said performance data gathered from said computer system to produce a trained neural network model of said computer system;

determining the response of said computer system from the output of said trained neural network; and

allocating the resources in said computer system among the plurality of job classes based on said response from said determining step and based on performance objectives for each job class specified by a user of said computer system.