International Business Machines Corporation

Karl Michael Solie

Mitchell Dennis Felton

Ray LeRoy Lucas

David W Alderman

A method and apparatus is described for dynamic relocation of a job executing on multiple nodes of a high performance computing (HPC) systems. The job is dynamically relocated when the messaging network is in a quiescent state. The messaging network is quiesced by signaling the job to suspend execution at a global collective operation of the job where the messaging of the job is known to be in a quiescent state. When all the nodes have reached the global collective operation and paused, the job is relocated and execution is resumed at the new location.

Dynamic job relocation in a high performance computing system

The exemplary embodiments of the present invention provide a method for efficiently identifying the bad component(s) in a multi-node system. The method includes assigning a unique ID to each of a plurality of nodes on the multi-node system, generating test statistics from a test on a plurality of nodes, and comparing the test statistics for the plurality of nodes against a first topology to generate a first number of clusters of bad nodes. The method further includes comparing the test statistics for the plurality of nodes against a second topology to generate a second number of clusters of bad nodes, and identifying the bad nodes by comparing the cluster sizes to a topology threshold.

Growing community of inventors

Rochester, MN, United States of America

Karl Michael Solie