ABSTRACT:
The degree of parallelism in emerging high-end systems is increasing dramatically. This is a multiplicative effect due to simultaneous trends towards System architectures with much larger numbers of nodes, multi-socket node architectures, and multi-core/multi-thread chips. The current generation of performance tools were designed for simpler systems; we believe that new approaches are needed to deal with the complexity of this degree of parallelism.
We will describe ongoing work at RENCI to address these issues. On a multi-core, multi-socket environment, the interactions of threads (cores) competing for shared resources is dominating
performance. We are therefore working on an introspective, resource-centric approach to performance monitoring. Combined with on-line adaptation in the operating system and in a user-level runtime system, this approach has the promise in most cases to make intelligent
choices between power consumption and throughput. In some cases, both power consumption and performance can be improved through adaptation. At very large scale, the volume and complexity
of data dictate that tools themselves become parallel. We are therefore working on parallel, in situ filtering and analysis to address the detection and diagnosis of performance problems at scale. We will present preliminary results for both domains.
BIO:
Rob Fowler directs RENCI's high performance computing research projects, including efforts to analyze the effectiveness of high-end systems in serving the needs of scientists and to develop software tools that enhance the performance of grid-enabled applications.
# # #