2018 Ninth International Green and Sustainable Computing Conference (IGSC), 2018
Constraints on power consumption are having per-vasive effects on high-performance computing (HPC... more Constraints on power consumption are having per-vasive effects on high-performance computing (HPC) systems, the facilities in which they are housed, and the application codes themselves. These constraints are driven by a variety of reasons including physical limits on available power within a facility, or are due to utility demand response etc. It is essential that power management must now be added to the traditional HPC goals of monitoring and optimizing application and algorithm correctness, scalability, and performance. Irrespective of the causes for these constraints, a holistic approach must be taken to effectively manage power or schedule jobs more optimally so data centers can remain below these power constraints. In this paper we propose results of implementing several strategies for throttling power and its impact on application performance while providing insights for managing overall system energy levels.
We present an eecient design and implementation for the solution of the Sylvester Observer matrix... more We present an eecient design and implementation for the solution of the Sylvester Observer matrix equation on a distributed memory computer. It consists of two main computational intensive segments, and one of a lesser complexity: the reduction of the system matrix A to a Hessenberg form, the solution of n shifted Hessenberg systems and a number of matrix-matrix multiplies. The Hessenberg reduction is solved eeciently by using appropriate routines from the ScaLAPACK package while the system solves are designed using a pipelined scheme on distributed memory computers and improved data reuse. We provide timings of our implementation on the Intel Paragon (TM) Supercomputer which demonstrates both speed and eeciency of the proposed algorithm.
2018 Ninth International Green and Sustainable Computing Conference (IGSC), 2018
Constraints on power consumption are having per-vasive effects on high-performance computing (HPC... more Constraints on power consumption are having per-vasive effects on high-performance computing (HPC) systems, the facilities in which they are housed, and the application codes themselves. These constraints are driven by a variety of reasons including physical limits on available power within a facility, or are due to utility demand response etc. It is essential that power management must now be added to the traditional HPC goals of monitoring and optimizing application and algorithm correctness, scalability, and performance. Irrespective of the causes for these constraints, a holistic approach must be taken to effectively manage power or schedule jobs more optimally so data centers can remain below these power constraints. In this paper we propose results of implementing several strategies for throttling power and its impact on application performance while providing insights for managing overall system energy levels.
We present an eecient design and implementation for the solution of the Sylvester Observer matrix... more We present an eecient design and implementation for the solution of the Sylvester Observer matrix equation on a distributed memory computer. It consists of two main computational intensive segments, and one of a lesser complexity: the reduction of the system matrix A to a Hessenberg form, the solution of n shifted Hessenberg systems and a number of matrix-matrix multiplies. The Hessenberg reduction is solved eeciently by using appropriate routines from the ScaLAPACK package while the system solves are designed using a pipelined scheme on distributed memory computers and improved data reuse. We provide timings of our implementation on the Intel Paragon (TM) Supercomputer which demonstrates both speed and eeciency of the proposed algorithm.
Uploads