Optimizing time and energy in MPI programs in a power-scalable cluster
Springer, Robert Coleman
MetadataShow full item record
Recently, the idea that power is a performance-limiting factor has gained traction in the high-performance computing community. This may be due to the fact that the cost of energy has become increasingly signi cant, or that the heat produced by higher-energy components tends to reduce their reliability. In addition, cooling and providing ample power to supercomputers is becoming more di cult as their power requirements increase. One way to reduce power (and therefore energy) requirements is to use high-performance cluster nodes that are frequency- and voltage-scalable (e.g., AMD-64 processors). The problem we address in this thesis is: given a power-scalable cluster and an upper limit for energy consumption, choose a schedule (number of nodes and CPU speed per node) that simultaneously (1) satis es an external upper limit for energy consumption and (2) minimizes execution time. We do this using a novel combination of modeling, execution and pro ling. Using our technique, we are able to nd a near-optimal schedule in just a handful of partial program executions.