Accelerating text mining workloads in a MapReduce-based distributed GPU environment
Journal of Parallel and Distributed Computing
| Hi-index | 0.00 |
This talk will present the Mont-Blanc project, an European initiative to build exascale systems using energy-efficient parts coming from the embedded market. The energy consumption of current general purpose and high-performance chips would require an unaffordable total power budget for an exascale system to be build using these parts. The Mont-Blanc project aims to lower the total power of exascale systems by using parts from the embedded market which have a much higher FLOPS/Watt ration than traditional general purpose processor, at the cost of a lower peak performance per chip. Hence, exascale systems built using embedded parts would require a very high number of processors. In this context, overlapping communications and computations is key for applications to reach the system peak performance. This would require highly tuned application code which most users would not be able to afford. The Mont-Blanc project heavily relies on the OmpSs programming model. OmpSs provide a simple parallel programming interface that most users can easily use, and an advanced runtime system that automatically overlaps computation and communication. Furthermore, the OmpSs runtime system is also able to dynamically adapt the load of each node to accomplish the overall system load balance.