Data-Intensive Workload Consolidation for the Hadoop Distributed File System
GRID '12 Proceedings of the 2012 ACM/IEEE 13th International Conference on Grid Computing
Using synchronization stalls in power-aware accelerators
Proceedings of the Conference on Design, Automation and Test in Europe
High-Resolution power profiling of GPU functions using low-resolution measurement
Euro-Par'13 Proceedings of the 19th international conference on Parallel Processing
Hi-index | 0.00 |
Enterprise workloads like search, data mining and analytics, etc. typically involve a large number of users who are simultaneously using applications that are hosted on clusters of commodity computers. Use of GPUs for enterprise computing is challenging because of poor performance and higher energy consumption compared to running enterprise workloads on CPUs. In this paper, we show that the GPU work consolidation can improve system throughput and results in significant energy savings over multicore CPUs. We develop a novel runtime framework that dynamically consolidates instances from different workloads from multiple user processes into a single GPU workload. However, arbitrary consolidation of GPU workloads does not always lead to better energy efficiency. We use new GPU performance and power models to make predictions for potential workload consolidation alternatives and identify useful consolidations. Our experiments on a variety of workloads (that perform poorly on a GPU compared to well optimized multicore CPU implementations) show that the proposed framework for GPUcan provide 2X to 22X energy benefit over a multicore CPU.