Cache-conscious scheduling of streaming applications
Proceedings of the twenty-fourth annual ACM symposium on Parallelism in algorithms and architectures
| Hi-index | 0.00 |
Streaming applications have become increasingly important and widespread, with application domains ranging from embedded devices to server systems. Traditionally, researchers have been focusing on improving the performance of streaming applications to achieve high throughput and low response time. However, increasingly more attention is being shifted to power/performance trade-off because power consumption has become a limiting factor on system design as integrated circuits enter the realm of nanometer technology. This work addresses the problem of scheduling a streaming application (represented by a task graph) with the goal of minimizing its energy consumption while satisfying its two quality of service (QoS) requirements, namely, throughput and response time. The available power management mechanisms are dynamic voltage scaling (DVS), which has been shown to be effective in reducing dynamic power consumption, and vary-on/vary-off, which turns processors on and off to save static power consumption. Scheduling algorithms are proposed for different computing platforms (uniprocessor and multiprocessor systems), different characteristics of workload (deterministic and stochastic workload), and different types of task graphs (singleton and general task graphs). Both continuous and discrete processor power models are considered. The highlights are a unified approach for obtaining optimal (or provably close to optimal) uniprocessor DVS schemes for various DVS strategies and a novel multiprocessor scheduling algorithm that exploits the difference between the two QoS requirements to perform processor allocation, task mapping, and task speed scheduling simultaneously.