The Power to Schedule a Parallel Program
The Power to Schedule a Parallel Program
复制标题
安排并行程序的能力
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Seth Gilbert
中科院分区：
文献类型：
--
作者：
Kunal Agrawal;Seth Gilbert
在本文中，我们考虑了在多层或多处理器机器上调度并行程序的问题，以最大程度地减少所消耗的能量。通过计算，我们考虑了两个版本的问题：在给定功率约束的情况下，将运行时间最小化，并在给定时间限制的情况下最小化。事先了解该程序的任何知识，但可以在程序执行时调整活动的处理器及其速度的灵活性。 ，即使只调整了有限的速度，我们也表明我们的解决方案具有最佳的静态千里眼解决方案 - 调度程序知道提前程序，但必须提前选择活动的处理器及其速度的数量，并且在程序执行时无法更改它。计算。
In this paper, we consider the problem of scheduling parallel programs on a multicore or multiprocessor machine so as to minimize the energy consumed. By adjusting the number of active processors and/or the speed of those processors, we can vary the amount of power used by the computation. We consider two versions of the problem: minimizing the running time given a power constraint, and minimizing the energy used given a time constraint. We consider the problem in the non-clairvoyant setting where the scheduler does not know anything about the program in advance, but has the flexibility to adjust the number of active processors and their speed as the program executes. We present a work-stealing algorithm that relies on a backoff-backon strategy to solve both of these problems, even while only adjusting the number of active processors and their speed a limited number of times. We show that our solution is competitive with the best static clairvoyant solution – here the scheduler knows the structure of the program in advance, but must choose the number of active processors and their speed in advance and can not change it while the program executes. Thus we conclude that with only a small number of adjustments, we can compensate for the lack of information about the future of the computation.