雲端計算中工作序列的虛擬機器排程方法

Abstract

這篇論文指出資料中心在能源保存上的重要議題。我們探討如何將伺服器分配給工作序列並且減少總能源消耗，在此我們使用效能評比標準為“浪費能源” – 伺服器額外供應給工作的計算資源超出實際工作需求的部分。我們提出了三種工作序列資源佈署策略，分別為最大機器優先方法、最佳符合方法和混合方法。我們並證明出最大機器優先方法與混合方法擁有效能保證，最多只會造成 2/n 浪費能源比值。換句話說，額外浪費能源與實際供應能源比值不會超過 2(1 + δ)/n，n 為工作序列中工作數目，而 1 + δ 為工作序列中工作最大執行時間與最小執行時間的比值。當δ 任意大時，我們也對浪費能源推導出一個精確界限 1/2 。最後我們藉由實驗去比較這三種方法在實際應用上的效能，實驗結果顯示出這三種方法都浪費相當少能源。混合方法優於最佳符合方法，且最佳符合方法優於最大機器優先方法。

This paper describes the important issue of energy conservation for data centers. We consider the problem of provisioning physical servers to a sequence of jobs, and reducing the total energy consumption. The performance metric is the wasted energy – the over-provisioned computing power provided by the physical servers, but exceeding the requirement of the jobs. We propose three new strategies for allocating servers to a sequence of jobs – a largest machine first heuristic, a best fit method, and a mixed method. We prove that both the largest machine first heuristic and the mixed method will only incur at most 2/n in over-provisioned energy. That is, the ratio between the over-provisioned energy and the total provisioned energy is bounded by 2(1 + δ)/n, where n is the number of jobs, and 1+δ is the ratio between the maximum and minimum execution time of jobs. We also derive a tight bound of 1/2 on the ratio of wasted energy if the ratio δ could be arbitrarily large. We also conduct experiments to compare the three algorithms in practice. The experiment results indicate that all three algorithms waste very little energy in over-provision. The mixed method outperforms the best fit method, which outperforms the largest machine first method.