考量最低能源消耗之雲端資料中心動態資源分配演算法

Abstract

隨著越來越多的公司將資訊服務轉移到公有雲資料中心，如何在滿足各個應用程式不斷變動的資源需求下，在雲端資料中心達成省電節能的資源分配，已經成為一個相當吸引人的問題。許多動態資源分配法被提出來解決這些問題，但他們大部分都缺少像同時調整VM及server分配量、橫跨時間軸上的最佳化、及最佳解演算法這些嚴重影響能源效率的因素。在這篇論文中，我們提出了一個新的應用程式層級的動態資源分配演算法稱作Time-Directed Dijkstra (TD-D)，它可以在現有的負載預測機制幫助下，藉由達成運轉耗電(operating cost)及開關耗電(switching cost)間的平衡來達成最低的能源消耗。我們主要的貢獻如下：(1)我們分類並整理了影響資源分配演算法能源效率的主要因素。(2)我們提出了一個最低能源消耗的資源分配演算法，它可以使用市面上常見的機器，在合理的時間內算出最佳解。(3)我們也驗證了這個新演算法即使在負載預測發生錯誤時，依然可以保持可靠(低資源分配不足率，low resource under-allocation rate)和有效(好的能源效率)。我們的模擬結果顯示，這個新的最佳解演算法可以比其他現有的代表性近似解演算法(local search)耗用更少的運算時間並省下平均9.5%的能源消耗。除此之外，我們的演算法在使用有預測錯誤的負載預測資料的情況下(7個單位時間後產生25%錯誤率)，仍然可以比現有的近似解演算法省下平均7.1%的能源消耗，並保持很低的資源分配不足(under-allocation)率(在1個單位時間內，每個應用程式平均0.03台VM)。

As more and more companies outsource their information services to public cloud data centers, how to perform dynamic resource allocation efficiently to reduce energy consumption while fulfilling each application’s fluctuating resource demands has become a very challenging task. Many dynamic resource allocation approaches were proposed to tackle this task, but most of them lack for some influencing factors that may impact the energy efficiency, such as resizing at both VM and server levels, optimization over time horizon, and the optimality of the algorithm. With the help of existing load prediction techniques, in this paper, we design an application-level dynamic resource allocation algorithm, called Time-Directed Dijkstra (TD-D), which can achieve minimum energy consumption, by seeking the best trade-off between operating cost and switching cost due to switching on/off resources. The main contributions of this paper are as follows. (1) We analyze and categorize the most influencing factors that should be addressed in order to build an application-level energy efficient resource allocation algorithm. (2) We develop a minimum energy consumption algorithm that can produce an optimal solution within reasonable computing time using commodity machines. (3) We demonstrate the proposed TD-D algorithm is fairly robust in terms of low resource under-allocation rate and energy-efficient to prediction errors. Simulation results show that, our optimal TD-D algorithm can save 9.5% of energy consumption in average using error-free workload data compared with a representative best-effort algorithm (local search), and consume much less computing time compared with the representative algorithm. In addition, using workload data with 25% prediction error after a prediction window of 7 time slots, our TD-D algorithm can save 7.1% of energy consumption than the representative algorithm and keep a very low resource under-allocation rate (0.03 VM / (time slot × application)).