Search for: All records

This paper studies a cloud datacenter (DC) consisting of two types of tasks with different priority levels. While non-priority tasks generally request the use of a single virtual machine (VM), priority tasks may utilize multiple available VMs to accelerate processing. We focus on determining whether to accept or reject non-priority tasks to maximize overall system benefits. By formulating the problem as a stochastic dynamic program, it is verified that the best approach for handling nonpriority tasks adheres to a control-limit framework. Both experimental outcomes and numerical evaluations highlight the efficacy of the proposed method, leading to the identification of the optimal threshold. The key contribution of this paper is the development of a stochastic dynamic program for DC resource management and the explicit derivation of an optimal control-limit policy. Both value iteration and linear programming methods are utilized to solve optimization problems. These results offer essential understanding for assessing the performance of various DC models, optimizing both rewards and resources efficiently. 

Cloud computing (CC), often necessitates dynamic adjustments due to its inherently fluid nature. In this paper, we introduce a novel dynamic task scheduling model that incorporates reward and holding cost considerations, leveraging the Continuous-Time Markov Decision Process (CTMDP) framework in heterogeneous CC systems. The primary goal of this model is to maximize the overall system reward for the Cloud Service Provider. By solving the Bellman Optimality Equation using the value-iteration method, we can derive an optimal scheduling policy for the dynamic task scheduling model. Additionally, to enhance its practicality in real-world scenarios, we incorporate a model-free reinforcement learning algorithm to obtain the optimal policy for our proposed model without requiring explicit knowledge of the system environment. Simulation results demonstrate that our proposed model outperforms two common static scheduling methods. 

Abstract Using an eastern tropical Pacific pacemaker experiment called the Pacific Ocean–Global Atmosphere (POGA) run, this study investigated the internal variability in sea surface salinity (SSS) and its impacts on the assessment of long-term trends. By constraining the eastern tropical Pacific sea surface temperature variability with observations, the POGA experiment successfully simulated the observed variability of SSS. The long-term trend in POGA SSS shows a general pattern of salty regions becoming saltier (e.g., the northern Atlantic) and fresh regions becoming fresher, which agrees with previous studies. The 1950–2012 long-term trend in SSS is modulated by the internal variability associated with the interdecadal Pacific oscillation (IPO). Due to this variability, there are some regional discrepancies in the SSS 1950–2012 long-term change between POGA and the free-running simulation forced with historical radiative forcing, especially for the western tropical Pacific and southeastern Indian Ocean. Our analysis shows that the tropical Pacific cooling and intensified Walker circulation caused the SSS to increase in the western tropical Pacific and decrease in the southeastern Indian Ocean during the 20-yr period of 1993–2012. This decadal variability has led to large uncertainties in the estimation of radiative-forced trends on a regional scale. For the 63-yr period of 1950–2012, the IPO caused an offset of ~40% in the radiative-forced SSS trend in the western tropical Pacific and ~170% enhancement in the trend in the southeastern Indian Ocean. Understanding and quantifying the contribution of internal variability to SSS trends helps improve the skill for estimates and prediction of salinity/water cycle changes.