在有限決策時間下的雲端機器學習計算工 作排程總完工時間最佳化策略研究;Optimization Strategy for Makespan of ML-Task Scheduling on the Cloud with the Constraint of Scheduling Decision Time

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/93535

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/93535

Title:	在有限決策時間下的雲端機器學習計算工作排程總完工時間最佳化策略研究;Optimization Strategy for Makespan of ML-Task Scheduling on the Cloud with the Constraint of Scheduling Decision Time
Authors:	張峻瑋;Chang, Chun-Wei
Contributors:	資訊工程學系
Keywords:	雲端;機器學習工作;總完工時間最佳化;快速排程策略;Cloud Computing;ML-Task Scheduling;Scheduling Decision Time;Makespan
Date:	2023-10-20
Issue Date:	2024-03-05 17:43:30 (UTC+8)
Publisher:	國立中央大學
Abstract:	由於機器學習的技術不斷的進步，任務對於雲端群集的需求逐漸增加，這些資源密集型的任務需要更多的計算資源來支持，更多的計算資源和任務就表示對於資源分配要有更好的處理。對於阿里巴巴所提出的模擬器，它是利用貪婪演算法進行任務排程，因為貪婪演算法存在著一些缺點，因此本論文提出以Min-Min演算法和Max-Min演算法以動態任務分配的方式來改善貪婪演算法的缺點。另外，所提出之排程演算法因具有高時間複雜度，這對於大規模分散式系統的任務排程造成了挑戰，所以提出了將任務已更小的單位切割任務資料集後再進行排程。實驗結果表明，Min-Min演算和Max-Min演算法之於阿里巴巴所提出的模擬器結果在任務的 Makespan 有良好的表現，而在切割任務後的結果來看也有良好的表現。;Due to the continuous advancements in machine learning technology, there is a growing demand for tasks in cloud clusters. These resource-intensive tasks require more computational resources to support them. More tasks and computational resources imply a need for improved resource allocation. In the case of the simulator proposed by Alibaba, it utilizes a greedy algorithm for task scheduling. However, since greedy algorithms have their limitations, this paper suggests enhancing the drawbacks of the greedy algorithm by employing the Min-Min and Max-Min algorithms with dynamic task allocation. Furthermore, the scheduling algorithms introduced here present a challenge due to their high time complexity, especially in the context of task scheduling in large-scale distributed systems. Therefore, it is proposed to break down the task dataset into smaller units for scheduling. Experimental results indicate that both the Min Min and Max-Min algorithms perform well in terms of the Makespan of tasks in Alibaba′s simulator. Moreover, when considering the results after task partitioning, these algorithms still demonstrate excellent performance.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	76	View/Open

社群 sharing

Loading...