針對巨量資料分析的異質計算提升加速器的使用率與吞吐量

Abstract

異質計算提供了Big Data和人工智慧在效能、成本和功耗上很大的改善空間。各種加速器被設計出與通用的中央處理器（CPU）協同處理大量的資料。然而各硬體架構的限制與軟體的設計使得效能瓶頸依然存在。本論文分析加速器的低使用率議題與提出解決方案，以求更好地利用加速器於大數據分析。首先我們探討即時系統中同步協定對於加速器使用率的影響並提出改善方案來提高其使用率，同時我們保證即時系統的性質。第二部份我們探討大數據分析中，演算法無法妥善利用圖形處理器(GPU)的問題。我們以經典的頻繁樣式探勘演算法-FP-growth當作研究案例，提出了適合GPU的資料結構以及演算法，藉此消除大量記憶體配置的開銷。最後，我們進一步探討多GPU系統中，在考量GPU的拓樸下如何有效地使用GPU。我們針對多工環境下的深度學習訓練，提出共享多GPU系統的排程機制以達到最小化平均工作完成時間的目的。本論文中的解決方案經由實驗與分析，均證實了對於聲稱的目標有顯著的效果。

Heterogeneous computing provides tremendous opportunities in performance, cost, and energy optimizations to Big Data and Artificial Intelligence applications. Various accelerators, such as GPU, or hardware architectures are designed to work together with general-purpose CPUs in large-scaled data processing. However, there are still inevitable processing bottlenecks between hardware components, due to architecture constraints and applications’ designs and behaviors. This dissertation is to address the utilization issues and their solutions to better utilize accelerators in large-scaled data processing. Frist, we exploit synchronization protocols for accelerators to improve the accelerator utilization as well as to guarantee the real-time requirements of the system. In the second part of the dissertation, we then explored the GPU-utilization problems in running algorithms behind Big Data processing. The classical FP-growth frequent pattern mining algorithm was taken as an example in the study, and a GPU-friendly algorithm was proposed by transforming recursive function calls into iterative ones and also by minimizing massive dynamic memory allocations. In the third part of the dissertation, we further explored both the GPU topology of servers and how effectiveness GPUs could be utilized by applications. A scheduling policy is presented for users in sharing GPU-powered servers for deep learning workloads, with an objective to minimize the average job completion time. The proposed solutions in this dissertation were all verified by experiments and/or analysis so as to show the effectiveness in resolving each respectively identified problem.