基於網宇實體系統整合的資料重用與工作替換

Abstract

一個多層次的系統需要整合伺服器和 IoT 設備，將 AI 應用整合到多層次系統中在現代是很有利的，而 AIoT 代表了在 IoT 設備中整合 GPU 進行 AI 計算，然而這會面臨幾個問題。首先，由於 AIoT 設備的規模限制，其設備也有計算限制。大多數 AI 算法都有巨大的重複性浮點數計算，以獲得基於機率的收斂近似解。GPU 使 AI 計算適合對大量的浮點數計算進行平行計算。其次，AIoT 設備還必須考慮工作的延遲要求，使 AIoT 設備控制的系統執行器在工作的關鍵事件中不會延遲，例如自動駕駛車輛中的剎車。第三，可以通過進一步整合邊緣伺服器、AIoT設備和現有的基礎設施來構建多層次系統，使數據重用具可行性且高效，多層次的 AIoT 系統在異質資源中會有更多的用途。本論文的貢獻是通過 CPS 概念將 IoT 平台計算和 AIoT 應用結合起來，然後探索 AIoT 設備在整合架構中的計算限制和解決方案，以實證方式來研究 AIoT 設備運行 AI 應用時適當的資源配置方法，在多層次系統中，不同的計算資源會導致不同的響應時間以滿足延遲要求。本論文的概念是數據重用，它來自於現有的基礎設施，並使用於新增加的 AIoT 設備，同時當代的 AIoT 設備配備的 GPU 對於AI 算法的局限性也是通過實證評估來衡量的。為了克服這些限制，本論文的架構提出工作替換和邊緣伺服器工作卸載的可行性。

A multi-tier system needs integration of servers and IoT devices. Modern, it is advantageous to integrate AI applications into a multi-tier system, which in AIoT terms stands for integrating GPUs in IoT devices for AI computing. However, there are several issues that need to be handled. First, AIoT devices also have computing limitations because of the scale of devices. Most AI algorithms have enormous repetitive floating-point computations for convergent approximate solutions based on probability. Using GPU, AI computing is suitable for parallel computing on enormous floating-point computations. Second, AIoT devices must also consider the latency requirements of the tasks, and system actuators controlled by AIoT devices will not get delayed in critical events. For example, the brake in autonomous vehicles. Third, multi-tier systems can be constructed by further integrating edge servers, AIoT devices, and existing infrastructure that makes data reuse feasible and efficient. Multi-tier AIoT systems will be more versatile in heterogeneous resources.The contribution of this thesis is to combine IoT platform computing and AIoT applications by CPS concepts, and then explore AIoT devices’ computing limitation and resolution in integrating architectures. The empirical results were investigated for the appropriate resource usage configuration for AIoT devices to run AI applications. In the multi-tier system, different computing resources result in different response time to meet latency requirements. The concept of this thesis is data reuse, which comes from the existing infrastructure, for the newly added AIoT devices. The limitations of contemporary AIoT devices equipped with GPU for AI algorithms were measured by empirical evaluation. For overcoming these limitations, this thesis concludes with the feasibility of using task replacement and edge server offloading.