適用於熱感知三維晶片網路之資源分配與可適性路由演算法

Abstract

本篇論文針對三維晶片網路之熱感知設計的效能降低做出改善。我們分別提出兩種設計技巧：1) 針對過熱關閉元件之三維晶片網路下的可適性路由演算法 2) 針對流量遷移的資源分配演算法。為了促進三維晶片網路之熱感知設計與相關研究，我們亦提出模擬平台：3) 針對三維晶片網路的流量-溫度共同模擬平台。對於過熱關閉造成的流量阻塞，我們提出一可適性路由演算法以改善網路效能。此可適性路由演算法基於奇偶路由法，能感知網路上的流量資訊以及過熱關閉資訊，進而自動選擇適當路徑，使封包能繞開過熱關閉元件以及平衡網路負載。由於能降低三維晶片網路的穩態溫度的流量遷移，帶來各層不平衡的網路負載並造成流量阻塞。我們針對流量遷移提出一資源分配演算法，藉由分配不同層上路由器上的緩衝器深度，平衡各層的網路負載，進而改善網路效能。因為現行模擬平台缺乏支援流量與溫度的共同模擬，我們自行開發一模擬平台，以支援上述兩種熱感知設計與相關效能改進技術。此模擬平台亦經由商用溫度模型工具驗證其垂直面的熱傳導模擬，以確保相對性上的正確度。

In this thesis, we proposed two techniques to improve the performance degradation of thermal-aware 3D NoC designs: 1) Adaptive routing for throttled 3D NoC. 2) Resource allocation for traffic migration. And we also proposed: 3) a traffic-thermal co-simulation platform for 3D NoC to facilitate the two kinds of thermal-aware design. To address the traffic congestion due to throttling of transient-temperature control, we proposed an adaptive routing algorithm. Based on the Odd-Even routing function, a traffic- and throttling-aware selection strategy is proposed to decide proper routing path to balance network traffic and detour throttled tiles. While the traffic congestion happens after the traffic migration of steady-temperature optimization, we proposed a design flow of resource allocation. Depend on the loading requirement at each layer, our allocation assigns suitable buffer depth for each layer under a constraint of total buffer cost. Our proposed simulation platform supports unidirectional coupling and mutual coupling. And we validate the proposed simulation platform with commercial thermal modeling tool, CFD-RC, to assure the relative correctness of vertical heat conduction.