透過您的圖書館登入
IP:18.118.144.69
  • 學位論文

在三維晶片內網路系統上適用於主動式路由演算法之熱感知動態緩衝配置

Thermal-aware Dynamic Buffer Allocation for Proactive Routing Algorithm on 3D Network-on-Chip Systems

指導教授 : 吳安宇

摘要


本篇論文針對三維晶片內網路交通與空間上溫度的不均所導致的效能降低問題,提出一種熱感知交通預先遷移演算法,並藉此減輕被節流的路由器降低的系統效能與提升在特定溫度限制下晶片內網路之吞吐量。傳統上,我們都認為溫度過高的區域是由於交通壅塞所引起的,所以為了均衡網路上的溫度,都會藉由平衡網路上的交通來達到空間上的溫度均衡。然而,溫度跟路由器切換封包的次數呈正相關,也就是說過熱的區域是由於穩定且大量的封包傳輸所造成的。另一方面,由於溫度取樣頻率有其限制,無法隨時得到當下的溫度資訊,過度依賴溫度去引導封包走向低溫區域來平衡空間的溫度分布反而會使得過多封包堵塞在低溫區域,如此一來便會使得吞吐量下降,降低系統效能。 本篇論文結合時間上與空間上的溫度資訊去限制路由資源,同時再從被限制的路由資源得到交通資訊去引導封包走向交通順暢且溫度較低的方向,藉此達到空間上溫度分布平均且提升系統效能。利用過去提出的溫度預測模型得到未來下一次溫度取樣可能的溫度,再將之與當下得到的溫度相減得到溫度上升速率,此即是時間上溫度變化的趨勢。再利用相鄰節點之溫度上升速率的大小關係來調整路由器各個方向最大可用的路由資源也就是暫存器深度。藉由調整暫存器深度與所提供的交通資訊來繞過潛在過熱區,進而減低過熱區域的交通量以及減慢溫度上升速度。由所提出的演算法可以使得系統的溫度分布標準差相較之前各種方法下降5.9%~33.2%,以及使得在熱限制之下的吞吐量進步4.6%~39.03%。

並列摘要


The thesis proposes a kind of proactive traffic migration algorithm to resolve the problems of performance degradation due to traffic and thermal imbalance in 3D Network-on-Chip (NoC). The proposed algorithm can mitigate the performance degradation due to the throttled nodes and improve thermal-limit throughput of the on-chip network. Conventionally, we believe that the overheating regions result from traffic congestion. To eliminate thermal hotspot, we balance thermal distribution spatially by balancing traffic on the network. However, thermal is highly correlated with the amount of packets switched by the router. That is, the overheating regions actually result from steady and a lot of packets transmitting. On the other hand, due to the limitation of sampling frequency, the system cannot get the thermal information at any time. Therefore, to balance thermal distribution, highly depending on temperature information easily causes traffic congestion in the low-temperature regions. The thesis integrates temporal and spatial temperature information to constrain the routing resource, and use proposed routing algorithm depending on the traffic information from the constrained routing resource to let packets go through non-congested and non-overheating regions. Through thermal prediction model, we can get temperature rising rate from the subtraction between the future possible temperature and the current temperature. Also, we use the difference of temperature rising rate between two adjacent nodes to control the routing resource (e.g., buffer depth). The information from the routing resource can help packets detouring the overheating regions. The proposed algorithm reduces temperature deviation by 5.9%~33.2% from worst case to best case and improve network throughput under a certain limit by around 4.6%~39.03%.

參考文獻


[4] V.F. Pavlidis, E.G. Friedman, “3D topologies for Networks-on-Chip,” IEEE Trans. Very Large Scale Integration System, vol. 15, no. 10 pp. 1081-1090, 2007.
[6] R.S. Patti, “Three-dimensional integrated circuits and the future of system-on-chip designs,” Proc. IEEE, pp. 1214–1224, 2006
[7] A. W. Topol, D.C. La Tulipe, L. Shi, et al., “Three-dimensional integrated circuits,” IBM J. Res. Develop., vol. 50, no. 4/5, pp. 491-506, 2006.
[8] B. Black, M. Annavaram, N. Brekelbaum, et al., “Die stacking (3D) microarchitecture,” Proc. International Symposium on Microarchitecture, pp. 469-479, Dec. 2006.
[9] B.S. Feero, P.P. Pande, “Networks-on-chip in a three-dimensional environment: a performance evaluation,” IEEE Trans. Computers, pp. 32–45, 2009

延伸閱讀