協同行為導向之系統模擬平台應用於平行運作系統設計、除錯

A Virtualization-Assisted Full-System Simulation Approach for the Verification of System Inter-Component Interactions

吳昕益(Hsin-I Wu)

指導教授 ： 蔡仁松

國立清華大學/電機資訊學院/資訊工程學系所/博士(2020年)

摘要

全系統模擬對於Embedded System設計驗證至關重要。其Flexibility與Early Availability性質特別適合設計初期探索與驗證系統行為。然而,傳統的模擬加速方法通常會遭遇Performance、Accuracy或是Scalability之困難。因此,本計畫提出VIRA (VIRtualization-Assisted)方法,建立Fast、Accurate與Scalable之全系統模擬器以克服上述困難。為了增加模擬效能,VIRA將Hardware-Assisted Component執行於Host HardwareDevice以得到Native Execution之效能。為了確保Accuracy與正確的Data Dependency,本計畫提出了一個包含Bus Contention Delay之Deterministic Timing Model。為了增加擴充功能,VIRA整合了Software-Modeled Component以支援新增功能,並使用快速Data Pass-Through機制以減少被模擬元件間的Communication Overhead。我們透過在市售的SoC(System-on-Chip)板上實作此一技術來驗證所提出的Virtualization-Assisted全系統模擬。實驗結果顯示,除了能夠得到正確的Inter-Component Interaction結果,執行速度也比Commercial Functional Simulator快了58〜625倍。

關鍵字

模型建構 ； 效能分析 ； 模擬 ； 虛擬化

並列摘要

We propose in this thesis a near-real-time performance full-system simulation approach with hardware acceleration using virtualization techniques. Traditional acceleration approaches generally cannot capture inter-component interactions due to unpredictable component simulation progress. Our approach leverages existing hardware virtualization framework and devises three key implementation techniques to achieve fast and accurate full-system simulations. First, our approach utilizes the virtualization framework trap mechanism and precisely intercepts inter-component interactions with no need to check every data access, but effectively maintains deterministic chronological orders of inter-component interactions. Second, VIRA provides very accurate system performance estimation for early system-level designs through effective integration of component timing models, interrupt effects, and bus contention analysis. Third, VIRA achieves near-real-time performance by having software and hardware simulated components executed on the same host machine to minimize the overhead of inter-component data exchange. We implement the proposed approach on a virtualization-enabled off-the-shelf System-on-Chip board to demonstrate the effectiveness of our idea. The experiments show that VIRA always produces deterministic results while running 58~625 times faster than a commercial tool and the system performance estimation is only 3~6% from real systems. Moreover, our deterministic full-system simulator is also verified to carry as little as 2~57% overhead compared to ideal native executions on the same host hardware devices.

並列關鍵字

Modeling ； Performance analysis ； Simulation ； Virtualization

參考文獻

[1]Wu, M. H., Wang, P. C., Fu, C. Y., and Tsay, R. S. “A Distributed Timing Synchronization Technique for Parallel Multi-Core Instruction-Set Simulation.” In ACM Transactions on Embedded Computing Systems. no. 54. 2013.

[2]Cai, L., & Gajski, D. “Transaction level modeling: an overview.” In Proceedings of the 1st IEEE/ACM/IFIP international conference on Hardware/software codesign and system synthesis. pp. 19-24. 2003

[3]Bellard, F. “QEMU, a Fast and Portable Dynamic Translator.” In USENIX Annual Technical Conference. pp. 41-46. 2005.

[4]Khaligh, R. S., & Radetzki, M. “A dynamic load balancing method for parallel simulation of accuracy adaptive TLMs.” In Specification & Design Languages. pp. 1-6. 2010.

[5]Chen, J., Annavaram, M., & Dubois, M. “SlackSim: a platform for parallel simulations of CMPs on CMPs.” In ACM SIGARCH Computer Architecture News. pp. 20-29. 2009.