異質多核心系統可結合各種不同功能的處理器在同一晶片上，使得此晶片上的負
載可於與之最適的處理器上被執行。而在這些處理器中，圖形處理單元(Graphic
Processing Unit, 後簡稱GPU) 為其中之一個使用頻率最高的處理器，不僅是因為其科技發展度足夠成熟，並且具備高運算效能。晶片網路(Network-on-Chip, NoC)常作為連結各個中央處理單元(Central Processing Unit, 後簡稱CPU)、快取記憶體(Caches)、及記憶體控制器(Memory Controller)，因此晶片網路的設計會影響到整個系統的效能。一般而言CPU 和GPU 扮演不同的角色，負責執行不同類型的工作負載，產生至晶片網路的流量也是有不一樣的特徵。如GPU 為產生一次產生大量的流量，易受到整個網路的吞吐量影響。所以對於GPU 於晶片網路中的放置位置和配置給GPU 的網路資源皆需考量不影響GPU 的效能。本文主要對於晶片網路在不同處理器的擺置與不同的網路資源配置上作效能分析。使用gem5-gpu 模擬器模擬執行在異質CPU-GPU 多核心系統的程式並記錄對於記憶體存取的狀況。之後修改Garnet2.0 並使用記錄好的記憶體存取資訊，使之模擬CPU 和GPU 核心在晶片網路上不同的放置，對其作效能分析。

Heterogeneous multicore systems integrate different types of processors on the same
chip in order to match the workloads with the most appropriate processors. Among
the different types of processors, Graphic Processing Unit (GPU) is one of the most
commonly used processors, not only because of their technology maturity but also
because of their high computing-power ratio. To interconnect the multiple CPUs,
GPUs, caches, and memory controllers on a chip, a network-on-chip (NoC) is often
used, whose design is very critical to the performance of the whole system. As
CPUs and GPUs often play different roles and execute different workloads, the traffic
injected into the NoC from CPUs and GPUs may have very different characteristics.
GPU cores tend to generate bursty high-volume traffic, which is throughputsensitive.
Therefore, the placement of GPU cores in the NoC and the network resources
allocated to the GPUs should be designed carefully in order not to hinder
the performance of GPUs. In this thesis, we evaluate the performance of NoC under
different processor placement and network resource allocation. We use gem5-gpu
simulator to simulate applications running in a heterogeneous CPU-GPU multicore
system and record the memory access trace. We then modify Garnet2.0 to take the
trace as input and evaluate the performance of different system configurations. It is
observed that GPUs are sensitive to the proximity to memory from the evaluations.
Therefore it can be referred as a critical impacting factor in terms of heterogeneous NoC design.

摘要ii
Abstract iii
1 Introduction 1
2 Methodology 3
2.1 The default gem5-gpu architecture . . . . . . . . . . . . . . . . . . . . . 3
2.2 The techniques of simulating Mesh topology for CPU-GPU heterogeneous
system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.1 NoC simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.2 Lifetime of memory access in gem5-gpu simulator . . . . . . . . 5
2.2.3 Ensnare and process the memory access event from Ruby memory
system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Implementation of simulating heterogeneous mesh NoC . . . . . . . . 6
2.3.1 Implementation issues . . . . . . . . . . . . . . . . . . . . . . . . 7
Issue 1: Virtual network mapping . . . . . . . . . . . . . . . . . 7
Issue 2: Response message delivery . . . . . . . . . . . . . . . . 8
Issue 3: Message delivery between cache controller and Garnet2.0
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Evaluation of the heterogeneous NoC design . . . . . . . . . . . . . . . 8
3 Evaluation 9
3.1 Evaluation Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Application Workloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Evaluation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4 Conclusions 19
vi
Bibliography 20

[1] Jieming Yin et al. “Energy-Efficient Time-Division Multiplexed Hybrid-Switched
NoC for Heterogeneous Multicore Systems”. In: Proceedings of the 2014 IEEE
28th International Parallel and Distributed Processing Symposium. IPDPS ’14.Washington,
DC, USA: IEEE Computer Society, 2014, pp. 293–303. ISBN: 978-1-4799-
3800-1. DOI: 10.1109/IPDPS.2014.40. URL: http://dx.doi.org/10.
1109/IPDPS.2014.40.
[2] Ali Bakhoda, John Kim, and Tor M. Aamodt. “Throughput-Effective On-Chip
Networks for Manycore Accelerators”. In: Proceedings of the 2010 43rd Annual
IEEE/ACM International Symposium on Microarchitecture. MICRO ’43. Washington,
DC, USA: IEEE Computer Society, 2010, pp. 421–432. ISBN: 978-0-7695-
4299-7. DOI: 10.1109/MICRO.2010.50. URL: http://dx.doi.org/
10.1109/MICRO.2010.50.
[3] Yaniv Ben-Itzhak, Israel Cidon, and Avinoam Kolodny. “Optimizing Heterogeneous
NoC Design”. In: Proceedings of the International Workshop on System Level
Interconnect Prediction. SLIP ’12. San Francisco, California: ACM, 2012, pp. 32–
39. ISBN: 978-1-4503-1437-4. DOI: 10.1145/2347655.2347670. URL: http:
//doi.acm.org/10.1145/2347655.2347670.
[4] Jason Power et al. “gem5-gpu:AHeterogeneous CPU-GPU Simulator”. In: Computer
Architecture Letters 13.1 (2014). ISSN: 1556-6056. DOI: 10 . 1109 / LCA .
2014.2299539. URL: http://gem5-gpu.cs.wisc.edu.
[5] Niket Agarwal et al. “GARNET: A detailed on-chip network model inside a
full-system simulator”. In: Performance Analysis of Systems and Software, 2009.
ISPASS 2009. IEEE International Symposium on. IEEE. 2009, pp. 33–42.
[6] Shuai Che et al. “Rodinia: A Benchmark Suite for Heterogeneous Computing”.
In: Proceedings of the 2009 IEEE International Symposium onWorkload Characterization
(IISWC). IISWC ’09. Washington, DC, USA: IEEE Computer Society, 2009,
pp. 44–54. ISBN: 978-1-4244-5156-2. DOI: 10.1109/IISWC.2009.5306797.
URL: https://doi.org/10.1109/IISWC.2009.5306797