高效能記憶體互連網路之設計

巨量資料處理的需求日益增加，處理資料所需的記憶體頻寬卻受限於記憶體處理能力的不足，高效能電腦系統的計算能力亦不能充分發揮。為了提升記憶體資料傳送頻寬，許多新型記憶體架構，例如Hybrid Memory Cube，即針對大量資料傳輸，透過更多的記憶體通道與序列化記憶體封包傳輸，提供更大量的資料吞吐能力。然而，大量的資料處理應用，需要更多的記憶體空間，電腦系統內亦須整合更多的Hybrid Memory Cube記憶體模組。用以連接處理器核心與多個Hybrid Memory Cube記憶體模組的記憶體互連網路，即成為一個影響效能表現的重要因素。有鑑於此，本研究提出了一個全新的記憶體互連網路－MemGrid，其本身具有高效能表現與可平面化的特性。MemGrid記憶體網路具有良好的擴充能力，相較於其他互連網路，在Hybrid Memory Cube記憶體模組數量增加時，其傳輸效能相當優異。相較於其他不同種類的記憶體互連網路，MemGrid可以獲得良好的效能表現；在連結成本的分析中，相較於其他互連網路，其所需之連結成本亦較低。

關鍵字

記憶體互連網路, 多核心架構, 高效能記憶體架構

並列摘要

With the growing requirement of data processing capability in big data science, the provided memory bandwidth is insufficient. The overall performance of the modern high performance computer systems are limited accordingly. Hence a new memory architecture, Hybrid Memory Cube (HMC), is proposed to improve the memory accessing performance while processing large amount of data in the big data application. By adopting multiple memory channels and serialized transmission of memory packets, the HMC memory module can provide more memory accessing bandwidth. However, the more computed data requires larger memory space. The number of HMC memory modules are increased accordingly. The computer system has to integrate more HMC memory modules and the processors via the memory network. The capability of memory interconnection network become an important factor of the overall computer performance. In this study, we develop a novel memory interconnection network, called MemGrid, to overcome the above challenge. The MemGrid network has the capabilities of high performance, high scalability, and high planarity. As the amount of HMC memory module increases, the MemGrid network can easy to scale up. Compared to other interconnection networks, MemGrid network can provide better performance than others. The required link number of MemGrid network is less than others.

並列關鍵字

Memory Interconnection Network, Multicore Architecture, High Performance Memory Architecture

參考文獻

[3] J. T. Pawlowski, “Hybrid Memory Cube (HMC)”, IEEE Hot Chips 23 Symposium, 2011.

[4] J. L. Hennessy and D. A. Patterson, “Computer Architecture – A Quantitative Approach”. 4th Edition, Morgan Kauffman, San Francisco, 2006.

[5] S. C. Woo, J. P. Singh and J. L. Hennessy, “The Performance Advantages of Integrating Block Data Transfer in Cache-Coherent Multiprocessors”, ACM SIGOPS Operating Systems Review, vol. 28, no. 5, pp. 219-229, 1994

[6] E. Rothberg and A. Gupta, “An Efficient Block-Oriented Approach to Parallel Sparse Cholesky Factorization”, SIAM Journal on Scientific Computing, pp. 1413-1439, 1994.

[8] J. P. Singh, J. L. Hennessy and A. Gupta, “Implications of Hierarchical N-Body Methods for Multiprocessor Architectures”, ACM Transactions on Computer Systems, vol. 13, no. 2, pp. 141-202, 1995.

國際替代計量

高效能記憶體互連網路之設計

未授權

主題瀏覽