在異質系統架構下加速SQL資料庫應用

Accelerating SQL Database Applications with Heterogeneous System Architecture

林冠儒(Kuan-Ju Lin)

指導教授 ： 洪士灝

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2015年)

https://doi.org/10.6342/NTU.2015.10771

摘要

資料庫系統在資訊領域中扮演著非常重要的角色, 先前的研究利用分離式(discrete) 圖形運算處理單元 (GPU) 大量平行的計算能力,能夠有效的加速資料庫系統。然而使用上卻會受到分離式 GPU 需要資料搬移的時間、記憶體容量上的限制。
新興的異質系統架構 (HSA) 致力於加強不同運算設備間的合作關係,在這種新的架構下,透過共享記憶體的機制,工作的分配以及資料的交換皆能有效的完成。在我們研究中,我們利用異質系統架構下的 GPU 作為加速器來提昇資料庫系統的效能,這個資料庫不再受到傳統分離式 GPU 資料庫使用上的限制,此外,我們還實做一個工作分配機制以及負載平衡機制,這些機制會根據使用者的搜尋 (Query) 特
性,並且觀測處理器的負載,將工作交給最合適的設備去處理。實驗結果顯示,我們的機制能夠有效的提昇資料庫的效能,使資料庫加速了 1.77 倍。

關鍵字

資料庫系統 ； 異質系統架構 ； 負載平衡

並列摘要

Database systems serve important purposes in the information technology landscape. Prior works showed that massive computing resources provided by Graphical Processing Units (GPU’s) could accelerate the processing speed for database systems efficiently. However, those works suffered from the
limitations posed by discrete GPU’s, including the data transfer time and the size of GPU local memories.
The emerging heterogeneous system architecture (HSA) aims to enhance the collaboration of heterogeneous processors on executing parallel programs, where different computing devices in such a platform can dispatch tasks and
exchange data quickly with shared memory space. In this work, we propose to use HSA-based GPU’s as an accelerator for improving the performance of database systems. This proposed database system offers better cost-performance as it is no longer bounded by the limitations of traditional discrete GPU. To further enhance its performance, we develop a job dispatcher and a load balancer for the proposed database system to characterize the properties of queries and schedule data queries appropriately to different devices. The experimental results show that our proposed algorithms are capable of improving database performance by 1.77 times.

並列關鍵字

Database system ； Heterogeneous System Architecture ； Load balancer

參考文獻

[1] HSA white paper. http://www.slideshare.net/hsafoundation/hsa10-whitepaper.

[2] Kenneth Lee, Heshan Lin, and Wu-chun Feng. Performance characterization of data-intensive kernels on amd fusion architectures. Computer Science-Research and De-velopment, 28(2-3):175–184, 2013.

[3] Stratos Idreos Fabian Groffen Niels Nes and Stefan Manegold Sjoerd Mullender Martin Kersten. Monetdb: Two decades of research in column-oriented database architectures. Data Engineering, page 40, 2012.

[4] Max Heimel, Michael Saecker, Holger Pirk, Stefan Manegold, and Volker Markl. Hardware-oblivious parallelism for in-memory column-stores. Proceedings of the VLDB Endowment, 6(9):709–720, 2013.

[5] HSA foundation. http://www.hsafoundation.com/.