電機資訊學院: 資訊工程學研究所指導教授: 楊佳玲許貴松Hsu, Kuei-SungKuei-SungHsu2017-03-032018-07-052017-03-032018-07-052015http://ntur.lib.ntu.edu.tw//handle/246246/275427由於優越的性能以及省電這兩個特色，以快閃記憶體為核心的固態硬碟已經被認定為機械式硬碟很好的替代品。因為固態硬碟的架構不同於傳統的硬碟，相較於硬碟而言，固態硬碟是透過內部的平行度以達到較高的性能。不過我們觀察到，雖然固態硬碟擁有著良好的平行度，現有的I/O 排程器在發送I/O 請求時，並沒有考慮到底層固態硬碟中的各個快閃晶圓的工作負載。這樣的行為導致在固態硬碟中出現了工作負載不均衡的現象以致於平行度可能無法被完全發揮。除此之外，既有的公平性I/O 排程器在為了保護公平性的同時，也犧牲了在固態硬碟裡的平行度。在這篇論文中我們提出一個跨層級的I/O 排層器，並且利用PCIe 設備的特性讓上層系統可以取得在固態硬碟中記憶體中的資訊，透過這種方式讓I/O 排程器能夠知道每個I/O 請求使用到的快閃晶圓，並且根據這些資訊在I/O 排程器中記錄固態硬碟中的工作負載情形。有了固態硬碟中工作負載的情形，I/O 排程器可以避免工作負載不均衡的問題，並且適當的放寬公平性的限制，最終得到最大化固態硬碟的平行度。實驗數據顯示，相較於傳統沒有考慮固態硬碟的排程器，我們的跨層級I/O 排程器可以改善20% 的性能。除此之外，在加入公平性的限制下，我們相較於現今的公平性排程器可以改善23%的性能。NAND flash-based SSDs have been regarded as great alternatives of mechanical HDDs due to their impressive performance and low power consumption. Since the architecture of SSD is different from that of HDD, SSD can achieve higher performance by exploiting internal parallelism. Despite of the rich parallelism, current I/O schedulers dispatch I/O requests without considering the loading of each die in the SSD. It aggravates unbalanced loading in SSD so that the parallelism may not be fully exploited. In addition, existing I/O schedulers use anticipation technique, which idles the device for anticipating soon-arriving request, to balance the usage of the SSD between read and write process. However, anticipation technique has negative effect on performance of the SSD. This thesis proposes a cross-layer I/O scheduler, which utilizes the feature of PCI Express device to enable the host system to access the mapping table of SSD. According to the information of mapping table, our cross-layer scheduler maintains a loading table in the host system to deal with unbalanced loading problem and selectively relax the limitation of fairness to maximize the parallelism of the SSD. Experiment results demonstrate our cross-layer scheduler can improve about 20% average I/O response time compare to conventional I/O scheduler which is not aware of the internal parallelism of SSD. Furthermore, we can achieve about 23% improvement compare to state-of-the-art fairness-oriented scheduler but without losing fairness.1106289 bytesapplication/pdf論文公開時間: 2018/8/20論文使用權限: 同意有償授權(權利金給回饋學校)固態硬碟跨層級排程器Solid State DriveSSDCross-LayerI/O Schedulerthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/275427/1/ntu-104-R02922108-1.pdf