考量固態硬碟內部平行度之跨層級公平性 I/O 排程器設計方式

Abstract

這篇論文提出一個跨層級排程器的設計，在保護公平性的前提下，增進固態硬碟的效能。而因為固態硬碟的架構不同於機械式硬碟，其可藉由利用固態硬碟內部的平行度來實現更好的效能。然而，目前旨在維護公平性之現有排程器中，並沒有考慮到固態硬碟其具有內部平行度的優點，因此沒對於這方面去做優化。因此，在本篇論文中，我們設計一個跨層級排程器以充分發揮固態硬碟其內部平行的特性。當在固態硬碟內部察覺到仍有晶粒處於空閒的狀況時，固態硬碟將會通知上層位於作業系統的 I/O 排程器，告訴其可分配多一點請求至固態硬碟內部。另外，如果 I/O 排程器在為了達到更好的效能，違反公平性原則時，也將通知固態硬碟內部，將此時所事先分配之請求其優先權降低以維持公平性。為了評估我們所設計之跨層級排程器，我們將其實作在 Microsoft Research 所提出的固態硬碟模擬器，並執行合成的工作流以及由 SNIA所發布之真實的工作流。而模擬結果顯示，我們設計的跨層級排程器其效能相較於 FIOS 能提升 11% 且能不失其公平性。

This thesis presents a cross-layer scheduler design to not only improve performance but maintain fairness in SSDs. Because the architecture of SSDs is different from that of HDDs, SSDs can achieve higher performance by exploiting internal parallelism. Currently, the existing schedulers aiming to maintain fairness are not aware of the internal parallelism that SSDs have. In this thesis, we develop a cross-layer scheduler design to fully exploit internal parallelism of SSDs. When there is any idle dies in SSDs, SSDs would notify the OS and the OS would then dispatch requests into SSDs. Also, if the OS breaks fairness limitation and dispatch requests into SSDs in advance, it would send these messages to SSDs and let these requests have lower priority in SSDs to maintain fairness. To evaluate our cross-layer scheduler design, we implement the cross-layer scheduler design in Microsoft Research’s SSD simulator and use synthetic workloads and real-world workloads published by SNIA. Simulation results show that the performance of our cross-layer scheduler design is 11% higher with less loss of fairness on average compared to those of FIOS.