以細粒度之寫入排程機制增進相變化記憶體於能源預 算限制下之效能

Abstract

作為被看好的新一代記憶體之一,相變化記憶體 (Phase-Change- Memory) 擁有多項優於傳統隨機存取記憶體 (DRAM) 之優點,如較高 的單元密度 (cell density) 及零靜態功率 (leakage power) 等。除了寫入耗 損問題 (wear-out) 及較長的寫入延遲 (write latency) 外,相變化記憶體 也面臨了因高寫入功率所帶來的能源議題。先前研究指出相變化記憶 體之內部能源供應並無法支援同時對所有寫入請求之操作並提出能源 預算限制 (power budget limitation),而我們發現此現象隨著最底層快取 記憶體 (last-level cache) 的快取線 (cache line) 之增大變得更嚴重。雖然 區段寫入技術 (division programming) 是個可行的解決方案,其造成之 效能耗損仍相當可觀。 在這篇論文中,我們提出細粒度之寫入排程機制以增進相變化記憶 體於能源預算限制下之效能。我們發現區段寫入技術主要有兩項缺點: 其一,所有屬於同一寫入請求之區段必須被共同地排程,意味著對無 法被發出 (un-issuable) 的區段之搶先 (preemption) 是不可行的而損害效 能。其二,將寫入請求分割為區段之方法為靜態地依照位元位置 (bit position),其喪失了動態調整各區段能源消耗之彈性。根據此觀察,我 們提出了兩個方法。首先,次請求排程 (sub-request scheduling) 使排程 器 (scheduelr) 可進行區段粒度之排程以支援對無法寫入區段之搶先。 第二,動態區段分割 (dynamic division) 可於運行時藉由搬移欲寫入之 資料以調整各區段間之能源配置 (power allocation)。我們的實驗結果展 示了這些機制能有效地增進系統效能達 31%。

As a promising candidate for future memory, Phase-Change-Memory (PCM) has several advantages over traditional DRAM, such as high cell density and zero leakage power. In addition to the wear-out problem and longer write latency, PCM also faces power issue with high write power. Prior works ad- dress the power budget limitation that the internal power supply could not sustain programming all the write requests for all banks simultaneously, and we find that this situation even gets worse with enlarged last-level cache line. Though division programming technique is a possible solution, the perfor- mance degradation is still substantial. In this thesis, we propose fine-grained write scheduling to improve the performance of Phase-Change-Memory under power budget limitation. There are two major problems with division programming of PCM write operation: (1) all the divisions of a write request are scheduled collectively, indicating that the preemption of un-issuable divisions is unfeasible and degrades the performance; (2) divisions are divided statically by bit positions, which is lack of flexibility to manage the power consumption of each divisions at runtime. Based on these observations, we propose two schemes. First, sub-request scheduling enables scheduler to schedule under the granularity of divisions to support the preemption of un-issuable write divisions. Second, dynamic division adjusts the power allocation of each division by shifting the data to be written at runtime. Our experimental results show that these techniques improve the system performance significantly by up to 31%.