Mitigating Disturbance Errors and Enhancing RMW Performance for PCM

Abstract

Phase-change memory (PCM) announces the beginning of the new era of memory systems, owing to attractive characteristics. Many memory product manufacturers (e.g., Intel, SK Hynix, and Samsung) are developing related products. PCM can be applied to various circumstances; it is not simply limited to an extra-scale database. For example, PCM has a low standby power due to its non-volatility; hence, computation-intensive applications or mobile applications (i.e., long memory idle time) are suitable to run on PCM-based computing systems.

The second problem of PCM is a long-latency compared to that of DRAM. In particular, PCM tries to enhance its throughput by adopting a larger transaction unit; however, the different unit size from the general-purpose processor cache line further degrades the system performance due to the introduction of a read-modify-write (RMW) module. Since there has never been any research related to RMW in a PCM-based memory system, this dissertation proposes a novel architecture to enhance the overall system performance and reliability of a PCM-based memory system having an RMW module. The proposed architecture enhances data re-usability without introducing extra storage resources. Furthermore, a novel operation that merges commands regardless of command types is proposed to enhance performance notably.

Despite these fascinating features of PCM, PCM is still far from the general commercial market due to low reliability and long latency problems. In particular, low reliability is a painful problem for PCM in past decades. As the semiconductor process technology rapidly scales down over the years, DRAM reaches 10 nm class process technology. In addition, it is reported that the write disturbance error (WDE) would be a serious issue for PCM if it scales down below 54 nm class process technology. Therefore, addressing the problem of WDEs becomes essential to make PCM competitive to DRAM. To overcome this problem, this dissertation proposes a novel approach that can restore meta-stable cells on demand by levering two-level SRAM-based tables, thereby significantly reducing the number WDEs. Furthermore, a novel randomized approach is proposed to implement a replacement policy that originally requires hundreds of read ports on SRAM.

Another problem is the absence of a full simulation platform for PCM. While the announced features of the PCM-related product (i.e., Intel Optane) are scarce due to confidential issues, all priceless information can be integrated to develop an architecture simulator that resembles the available product. To this end, this dissertation tries to scrape up all available features of modules in a PCM controller and implement a dedicated simulator for future research purposes.